As new AI agents accelerate their entry into the workplace, cybersecurity platform Zscaler is extending its Zero Trust Exchange to govern how these autonomous systems connect, access data, and operate across enterprise networks. On Tuesday, the company unveiled three new capabilities—AI Broker, AI Access Graph, and Endpoint AI Security—that together constitute what Zscaler calls the industry’s first complete Zero Trust platform for agentic AI.

The announcement, made at the company’s Zenith Live conference, targets a governance gap it argues most enterprises haven’t fully reckoned with yet. Unlike human employees who authenticate through established identity systems, AI agents create ephemeral identities, spawn sub-agents, and access sensitive systems at machine speed. They can be largely invisible to the current security tools organizations already have in place.

“Security has always evolved as an incremental thing,” said Swamy Kocherlakota, Zscaler’s executive vice president of agentic AI, in an interview. He likened traditional enterprise security to building a house: first, you add a door; then a fence to keep people out of the yard; then a path connecting the two. Each addition makes sense in isolation. The cumulative result, he said, is an industry riddled with disconnected point solutions: ”Scotch tape and chewing gum to keep it together.”

The Zero Trust Exchange is Zscaler’s solution to that problem, built on a single principle: never assume trust, always verify it. It’s a cloud-native platform that sits between users and whatever they’re trying to reach, inspecting every connection in real time rather than routing traffic back to a corporate data center for review. At its current scale, the platform inspects more than 750 billion transactions daily. Until today, every one of those connections involved a human on one end.

Creating Agent Visibility for CISOs

The problem, according to Kocherlakota, is that HTTP traffic doesn’t reveal whether it’s from regular web browsing or AI traffic. Because they’re technically indistinguishable, organizations struggle to govern traffic using their existing toolset. He claimed that since Zscaler sits between the user and the internet, it can intercept HTTP packets, identify those containing AI-related traffic, and apply AI-specific governance policies. That gives organizations something they currently don’t have: visibility into what employees and agents are sending to AI tools, and the ability to enforce business rules on the output.

Two new capabilities from Zscaler are designed to close this gap: AI Broker and Endpoint AI Security. With the former, the company’s Zero Trust Exchange becomes the intermediary for Model Context Protocol (MCP) and Agent2Agent (A2A) communications. It includes an Agent Registry to help organizations track what each agent is authorized to access and enforce permissions specific to its designated function, rather than applying a blanket policy across all agents. An agent built to read financial data, for example, shouldn’t also be able to write to HR systems or approve purchase orders.

The latter targets risks and threats on any device connected to an organization’s network. This includes laptops, desktops, smartphones, and tablets. Zscaler’s Endpoint AI Security will identify and mitigate any security issues, whether they’re hidden in browsers, plugins, extensions, or locally hosted AI tools.

As Kocherlakota put it, it’s all about giving CIOs and CISOs visibility into what’s being installed on company devices, letting them manage what an employee can and cannot do.

Zscaler Gets an AI Graph

Zscaler is also introducing the AI Access Graph, a real-time intelligence map for every identity, application, and data source within an enterprise. It’s designed to solve the problem of accumulated complexity. In a typical enterprise with 1,000 employees, Kocherlakota said the number of individual roles and entitlements can reach 10,000 or more across systems like Snowflake, ServiceNow, Workday, Oracle Financials, and Microsoft 365. Adding AI agents to the mix only compounds the number of connections with agents accessing databases, communicating with MCP servers, and operating across cloud environments. Nobody has a single view of what has access to what.

The AI Access Graph draws on technology from Symmetry Systems, which Zscaler recently acquired. Symmetry’s core capability is building knowledge graphs of entitlements, mapping which users and non-human identities have access to which systems, and whether those permissions are actually being used. It’s the intelligence layer that distinguishes the AI Access Graph from a basic monitoring tool by tracking excessive access and enforcing tighter permissions in real time. Zscaler Chief Executive Jay Chaudhry said such a capability wasn’t necessary until now, thanks to the rise of AI agents that have made the complexity of enterprise access too great to govern with existing tools.

That inability to govern AI with existing tools is, for Kocherlakota, the heart of the matter. Securing AI is not a refinement of current security practice but a genuinely new discipline, and one he believes most enterprises are underestimating. “Protecting AI is the net new,” he said. “If you don’t do protection, you will learn the hard way, and the cost of learning this hard way is not something enterprises are ready for at this time. That’s why we want to get ahead in protecting AI.”

Disclosure: I attended Zscaler’s Zenith Live conference as a guest of the company, with some travel expenses paid for. However, what I write reflects my own reporting and analysis. No one reviewed or approved this piece before publication.

The definition of an AI agent has evolved rapidly since the term first entered the zeitgeist. The earliest versions were chatbots with better manners: systems that answered questions, drafted text, and waited for the next prompt. Tool use and function calling took them a step further, allowing models to query databases or call APIs, but humans still drove every exchange.

In the past couple of years, the software vendor-led wave of agents, from OpenAI’s Operator and Microsoft Copilot to Salesforce’s Agentforce, promised autonomy but mostly delivered supervised task execution within someone’s walled garden. Real autonomy did arrive, but on vendors’ terms. Cognition’s Devin and Anthropic’s Claude Code demonstrated that agents could carry out coding tasks end to end; Manus generalized the idea beyond code; and OpenAI’s ChatGPT Agent enabled multi-step execution. Each ran in a sandbox controlled by its maker, on a session that ended when the task did.

OpenClaw: The Agent at Arm’s Length

OpenClaw broke that pattern. Launched in November 2025 by developer Peter Steinberger, the personal AI assistant runs on a user’s own hardware, lives in the messaging apps they already use, and acts without waiting to be summoned. It became one of the fastest-growing open-source AI agents, with its GitHub repository receiving more than 377,000 stars. OpenClaw recast the agent as a proactive, always-on digital worker rather than a tool you open and close.

Despite its promise, significant security concerns slowed adoption. The project reportedly leaked plaintext API keys and credentials, and its skills marketplace was found to be riddled with malware. Researchers demonstrated how a prompt injection attack could backdoor a user’s machine through something as routine as an email. And in one widely circulated incident, the agent wiped out the entire email inbox of its own user, a Meta AI alignment director.

The OpenClaw ecosystem spawned imitators. Today, there are at least four “Claw” variants: NanoClaw, Nanobot, PicoClaw, and Nvidia’s NemoClaw. Each one attacks a different limitation of the original, from security architecture and hardware resource requirements to codebase complexity.

Tech companies have also been paying close attention to OpenClaw. Executives may not be saying the name on stage at their conferences, but the agent has started appearing on their slides, tucked into ecosystem diagrams and integration lists, a quiet acknowledgment of what early adopters were already experimenting with. These companies sensed an opportunity but haven’t figured out how to adopt the technology without inheriting its risks.

That uncertainty has created an opening for other claws. NanoClaw, a stripped-down variant built around the security discipline OpenClaw lacked, has drawn praise from OpenAI co-founder Andrej Karpathy, and in May, its creator turned down a $20 million buyout offer, raising $12 million in seed funding instead.

Microsoft + OpenClaw

None of it dented OpenClaw’s momentum. At this year’s Build conference, Microsoft handed the project its biggest endorsement yet, announcing that OpenClaw will power Scout, its new assistant for Microsoft 365, and that the agent will run natively on Windows. It’s a striking turn for Microsoft, given its chief executive, Satya Nadella, once said he couldn’t ship OpenClaw because “that would be considered Microsoft launching a virus,“ even as he called it a fantastic innovation. What changed wasn’t his assessment of the technology. Microsoft believed it had built a cage strong enough to hold it.

To assuage organizations’ concerns about using an OpenClaw agent on their networks, Microsoft has placed it inside its Microsoft Execution Container (MXC), a policy-driven execution layer that “lets developers declare what an agent can access with containment boundaries enforced at runtime.” These protections are designed to keep OpenClaw from, say, erasing every file on the machine, with containment enforced at the operating-system level.

Nadella framed the containment work as what made the embrace possible. “It’s so wonderful to see OpenClaw come to Windows and have all that capability in terms of the security and that comfort to be able to have these long-running Agents and unmetered intelligence come together,” he remarked during his Build keynote.

For most people, though, the first encounter with OpenClaw will be Scout. It’s Microsoft’s first ‘Autopilot’ agent, the company’s label for agents that work autonomously over long stretches and meet enterprise compliance requirements. It also runs entirely within the user’s Microsoft 365 environment. “Scout works where you work, joining group chats and teams, handling threads in Outlook,” Nadella explained. But while OpenClaw on Windows can run either locally or in the cloud, Scout is cloud-based.

Its development shouldn’t be that surprising, though, given that Microsoft has openly telegraphed it for the past several months. Omar Shahine, who previously led Microsoft Word, leads the team tasked with building what he called “a new generation of proactive assistants, ones that lighten your load by taking on tasks end-to-end, and that can also step in proactively when they can help.”

For all the ways Microsoft reshaped OpenClaw to operate on its platform, the company is also giving back to the ecosystem it borrowed from. At Build, it said it is contributing its policy conformance system back to the project, which would let organizations running the open-source agent check their deployments against security and compliance requirements and produce an audit trail to prove it.

The logic behind Microsoft’s reversal appears to be competitive. Developers were already running OpenClaw on Windows, with or without Redmond’s blessing, and given where agentic development is headed, Microsoft had more to gain from making its platform the safest place to run the agent than from pretending it wasn’t there. And if OpenClaw is indeed the future, Scout and future Autopilots could help boost Microsoft 365 subscriptions.

Absorbing OpenClaw also buys Microsoft influence without ownership. By contributing its conformance system upstream, the company gets a hand in defining what a “secure” OpenClaw deployment looks like for every organization running it, which will help the project further stand out against other claws tackling the same issue.

Google Rebuilds OpenClaw in Its Own Image

Google is another Big Tech company that has let OpenClaw in. Sort of. In March, the company released a way for developers to integrate OpenClaw agents into Google Workspace, a tacit acknowledgment that users were bringing their own claws, whether Google built for them or not. Then, at its I/O conference in May, the company introduced its rival: Gemini Spark, an always-on agent that can compose emails, update study guides, and monitor credit card statements for hidden subscription fees, but runs entirely on Google’s stack, powered by Gemini 3.5 with the company’s Antigravity harness underneath. The capabilities are OpenClaw’s. The code, however, is not.

This suggests Google sees value in OpenClaw. The company is open to letting developers connect their agents to its apps, but it isn’t yet comfortable making the technology a native part of its platform. Or Google believes it has a better approach: an agent powered by its own model, with the security and governance controls already built into Workspace. There’s also a simpler explanation. Whatever agent sits in front of Gmail, Docs, and Calendar controls the relationship with the user, and Google has no intention of letting that be someone else’s software.

Google’s broader direction reinforces the point. The same month it previewed Spark, the company announced that its open-source Gemini CLI, an Apache-licensed tool with hundreds of contributors, would stop working with Google AI subscription plans on June 18, to be replaced by the closed-source Antigravity CLI. Even the Antigravity SDK, the developer-facing piece, is an open wrapper around a closed binary. While Microsoft moves toward the open agent project, Google is retiring its own open agent tooling.

The result, as one developer put it, is a vision in which “every agentic workflow you build runs on infrastructure Google controls, through interfaces Google designs, against models Google hosts.”

The split between the two companies comes down to how each wants to handle the same risk. Microsoft is offering an open foundation within a Microsoft-managed cage: build on the community’s project, contribute security work back, and sell the governance layer to enterprises on top. Google may be betting that customers would rather not touch the open project at all. One approach asks organizations to trust a caged version of software they’ve read alarming headlines about. The other asks them to trust Google with the whole thing.

Google has spent years building the full stack, its own chips, its own models, its own agent harness, so constructing a homegrown OpenClaw was the obvious move. Microsoft’s in-house models are far younger, and for most of the AI boom, its products have run on OpenAI’s technology. Neither side has the upper hand as it seems to be part of the ongoing debate about what’s better: open-source or closed-source.

Coming Soon: Meta

Facebook’s parent company could soon join the list of Big Tech companies drawing inspiration from OpenClaw. The Information reported that Meta is developing an OpenClaw alternative called Hatch. But unlike Microsoft’s Scout, this agent is aimed at consumers. The reason: OpenClaw is supposedly too technical for non-technical users. However, Hatch, or whatever name it’ll eventually go by, could perform functions similar to OpenClaw’s.

The Information further reported that Meta has tested Hatch on simulated versions of apps such as DoorDash, Reddit, Etsy, and Outlook. That said, consumers will have to really open up their wallets to use Hatch, as it could cost up to $200 per month. We could see it introduced next month.

What’s unclear is whether Meta will use a forked version of OpenClaw tailored for Hatch or will follow Google and use its in-house models to power the agent. As Engadget noted, Meta has a relationship with OpenClaw: its Chief Executive, Mark Zuckerberg, once offered Steinberger feedback and even tried to hire him—Steinberger ultimately went to work for OpenAI.

It’s Now an OpenClaw Agentic World

During his appearance at Microsoft Build, Steinberger celebrated Microsoft’s integration of OpenClaw. “I built OpenClaw to have access to everything—my files, machines, chats—always on and fully open source. That’s what makes it so powerful!” he said, while acknowledging that that’s what also makes companies nervous about the technology. Steinberger and his team of project contributors spent months working with Microsoft, OpenAI, GitHub, Nvidia, and others to create an OpenClaw that the enterprise would be more comfortable with. They added observability, auto-mode for permissions, and changed how access works. “It’s not all or nothing anymore,” he said. “You can pick which folder should be read-only, which ones should be hidden.”

Steinberger, however, envisions OpenClaw becoming something much bigger. He sees it as a movement, a community. That’s why he started the OpenClaw Foundation, a nonprofit designed to keep the agentic technology he created “open and neutral” and to work on any model or operating system. “We are entering an era of building these agents, more capability for the people who don’t code, and more power for those who do.”

A year ago, the world’s largest tech companies wouldn’t touch OpenClaw. Now Microsoft runs on its code, Google and Meta are rebuilding its blueprint, and all of them are arguing over who gets to define what it becomes. With validation from Big Tech, expect the floodgates to open over the coming year, with a wave of enterprise software vendors debuting their next-generation agents, powered by OpenClaw…or by one of its alternatives.

Disclosure: I attended Microsoft Build as a guest of the company, with my travel and expenses paid for. However, what I write reflects my own reporting and analysis. No one reviewed or approved this piece before publication.

Search engines learned what a good result looked like by observing how people behaved. A click followed by a long page stay signaled satisfaction, while a quick bounce back to the search results signaled a miss. Microsoft built its Bing search engine on those patterns. But in the AI era, there’s a type of user that can’t click: agents. That led the company to try something that sounds strange: interviewing the AI.

“We’ve been interviewing the agent,” Tim Frank, Microsoft’s corporate vice president for monetization, commerce, and the AI economy, told The AI Economy on the sidelines of this year’s Build conference. “What did you like of this answer? How would you like it different for you?” The AIs’ response? They want structured answers, tight token budgets, speed for follow-up calls, and disclosure of how deep the search went. Those preferences, not what we humans do, are what Microsoft’s new Web IQ service is built to satisfy.

Introduced at Build, Web IQ is a set of grounding APIs built for agents. The service supplies AI systems with real-time context pulled from across the web, spanning webpages, news, images, and videos. It’s designed to help agents not only surface the right information but also convert it into useful evidence and apply it to their reasoning. It’s the latest addition to Microsoft’s line of IQ products, joining Work IQ, Fabric IQ, and Foundry IQ.

For developers, Web IQ isn’t replacing the model underneath an agent. It’s intended to complement the LLM, but addressing two things no model can fix on its own. First, a model’s knowledge stops at its training cutoff, so it knows nothing about what happened after. And second, because a model can only hold so much, the long tail of obscure detail never makes it in.

“Microsoft is one of a very small number of companies that has invested in the infrastructure to build an index of the web, and then to make a retrieval layer that says, ‘I heard what you wanted, let me give you back—not the whole internet—the slice of what’s useful based on what you told me,’” Frank said. “It’s a very technically challenging thing to build, which is why you don’t see tons of people doing it.”

Bing for AI Agents

The company calls Web IQ a “search engine for AI systems.” The description fits because Web IQ is less a new product than an extension of the Bing platform. It sits on the same global index Microsoft has refined for over two decades, relies on the same crawlers, and inherits the same rules of the road, from honoring robots.txt to enforcing publisher access controls.

Where the two diverge, however, is everything above the index. While Bing’s retrieval and ranking machinery was tuned for a person scanning a page of results, it was rebuilt in Web IQ for machines that read differently. When a human searches for something on the internet, they’re presented with the traditional results page featuring ten blue links. For agents, this isn’t the way.

“They don’t have a limit of ten [per page] because they don’t have a viewport on the screen,” Frank explained. “For the human, it’s quite important that you get the right answer in one of the first ten links. For the agent, you actually could give many more links because it’s not very expensive for the agent to consider 15 or 1,500 items. The cost is in how many tokens you give it. Then it needs to rationalize and understand the tokens and decide if it wants to make another call or not.”

Under the hood, Web IQ is itself built from AI models, just not the kind anyone chats with. Microsoft says it chose a handful of in-house models over a sprawl of specialized ones, each with a defined job: one converts web content into mathematical representations so the system can match meaning rather than keywords, while others read, rank, and select the passages worth handing to an agent. The telling detail is how they’re trained. Microsoft optimizes them not for standalone benchmark wins but for how their output performs inside another model’s reasoning. The practical effect is worth sitting with: a small set of Microsoft models now decides what evidence everyone else’s models get to reason over.

All that evidence comes from somewhere, which raises the question that publishers have been asking for two years: does letting an AI system touch your content mean it gets permanently absorbed into someone’s model? Frank said Web IQ doesn’t work that way. The licenses underpinning the service cover inference only, meaning a publisher’s content is fetched only when an agent needs it to answer a question, is used, and is traced back to its source, with nothing baked into a model. “None of these licenses are for model training,” Frank said. Anyone seeking content to train models, he added, has to strike deals outside of Web IQ’s scope.

Why did Microsoft release Web IQ when Bing Search API offered similar capabilities? Simply put: the now-retired developer tool wasn’t made for AI agents. It was designed for human use, Frank said, and delivered a human-style SERP experience. The company rebuilt it for the AI era and deliberately gave it a new name. “We needed to be able to cleanly break from the convention and decisions we made to optimize for a human [and] to make it work excellently for AI systems. The name is intentionally separate so people don’t confuse the stuff designed for humans versus what is explicitly designed for AI.”

Best in Class For Quality, Latency, Token Density

Microsoft claims Web IQ beats every rival on quality, speed, and token efficiency. “The company that’s number two on some of [the benchmarks] is not two on all of [them], so it’s not like there was a leader,” Frank said, claiming that before Web IQ, developers needed to decide between accuracy, speed, and efficiency. “Depending on which of those heuristics you would have chosen, you’d have to work with three different suppliers. For the first time, you can actually get all three best-in-class with one supplier that has 20 years of experience in this space.”

In an evaluation of grounding satisfaction, a metric that captures whether grounding actually meets user intent across completeness, freshness, and authority, Microsoft reported that Web IQ has a higher score than competitors. This means users trust its response and achieve stronger downstream outcomes.

Speed determines whether an agent can take multiple retrieval-and-reasoning steps or must cram everything into one attempt. Microsoft reported Web IQ returns responses in under a sixth of a second, nearly 2.5 times faster than its peers.

The last variable is token consumption, something many developers and organizations are concerned about today. Microsoft claimed that Web IQ can reduce the amount of context required to produce a “given level of quality” thanks to its use of passage-level evidence and its high information density per token.

It’s worth noting, though, that despite these claims, Microsoft doesn’t name the competitors it tested against. Moreover, these comparisons are ones the company ran itself.

Impact on Enterprises

Enterprises have spent the past two years carefully grounding their agents in clean internal data. Inviting the open web into that loop sounds like a step backward. Why would a company risk polluting trusted retrieval with the same stale pages and SEO-gamed content that made the web hard to trust in the first place?

“A lot of AI applications will want to have fresh, reliable, ground-truth data that they can rely on,” Frank said. He offered a shopping app as an example. Build one on today’s AI alone, and users will hit links to products that no longer exist. What developers want underneath it is a source of honest, current answers about availability, local pricing, and promotions. That, in Frank’s telling, is the kind of use case Web IQ exists to serve, and not only for apps built around a model.

Worth remembering: when Web IQ pulls from the web, it doesn’t dump the internet on an agent. It returns just the slice that’s useful, passages the agent can reason over and trace back to a source. That design targets AI’s most corrosive habit—the answer that sounds confident and turns out wrong. Frank said he’s been burned that way “many a time,” and that the stakes vary, with a wrong answer sometimes costing nothing but other times carrying real costs. Get tricked enough times, he said, and no amount of model improvement will win back the trust.

While Web IQ can be useful for consumers, when it’s part of the enterprise, it’s not intended to stand alone. As mentioned earlier, it’s part of the broader Microsoft IQ family. Frank used a Windows 11 upgrade to illustrate how the suite works together. For a consumer weighing a purchasing decision, Web IQ alone can surface official support details from Microsoft, corroboration from independent reviewers, and alternative comparisons. But when applying that same scenario to a CIO planning a company-wide upgrade cycle, one layer isn’t enough. Web IQ informs them about what’s currently supported and sold in the public marketplace, but Fabric IQ catalogs the hardware currently deployed in the company and Work IQ shows how heavily workers use it.

“As you start to bring all these different layers together, now you can stack that to make a really good business decision. That’s what’s special about delivering it through Azure and through the Microsoft IQ suite,” Frank said.

Web IQ is currently available to a limited group of enterprise customers, with priority given to those working with Microsoft’s account teams and “developing production AI workloads that require high-quality, current, real-world grounding.”

Disclosure: I attended Microsoft Build as a guest of the company, with my travel and expenses paid for. However, what I write reflects my own reporting and analysis. No one reviewed or approved this piece before publication.

Asana has spent over 18 years telling companies how to manage work. Now, it wants to tell them how to manage AI. On Thursday, the company unveiled its Agentic Work Management Platform, a system designed to align humans and agents around the same plan, context, and governance. Asana calls it the “most significant product evolution” in its history, and it comes with three parts: a personal chief-of-staff agent called Dash, a new generation of AI Teammates, and three vertical applications targeting IT, software development, and professional services.

While many enterprise software vendors envision AI agents as being autonomous workers and humans as supervisors, Asana’s approach is more constrained and structured. It reports that although three-quarters of knowledge workers use AI in their jobs, only five percent of organizations report meaningful gains in productivity. A key reason why agents fail in the enterprise: A lack of organizational memory with no understanding of who does what, by when, or why, explained Chief Product Officer Arnab Bose. Asana’s argument is that it has spent years building exactly that infrastructure, even before agents existed to use it.

“It enables us to create this system where human beings and AI agents can work side-by-side in the flow of business,” he told The AI Economy in an interview, arguing it delivers better performance and quality outcomes compared to agents operating through chat-based tools like a Copilot sidebar. “We created the collaborative work management category,” Bose said. “Now we’re creating a new category: agentic work management.”

The ‘Easy Button’ Enterprises Are Looking For?

Asana isn’t the first to talk about how AI agents need more context to be successful. Glean, Slack, ServiceNow, and Atlassian are among those sounding the alarm lately. But where Asana has a structural edge is in four key areas, it argues.

The first is the Work Graph itself, which functions as a true context graph, tracking not just what work exists but who owns it, when it’s due, and why it matters. The second is its multiplayer interaction model, which makes all agent activity visible to the entire team, not just to the person who invoked it. Third is shared memory—as AI Teammates interact with multiple workers, they retain what they learn, so a new team member doesn’t have to retrain the agent from scratch. Finally, every agent action is logged in a full audit trail, including what was accessed, who owns the agent, and how much it costs to run.

Together, these four capabilities can help bestow the organizational memory that AI agents need to transform a productivity experiment into a reliable system of record. “There’s tremendous pressure on all businesses worldwide to become AI native,” Bose said. “Everybody feels like the technology space is moving super fast, that they’re missing out on an ability to accelerate their business, accelerate work, and gain an edge.” But while technology continues to improve, there has yet to be a so-called “easy button” for the agentic enterprise, he contended.

Asana thinks it has cracked the formula and has the “easy button” companies need.

“Not only are we giving you this easy button to identify those processes, the AI agents are making it easier than ever before to create the projects, keep them updated, and create this compounding benefit for your company,” Bose said. “Don’t think of Asana as this work management tool where you have to track and log work. Think of it as this operating system that gives you the easy button to make your enterprise truly agentic.”

Asana Dash

Asana Dash is a personal AI assistant built for the individual worker. It’s distinct from the company’s AI Teammates, which are multiplayer by design. Where Teammates operate as collaborative agents within a team’s workflow, Dash is yours alone. It understands your goals, priorities, and pending work, handling what it can and surfacing what needs your attention.

“We are explicitly not calling [Dash] a Teammate because we want people to believe it’s their assistant—it’s your Dash,” Bose said. “It has all the capabilities of a Teammate, and it can invoke Teammates for you. It can update projects, tasks, portfolios—things like that. But it really is for you.”

Asana subscribes to a vision of digital labor in which AI agents handle the busy work, freeing human workers to focus on higher-order decisions and judgment calls. But as the pace of work accelerates, every worker increasingly becomes a manager of AI agents, and managing agents while staying in the flow of work requires its own layer of support. That’s where Asana Dash fits in.

Bose shared that he’s using Dash to generate a morning brief for himself, scanning his email inbox for unread emails over the past 24 hours, unread Slack messages from select channels, and updates and activities on critical projects he’s tracking. Each morning, he receives a notification with a summary, action items, and required approvals. “All of these things are just helping me get going with my game much faster,” Bose said.

New AI Teammates

Asana’s AI Teammates were introduced two years ago as customizable workflows that can take action when needed to get jobs done. They’re also powered by the company’s Work Graph data model. In September 2025, Asana upgraded its Teammates so they function as actual coworkers and operate actively rather than passively. They only recently became generally available in March.

The next-generation Teammates announced on Thursday are more capable, more connected, and easier to adopt. Asana has redesigned the chat interface for interacting with these agents and added in-product recommendations that surface the right Teammate for the job. A new Skills library lets teams extend Teammates with additional capabilities tailored to their specific workflows, capturing repeatable work patterns. And new integrations with Gmail, Outlook, Slack, HubSpot, Figma, and Canva allow Teammates to handle multi-step work across apps workers use daily.

Asana is also launching industry-specific AI Teammates for manufacturing, retail, and other sectors, expanding beyond its existing marketing, IT, and operations agents. Bose said one such Teammate is designed to address launch planning: “People are using Asana to look at a marketing or product launch. There are many different stakeholders publishing updates, and you have to go through as a human being and look at tech space status updates to glean if a project is on track or not. These launch planner Teammates can automate all of that process.” In other words, it’ll proactively scan through status updates and flag any risk or remediation steps before guiding the worker through the process. The Teammate will then be responsible for chasing down stakeholders for updates, analyzing them, and flagging any additional risks.

“The human being is elevated to being the tastemaker, where they can evaluate if the quality of the remediation is what they want, if they want to run a different play, [or] how they want to do stakeholder management with other human beings,” Bose explained.

Apps for the Human-Agent Operating System

Beyond the core platform, Asana is making its first foray into verticalized applications, packaging its agentic capabilities for the teams whose work most depends on coordinated execution.

Asana Service Management brings together support ticketing and project management and is aimed at IT, HR, facilities, and other service workers. The app features a self-learning knowledge base that can improve case deflection over time. The company claimed that if another team needs to get involved, Asana can move from a ticket to a project without sacrificing context.

The move puts Asana in direct competition with Atlassian, ServiceNow, and Salesforce, and Bose doesn’t shy away from it. “We are seeing a tremendous amount of interest from our customer base in replacing their existing IT ticket management tools with something more modern,” he said. Many customers, Bose added, already run Asana alongside ITSM tools, using it for project management. “We are combining those two things, and in the cases where people buy us, they will be displacing an existing ticketing product.”

Command by Asana is a software development system designed to turn ideas into outcomes. It is not a coding agent or a standalone command-line interface (CLI), but rather a framework orchestrating the entire development lifecycle, from generating product requirements documents (PRDs) and well-documented tickets to coordinating with coding agents, tracking pull requests, and keeping engineering and product leaders current on velocity, risk, and delivery timelines.

Asana Client Management addresses new client onboarding by coordinating all deliverables and requested approvals. It features six pre-built specialized AI Teammates designed for agency work, such as client intake, project staging, producing deliverables, and communicating status updates.

Asana 2.0 Coming Later This Summer

Bose described the Agentic Work Management Platform as Asana 2.0, elevating collaborative work management to a mode in which humans and agents work together from a shared system. The full suite is expected to be available later this summer, at which point every Asana customer will automatically receive a baseline allocation of AI Teammate executions and Dash access, no procurement required.

Thursday’s announcement caps off a busy week for Asana. Earlier, the company revealed it had acquired the no-code, agentic development platform startup StackAI for $75 million. However, Bose noted that today’s announcement didn’t involve StackAI at all, at least for now. Once Asana’s Agentic Work Management Platform launches, the company plans to let customers build agents in StackAI and then integrate them with AI Teammates within Asana.

“That agent that is doing multi-party orchestration will now get the benefits of organizational context with the Work Graph, shared memory, a multiplayer experience for human beings, and the governance and activity tracking that we have within Asana,” he said.

There’s a lot riding on this bet. Asana has been heavily impacted by generative AI, as have most SaaS companies—SaaSpocalypse, anyone?—with its market valuation sliced nearly in half. Whether its pitch as the operating system for human-agent teams wins over more customers remains an open question. Bose said that Asana 2.0 is about “reinventing our existing $800 million customer base.” Still, it’s hardly a unique position in a market where ServiceNow, Salesforce, Microsoft, and others have all staked their own claims to owning the agentic enterprise.

For most of the generative AI era, Microsoft has been the most prominent buyer of someone else’s intelligence. Its Copilot stack runs on OpenAI’s GPT models. Azure customers reach for frontier reasoning through the partnership. The arrangement has worked, and it has also been a structural dependency that the company has spent the last year visibly working to reduce.

At its Build developer conference this week, Microsoft took another step towards that independence, launching seven new in-house models from its AI Superintelligence Team. The headliner is MAI-Thinking-1, Microsoft's first reasoning model, that the company says matches Anthropic's Opus 4.6 model in coding according to its own testing, and is designed for complex multi-step instructions, long-context reasoning, and code generation, all at what the company calls a low token cost.

In addition to MAI-Thinking-1, Microsoft is releasing new versions of its homegrown image generation, transcription, and voice models. It’s also debuting a new coding model that could rival coding models from OpenAI and Anthropic.

These seven models are part of the company’s push to create what it calls “humanist superintelligence,” state-of-the-art AI with capabilities “explicitly designed” to serve people and organizations rather than replace them. “The type of AI that we create really does matter,” Mustafa Suleyman, the chief executive of Microsoft AI, proclaimed on the Build stage. “We need an AI that places humanity first, that always prioritizes human well-being and human progress.”

He added that the MAI models “are all built with real attention to detail and a commitment to making very practical and efficient tools that are tuned to just how you work in the real world.”

Microsoft’s First Homegrown Reasoning Model

MAI-Thinking-1 is a mid-sized reasoning model with 35 billion active parameters and a 128K context window. Microsoft says it was trained from scratch, with zero distillation, on enterprise-grade, clean, and commercially licensed data. In other words, the model’s capabilities are its own rather than borrowed from another lab’s outputs, such as OpenAI, and the training corpus is the kind Microsoft can defend if a copyright claim arrives.

Through reasoning, AI models can work through complex problems step by step rather than making a prediction in a single pass. With MAI-Thinking-1, Microsoft joins a field that has gotten crowded fast. OpenAI, Anthropic, Google, xAI, DeepSeek and Alibaba have all shipped reasoning models since OpenAI introduced the form factor with o1 in late 2024. Microsoft was the conspicuous absence in that group; today it isn’t.

Suleyman noted that MAI-Thinking-1 was preferred by independent human raters on search over Anthropic’s Sonnet 4.6. “It’s achieved 97 percent on AIME 2025, which is obviously the key measure of its general-purpose reasoning abilities, but most importantly of all, it’s now at 53 percent on the SWE-bench Pro, which places it right alongside Opus 4.6, at least on the toughest coding benchmark that’s out there.”

MAI-Thinking-1 is now available in Microsoft Foundry in private preview.

The Rest of the MAI Model Slate

Beyond MAI-Thinking-1, Microsoft used Build to update four model families it has been building out since last summer.

To start, Microsoft is broadening MAI-Image-2.5’s reach. The base model launched a week ago, ranked third on Arena’s text-to-image leaderboard, with Microsoft citing gains in text rendering and commercial imagery. Today’s news adds a flash variant tuned for speed and cost, the family’s first move into image-to-image work, and availability across PowerPoint, OneDrive and Foundry. “Flash is here for super-efficient production workloads, while 2.5 gives you that maximum fidelity and professional-grade performance,” Suleyman said. Microsoft also says the model surpasses Google’s Nano Banana Pro on ELO.

The pace behind that announcement is worth noting. Microsoft introduced MAI-Image-1 earlier this year, debuting in the top 10 of LMArena’s text-to-image leaderboard and shipping into Bing Image Creator and Copilot. MAI-Image-2 followed in March, with an emphasis on photorealism and readable in-image text. In April, the company released MAI-Image-2-Efficient, a faster, cheaper variant for high-volume production workloads.

Next, Microsoft is extending MAI-Transcribe’s footprint. The base model launched on Foundry on April 2, supporting the 25 most-used languages at $0.36 per hour of audio and delivering 2.5 times faster performance than Microsoft’s Azure Fast transcription for batch workloads. Today’s update brings coverage to 43 languages, with streaming support promised in a later release. Suleyman boasted that MAI-Transcribe-1.5 outperforms Google’s Gemini and OpenAI’s flagship transcription models. Microsoft has integrated the new model inside GitHub, Teams, Copilot, and the Dynamics 365 contact center. It’s also available inside Foundry.

MAI-Voice is perhaps Microsoft’s oldest model. It debuted in August 2025 in Copilot Daily, Podcasts, and Copilot Labs, capable of generating a full minute of audio in under a second on a single GPU, before reaching commercial availability on Foundry in April. “It has beautiful prosody, natural-sounding delivery, fine-grain emotional control, and it’s available in 15 languages, with many more coming soon,” Suleyman said. Microsoft has also unveiled Voice to Flash, which supports ultra-latency-sensitive voice agents, something that’s “the big thing in 2026.”

Meet Microsoft’s Coding Model

Microsoft owns one of the most widely used developer platforms on the market, GitHub, along with GitHub Copilot, the most widely deployed AI coding assistant. Since Copilot launched in 2021, it has run primarily on OpenAI’s models, with Claude and other options added to its model picker over the past year.

Today, Microsoft is putting a first-party option into that stack with MAI-Code-1, a five-billion-parameter coding model tuned for GitHub and built to run quickly and cheaply. Microsoft says it’s now available in Copilot and VS Code, though it didn’t specify whether the Copilot reference here means GitHub Copilot, the developer assistant, or the so-called super Copilot app. Whichever it turns out to be, MAI-Code-1 gives Microsoft something it has lacked for the entire history of its AI coding work: a model it owns running inside the products that defined AI-assisted coding.

The timing of an ‘ultra-efficient’ coding model is hard to read as anything other than intentional. On the day before Build, GitHub moved Copilot to usage-based billing, replacing flat-rate subscriptions with AI credits consumed by token use. Developer reaction has been sharp; some users have shared projections of 10x to 50x increases on agentic workflows, and TechCrunch declared the end of Copilot’s ‘golden age.’ A model designed to consume fewer tokens per request is the most direct technical answer Microsoft can offer to a problem its own pricing change just made acute. Whether MAI-Code-1 actually delivers on that depends on numbers Microsoft hasn’t published. The positioning, though, is unambiguous.

It’s worth asking, given the visible reshaping of the Microsoft-OpenAI relationship, why today’s slate doesn’t include a head-on competitor to GPT-5, Claude or Gemini at the general-purpose flagship tier. The likely answer is enterprise demand. These organizations already have access to a flagship LLM, and dislodging an entrenched default inside the same buyer can be an expensive, low-margin fight. Instead, Microsoft’s strategy is about differentiation by category: a reasoning model, a coding model, refreshes across image, voice and transcription. Build the models for categories where the market hasn’t settled on a default. Then, over time, as organizations embrace Microsoft’s humanist AI philosophy, they may want to use its general-purpose LLM whenever that is released.

Disclosure: I attended Microsoft Build as a guest of the company, with my travel and expenses paid for. However, what I write reflects my own reporting and analysis. No one reviewed or approved this piece before publication.

Kevin Scott always has something to say on the eve of Microsoft Build. The company’s chief technology officer shares his read on where technology is headed, typically in step with whatever the company is about to announce. This year, I joined executives, developers, and my fellow journalists in a downtown San Francisco bar as Scott laid out five observations about AI and the gap between what people expect from it and what it can actually deliver. His message: there’s a lot of work ahead if we want AI to live up to its hype.

Capability Runs Ahead of Deployment

The first observation centers on the concept of capability overhang: the idea that AI models are more capable than the tasks for which we use them in the real world. “Just because an AI is highly capable of something does not necessarily mean that quick deployment will follow,” Scott said. In other words, when it comes to using AI, we’re leaving a lot more potential on the table.

By his read, agentic coding is the area where AI’s capabilities are “furthest ahead.” It’s letting people write more code, making strong programmers more effective, and giving people who have never written a line the means to build something for the first time. But even where AI is at its most capable, deployment can still lag.

“It’s not necessarily like ‘Field of Dreams’—build it, and they will come,” Scott said. “In some places, yes, and in some places, no, and so we just shouldn’t have uniform faith that as the AI model capabilities improve, we’re going to get this crazy fast deployment everywhere.”

He admitted that it could be a bit of a controversial take, at least for those in Silicon Valley.

Progress Follows the Feedback Loop

Setting that debate aside, Scott moved to his second observation, though it’s really the engine underneath the first. Coding races ahead, he argued, because it has the tightest feedback loops of any domain. A closed feedback loop is a cycle in which the output of a process gets measured and fed back in to sharpen the next round, over and over.

With agentic coding, the model writes code. You can immediately check it with existing tools: does it compile, and is it well-formed? Models can also generate the tests themselves, so you can see whether the code actually does what was asked. Then, experts tune the models in post-training to prefer better outputs. And finally, the developers using the product generate even more feedback.

“The capability of the models is going to improve fastest where these feedback loops exist,” Scott said. “In some cases, the feedback loops, to the extent that they exist at all or can exist, are not going to necessarily be as fast as some of the ones that we have with things like software development.”

Take particle physics, for example. The models can already propose new experiments, Scott contended, but that’s where the loop stalls. Running an experiment requires expert labor and access to expensive, oversubscribed equipment, and the only way the results reach the model is through published research. There’s no fast, automatic check the way there is with code. And this isn’t a matter of waiting for the next, bigger model to close the gap. Absent a breakthrough, like simulators accurate enough to the physics of the world to generate usable training data, some fields just won’t move at coding’s pace. “You shouldn’t expect some things to go terribly fast,” Scott said.

Fast Software, Slow Organizations

His third observation is one anyone who works in the enterprise will recognize: the software can now move faster than the organization around it. “Just because we can make software go fast doesn’t mean we can make organizations go fast,” Scott said.

Sure, plenty of workers are excited about these tools and eager to start building. But most real work runs into constraints that a quick demo never does. One reason is that the work might sit in a regulated industry like finance or healthcare, which limits what can be shipped. Internal systems sit behind walls that AI agents can’t reach, locked inside infrastructure never built for them. It could also be a last-mile problem, where a tool makes one task “near infinitely faster,” in Scott’s words, only for the next bottleneck to surface right behind it. A lot of the work ahead, he said, looks less like building and more like plumbing.

He framed it as both bad news and good news for developers. The plumbing is tedious, but it’s also where the work is. There’s “a lot of opportunity out in the world right now,” he said, in making systems more accessible to agents.

Resistance to change is another reason organizations move slowly. People can feel threatened by a new process, one that upends how they’ve worked for years. As Scott pointed out, it takes “a lot of proof” that the new way is better before people will change, and he didn’t fault them for it. “That is a perfectly reasonable and valid point of view to come from,” he said. There’s also a perception problem. When capability moves as fast as it has, fast enough that the agentic coding he’d been describing wasn’t possible just months ago, people struggle to register how quickly the ground is shifting beneath them.

Activity Isn’t Value

Another observation lands more like a warning: just because you can use AI to build something doesn’t mean what you’ve built is valuable to anyone else. It’s a trap businesses of all sizes fall into, and startup founders know it best. Plenty begin as a fix for someone’s own problem, only to learn that solving it for yourself isn’t the same as solving it for a market.

Scott shared an example of an app he built during a recent flight back from Japan. It was a “frivolous” meme chat app that he could use to “irritate” his two teenage kids. Yet despite the 65 pull requests he churned through building it, Scott realized that it was “valueless,” good for nothing beyond his own enjoyment. As Scott told the room: “We now have this new tool, and we can just do a lot more with it—we can have a lot of output, build more complex things. That doesn’t necessarily mean that the things that we’re building are super valuable, that they’re going to land in users’ hands, solve problems that they have, [or] increase the top line on the businesses that we’re running.”

His advice: Developers need to pay close attention to how value is measured and to the feedback they’re receiving.

Autonomy Doesn’t Earn Trust

Scott’s last observation takes on the word the industry has fallen in love with: autonomy. Just because you can build a system that handles a task end to end, he argued, doesn’t mean you can trust it to.

“You have to build systems in a way that are doing complex things, where people can trust that the systems are doing them correctly and in a way that’s aligned with their interests and values,” he said. “That’s a new way of thinking about software.”

Scott pointed out that the definition of what developers do has shifted. The old job was to build software that works, a clean user experience, and a product that passes its test suite with good coverage. That’s no longer enough. The new bar, he said, is software people can trust to do complex things correctly and in line with their interests. “What does it mean to build trustworthy software?” he asked. The answer carries real weight, because “until you get to real trustworthiness, no one’s going to give work to these software systems to do fully autonomously on their behalf.”

No Silver Bullet

None of this gets solved by waiting for the next model. “There is no silver bullet,” Scott said, and the problems he’d laid out all evening don’t disappear simply because AI gets more powerful. Solving them requires a blend of technical, societal, and organizational work, as well as addressing legacy systems and plumbing issues.

He warned that in some cases, there’s going to be “more work than we’ve ever seen before,” requiring developers to “engage in ways that are even more intense than what we see right now.”

The capability of AI is real, and so is the promise. But Scott’s point, the night before Microsoft spent its keynote selling what AI can do, was that none of it pays off on its own. The hype assumes scale does the work. The reality is that we still have to.

Disclosure: I attended Microsoft Build as a guest of the company, with my travel and expenses paid for. However, what I write reflects my own reporting and analysis. No one reviewed or approved this piece before publication.

Glean has reached $300 million in annual recurring revenue (ARR), the company said Thursday, an amount that tripled in just 15 months. It attributes the growth to stronger enterprise data context, broader departmental deployment, and higher-than-average user engagement.

This milestone signals that Glean is holding its ground, even as enterprise software vendors like Slack, Dropbox, HubSpot, and Atlassian launch competing features. Underscoring its appeal, the company reported that its Fortune 500 customers “nearly doubled” year over year, though it offered no further specifics. It did note that more than 85 percent of customers are using Glean across five or more departments, suggesting that the enterprise search platform is not a point solution for just one team or workflow. And nearly half of Glean’s monthly active users return on a daily basis, more than twice the rate typical for enterprise SaaS products.

The product numbers reinforce the revenue story. Compared with generic alternatives, Glean reported that users preferred its output 2.5 times more often. One key reason: It’s cheaper. The company claimed its approach used 30 percent fewer tokens. And in a time when token consumption is being closely tracked by finance and IT teams, every token saved is a dollar earned.

For a platform employees are expected to rely on as heavily as they rely on a search engine, token costs are not a minor implementation detail—they are a direct constraint on how often workers can use it. If workers are forced to ration tokens, the daily engagement Glean is celebrating today likely becomes the first casualty. That makes affordability not just a competitive talking point, but an existential consideration for any AI platform positioning itself as workplace infrastructure.

“Glean’s moat is its unmatched ability to capture enterprise context across complex environments and convert it into high-quality, action-oriented intelligence that provides organizations secure access to the knowledge, people, workflows, and systems they need to drive business outcomes,” the company wrote in a release.

Over the past 15 months, Glean has expanded well beyond indexing files and surfacing answers to become a platform for deploying autonomous AI agents, tools that don’t wait to be queried but proactively orchestrate workflows, surface relevant knowledge, and take action across an organization’s connected tools and data sources. Its strategy has worked so well that the company has even released a framework for Chief Information Officers and IT leaders to help them create a path for scaling agents across their organizations.

That said, Glean’s figures warrant the standard caveat. ARR has no standardized accounting definition, and there’s growing scrutiny of AI startups on how they calculate and report this figure. Critics point to the ease of annualizing a single month’s revenue, variable pilot-to-production conversion rates, and high churn as factors that can inflate numbers. Glean’s $300 million figure is self-reported, and as TechCrunch noted, the company uses a consumption-based model in which customers pay per use—it “by definition doesn’t have a strictly recurring component.”

What’s clear is that enterprises are spending real money on Glean and coming back to use it daily. But the competitive window is narrowing. “The first four or five years of our existence, we had no competition. Given how important search is to make AI work in the enterprise, every single company in the world wants to be in this space,” Arvind Jain, Glean’s chief executive, told TechCrunch. The next 15 months will show whether Glean can keep outrunning them.

We won’t have to wait long to see what the company does next, however. Next month is the startup’s Glean:Go conference, where it is expected to unveil new enterprise AI and connected workplace products and features.

Building reliable AI agents has traditionally meant doing most of the hard work before anyone uses them. Developers run lengthy offline evaluations against labeled datasets, measure performance across quality, accuracy, cost, and style benchmarks, make improvements, and repeat the cycle until the numbers look acceptable. Only then does the agent get deployed to users.

CoreWeave thinks the sequencing is wrong. Labeled datasets can’t cover every real-world scenario, and real users reliably find the gaps. The result: agents that perform well in testing, but disappoint in the wild. The GPU cloud provider’s new agentic AI platform flips the model: deploy agents to users immediately, then let real-world usage generate the signals that drive improvement.

The platform combines CoreWeave’s serverless reinforcement learning and production inference with two products from Weights & Biases, the AI development tool provider it acquired in 2025: W&B Weave for observability and W&B Skills for autonomous improvement. Together, they form what CoreWeave calls the Superintelligence Loop, a closed feedback cycle between training and inference that helps agents compound their reliability over time.

In practice, agents are deployed immediately, bypassing lengthy offline evaluation cycles. W&B Weave tracks production behavior by capturing and classifying user interactions and surfacing failure modes. Those signals feed into CoreWeave’s Serverless RL, which post-trains the model on real-world data. CoreWeave boasts that its backend has been proven to reduce costs by up to 40 percent and accelerate training by approximately 1.4 times, with no loss in quality.

In the final step of the cycle, the improved agent returns to production before the process repeats.

While it may seem unorthodox to deploy agents without extensive prior training, the approach has precedent. Recommendation systems from Netflix and Spotify, for instance, have long operated on a similar principle, launching with baseline models and continuously improving based on real-world usage rather than waiting for perfect pre-trained accuracy.

The critical difference with AI agents is what happens after deployment. Without reinforcement learning driving continuous improvement, shipping early just means failing in production. RL has historically been out of reach for most enterprise teams as it’s too GPU-intensive and operationally complex. CoreWeave’s platform puts Serverless RL at the center of the loop, making that continuous improvement mechanism accessible to enterprises that couldn’t previously run it.

“The gap between development and production has always been where agent projects stall,” Phil Gurbacki, Weights & Biases’ vice president of product, wrote in a blog post. CoreWeave’s platform is designed to close that gap — using real-world production data to continuously improve agent reliability rather than relying solely on pre-deployment evaluation cycles.

As Asana positions itself as the “operating system for human-agent teams,” it has faced a real vulnerability: its agents couldn’t execute work end-to-end across the enterprise systems where business actually runs. That changes today with the company announcing the acquisition of StackAI, a no-code platform that lets enterprises build and deploy AI agents across critical business systems. According to Asana’s filing with the SEC, the purchase price was approximately $75 million. In addition, StackAI will continue to operate as its own product and brand.

“This acquisition accelerates our roadmap substantially and unlocks the next phase of human-agent work,” Dan Rogers, Asana’s chief executive, said in a statement. “StackAI gives our customers what they’ve [been] asking for: the ability to orchestrate their most complex business processes end-to-end, across every enterprise system they run on.”

The addition of StackAI completes a trifecta Asana has been building toward: AI Teammates for day-to-day human-agent work, AI Studio for simpler automations, and now StackAI for orchestrating complex end-to-end workflows. Here’s how this flywheel could work: StackAI is the execution layer that reaches into the enterprise systems where the data actually lives, pulling it from CRMs, ERPs, document repositories, and databases, synthesizing it, and kicking off the workflow. That logic is defined within Asana’s AI Studio, establishing the rules, triggers, and automations that govern how work moves between systems and people. AI Teammates are where human-trained agents take action—already shaped by their human counterparts, they execute tasks and make decisions at the right checkpoints.

It’s a recognition that work isn’t confined to a single app but spans many tools, including customer support, ITSM, compliance, and business operations. Asana noted that StackAI can natively integrate with Salesforce, Amazon Web Services (AWS), Docusign, databases, document systems, and industry applications using bi-directional sync; it can read and execute across all of them in real time.

“This is the first operating system where an entire business can run all of its human-agent workflows—no matter which systems, agents, data, or teams they touch,” Rogers said, adding that the enduring enterprise value will go to platforms that can coordinate everything inside the flow of real operational work. “We are closing the gap from pockets of individual productivity to enterprise-wide workflow productivity—humans and agents working together at the right checkpoints, on the workflows that actually matter.”

Founded in 2022 by MIT PhD students Toni Rosinol and Bernard Aceituno, StackAI set out to help organizations create specialized agents to tackle some of the most difficult and important processes. Rather than chase easier wins, the startup deliberately targeted heavily regulated industries where the bar for security, governance, and reliability is highest. That track record—more than 130 enterprise customers including BAE Systems, LifeMD, and Raiffeisen Bank—now gives Asana credibility with a class of enterprise buyers that are the hardest to win and most lucrative to keep.

Leading up to the acquisition, the startup raised over $19 million in funding from investors, including Y Combinator, Lambda, Gradient, Soma Capital, True Capital Management, Epakon Capital, Beat Ventures, and Weaviate CEO Bob van Luijt. Its most recent financing round—its Series A—was announced in May 2025.

Today’s announcement comes on the same day as Asana’s Q1 FY2027 earnings report.

Updated on May 28, 2026: This post has been modified to include how much Asana paid for StackAI.

Walking through the Venetian Convention and Expo Center in Las Vegas on a Wednesday afternoon in May, I navigate through the crowds and vendor booths before arriving at a section called the “Playground.” It’s here where I find Jayney Howson, ServiceNow’s chief learning officer.

This wasn’t our first meeting: In 2025, we spoke on the sidelines of the company’s Knowledge conference ahead of the launch of ServiceNow University. We were back to talk about year two. But every so often, our conversation would naturally pause, and in those moments, conference attendees would wander over just to say hello to her. Around here, Howson is something of a celebrity—and she took it in stride, welcoming each person warmly before turning back to me. It’s not hard to understand why: she and her team have built something that now counts two million learners, two-thirds of the way to the three million ServiceNow initially set out to reach.

At this rate, the company will likely exceed its goal by next year’s conference. Howson acknowledged the possibility and wondered if she should start thinking about a stretch goal. However, that wasn’t really her main focus. “We’re going to keep getting bigger. We want to go deeper with that learning, though,” she said. “It’s good enough to get people [at the] top of the funnel. I want to pull them down and get more of them credentialed, more of them certified.”

The platform isn’t an original concept. It competes alongside similar offerings from Salesforce, Microsoft, Google, SAP, AWS, and IBM, not to mention LinkedIn Learning and Coursera. But ServiceNow is taking a different approach, one deliberately centered on play—the idea that failure is part of learning, not a detour from it. At this year’s Knowledge, the company doubled down on that philosophy, transforming an entire section of the expo hall into what could be colorfully described as Nintendo World for enterprise tech.

“We call ServiceNow University a playground for learning very purposely because it uses the science of play to trigger our prefrontal cortex,” Howson explained. “We wanted to create a space that feels almost child-like in its awe and wonder. This is the first time that we’ve combined ServiceNow University with community and brought it all together. What we realize is that all these people have so much knowledge, we want them to learn all together, from each other, all in one place.”

ServiceNow’s Playground consisted of a dozen activations, described as being “faithful versions of what’s happening online,” and was structured as a gamified system. Each activity earned participants 50 points—the more activities completed, the higher the level reached and the better the prizes. “It creates the idea that you want to keep coming back and learning more,” Howson said. And the points aren’t lost now that Knowledge is over—they became part of each participant’s learner profile.

An AI Guide to Help With Learning

This interactive area also gave ServiceNow a chance to showcase two new features added to ServiceNow University: AI Learning Guide and SimStudio. Both are designed to give learners a clearer path through the platform, so they feel they’re getting real value rather than blindly choosing courses they think best suit their role.

The AI Learning Guide functions as a personal coach embedded directly into ServiceNow University. Rather than leaving learners to scroll through a catalog of 1,500 courses and guess what’s relevant to their role, it asks questions, surfaces a personalized path, and provides real-time feedback along the way. A system administrator in financial services, for example, gets different guidance than one working in manufacturing. Powered by Otto, ServiceNow’s conversational AI experience, the tool can also help learners think through career transitions—mapping out what skills they need to move from an admin role into a developer or platform owner position. It’s less a course catalog and more a conversation with someone who already knows the curriculum.

SimStudio is described as a hands-on simulation environment where learners can practice real ServiceNow tasks rather than simply watching videos or completing quizzes. It observes not only whether a learner finished a course but also how they worked through the exercises. It flags whether best practices were followed and offers feedback and alternative methods accordingly.

To start, however, SimStudio is being piloted for two courses—ITSM fundamentals and administration fundamentals—, but ServiceNow said it plans to roll the tool out to more courses in the future.

AI has always been a part of ServiceNow University, but it functioned more like a recommendation engine. It understood what a learner knew or didn’t know and what was important to them. Then, it surfaced relevant courses for them. Now, AI is embedded directly into the platform where the learning actually happens. It’s not part of a separate course or lab—it’s watching the learner work through the skill in real-time. “Previously, we were AI-powered,” Howson said. “This is a completely AI-native way of looking at learning.”

She suggested that with this approach, technical training that once may have taken five weeks could now be completed in roughly five hours, “with an improved observability. We have improved confidence in someone’s ability versus them doing the course.” It’s an ambitious claim, and one that ServiceNow will ultimately need real-world data to back up.

A Crowded Playground

ServiceNow isn’t the first software company to turn its learning platform into a conference spectacle. Salesforce, for example, has made Trailhead a centerpiece of its flagship Dreamforce conference for years, even creating a visible symbol of achievement in the form of gold jackets awarded to standout Trailblazer and Agentblazer community members. The idea is the same: use the energy of a live event to make learning feel like something worth showing up for.

Other enterprise tech firms, such as Okta and Adobe, also promote learning at their conferences, though the emphasis tends to be on certification pathways rather than the kind of immersive experience ServiceNow has built. Howson keeps an eye on the competition, but she doesn’t lose sleep over it.

“I am not here to make money from training,” she said. “I am here to make sure that our customers can unlock the platform, and that means they need individuals who are as informed as possible. My North Star is [that] I need to keep redesigning learning so it’s as easy as possible to get there.”

Howson noted that she put most of her focus closer to home, making sure ServiceNow University keeps pace with a platform that ships new releases each month, with “thousands of products, innovations” that need support. In addition, Howson wanted to ensure that her learning platform remains like a playground, encouraging people to “feel safe to be outside their comfort zone.”

“We all can remember being a kid and feeling like we were safe,” she said. “This needs to feel like you’re safe to push yourself and not get it right the first time.”

The stakes for getting this right extend beyond the playground. Despite the rapid proliferation of AI tools across the enterprise, adoption remains uneven—is it because of the disconnected systems? Data problems? Slow-to-respond management? Employee skepticism? Howson believed organizations aren’t doing enough to bring their workers along. “Humans want to be as proficient as possible,” she said. “But they’re scared.” The solution, in her view, was straightforward: leaders need to understand the moment, provide the right tools, and then get out of their employees’ way.

It’s an argument that conveniently doubles as a pitch for her own platform. ServiceNow University, with AI Learning Guide surfacing personalized learning paths and SimStudio observing competency in real time, is designed to do precisely that, guiding workers through the process without requiring hand-holding.

More Than a Participation Trophy

But getting people into ServiceNow University is only half the battle. The harder question is what a certification from the platform is actually worth, and whether accelerating the path to one diminishes the credential itself.

“A skill without a signal has no economic value,” Howson said. She noted that there are hundreds of ServiceNow’s most elite certifications in existence today, each commanding significant premiums in the market precisely because earning one is genuinely difficult. Speed up the process too much, and it risks undermining the value entirely. At ServiceNow, that concern has a name: Minimum Viable Duration. “We need to make sure that learning doesn’t become so short that it’s got no value,” Howson added. “We need to make sure it’s got proof of validation. It’s an art and a science.”

Testing and validation remain “critically important” for assessing capabilities, she insisted, but the time spent should not define readiness. She pointed to the legal profession as an example: Could AI compress the path to becoming a lawyer from five years to five months? You might get there faster, Howson reasoned, but you wouldn’t necessarily be a great one. “It’s experience and wisdom that takes time to build,” she said.

The playground, it turns out, is the easy part. What happens after–the certification, the credential, the signal that tells the market you actually know what you’re doing—that’s where the real work begins.

Updated on 5/26/2026: Corrected Jayney Howson's title from senior vice president and head of global learning and development to chief learning officer.

Disclosure: I attended ServiceNow’s Knowledge 26 as a guest of the company, with my flights and hotel stay paid for. The AI Economy’s coverage is editorially independent from those that it covers. These words are my own.

This story originally appeared in the May 22 issue of The AI Economy on LinkedIn. Read the full issue here.

Google’s AI ambitions have finally caught up with Search, and the marketing playbook that ruled the web for decades may need a complete rewrite. At Google I/O this week, the company unveiled what it calls the biggest upgrade to its search box in over 25 years—an AI-powered reimagining of Search’s most fundamental interface. It’s indeed a historic change to something Chief Executive Sundar Pichai calls the “ultimate moonshot,” and that’s long been synonymous with Google’s very identity.

This isn’t AI appended to Search. It’s Search rebuilt around AI, and the difference has profound implications for how businesses think about being found online. The search engine results page (SERP) we’ve come to know has already begun transforming from generic results into something personalized to the individual. But this didn’t happen overnight—it started in earnest in 2023 with the introduction of the Search Generative Experience. When you type in a query, an AI-generated response now sits above the traditional results, changing not just what you see first, but whether you scroll further at all.

Now, Google is pushing that reimagining further than ever, leaving marketers to wrestle with a question that would have seemed radical just two years ago: Is it time to prioritize Answer Engine Optimization (AEO) over Search Engine Optimization (SEO)?

“Search is no longer just about finding links,” Jim Yu, CEO of enterprise search platform BrightEdge, told The AI Economy. To him, Google’s latest moves are confirmation of something he’s been tracking for some time: the rise of what he called the “answer” economy. “Users want a conversational, AI-led experience where answers and recommendations happen immediately, inside the search environment itself.”

Yu reasoned that this is positive news for businesses since “it creates more ways to be discovered when people are asking high-intent questions.” Still, he doesn’t think marketers should throw away their SEO playbook. The fundamentals that made this practice work—strong content, technical discoverability, domain authority, trustworthiness, etc.—are the same ones that make content visible to AI systems.

“The companies that win here will be the ones that build on a strong SEO foundation and extend it into AEO. It’s still search, but its shape has changed,” Yu said.

Anne Ahola Ward, CEO of digital marketing firm CircleClick, doesn’t mince words: “Some businesses will thrive because of these changes, and others will perish.” Consider the numbers: She said around three-quarters of queries are now resolved without users ever leaving Google, and AI’s relationship with web traffic tells an even starker story. OpenAI’s scrape-to-visitor ratio is estimated to be at 1,500-to-1, while Anthropic’s is at 6,000-to-1. AI is consuming the web’s content without returning its traffic. For marketers, the implication is uncomfortable: content visibility and site traffic are no longer the same thing.

“To be an SEO means your world is constantly changing,” she wrote in an email. “You’ve got to be a Great White Shark and keep swimming, or you will die. I’ve seen more change in the last year than I have in the past ten or more—it’s quite exciting.”

How Google Just Transformed AEO

The rise of AI chatbots like OpenAI’s ChatGPT, Perplexity, Anthropic’s Claude, and Google Gemini created what Yu has previously described as a “searchquake” in marketing. As users increasingly turned to these tools for answers, companies were forced to rethink their content strategies as a growing amount of referral traffic from generative AI began to reshape how audiences found them online.

For Google, the adoption numbers are striking. AI Overviews, the AI-generated summaries appearing atop Google’s SERP, now reach over 2.5 billion users each month. AI Mode, its fully conversational search experience, has surpassed one billion monthly users, with queries more than doubling every quarter since launch. “Google Search is AI search,” Liz Reid, Google’s vice president of Search, declared during her I/O keynote.

Now that Google has embedded AI via Gemini into the fabric of Search—it’s no longer part of the experience, but all of it—it has direct consequences for how businesses think about visibility. “Google’s latest moves make it clear that AI-led discovery is becoming part of the core search experience,” Yu said. “Marketers have to understand how AI systems interpret brand authority, summarize content, compare options, cite sources, and surface recommendations.” He suggested that marketers start to measure their AI visibility alongside classic SEO rankings.

Consumers will find the new Google Search more responsive to how they actually think and communicate. The clipped, keyword-heavy queries that once dominated Search are replaced by something more conversational. Google is now built to handle both, but increasingly optimized for the latter. It’s also deeply personalized for the individual—the SERP of old was essentially the same for everyone with identical results, though ranked by signals like domain authority, backlinks, and keyword relevance. It was optimizing for the query, not for you. The Google I/O announcements are designed to surface information relevant to your context, history, and intent.

Google isn’t just reimagining the search bar—it’s also reimagining what a search result can be. Rather than returning a static page of links, consumers will be able to generate custom layouts, interactive visuals, and functional mini-apps tailored to their queries in real time. Generative UI in Search is powered by Google’s agentic development platform Antigravity and its newest AI model, Gemini 3.5 Flash.

“Search plans the ideal response from scratch. It designs the layout, decides what custom components to build, fans out to do the research, and then finally deploys the code to build custom components…in the response,” Robby Stein, Google’s vice president of product for Search, said during the Google I/O keynote.

Google Search is also entering its agentic era. The first agents it’s introducing are a natural fit for a company built on helping people find information, operating in the background while continuously monitoring the web for changes relevant to your specific interests. And rather than waiting for you to search, they proactively surface synthesized updates and can take action on your behalf, flagging new apartment listings that match your criteria or alerting you to a sneaker drop from your favorite athlete, for example.

Ahola Ward cautions that these are features marketers haven’t been considering. “Agent search favors content where AI systems can directly extract information, not content optimized for human readability,” she said. “That’s a fundamental inversion of how we’ve been building pages for decades.”

The implications extend beyond organic search, impacting search engine marketing. “Paid search will not disappear, but the context around ads will change as more interactions happen inside LLMs,” Yu noted. “Marketers need to understand how paid, organic, AI summaries, and agentic actions work together in a singular customer journey.” The future of search, he argues, is all about retrieval, reasoning, recommendation, and action, powered by generative and agentic AI.

For marketers, this shift leaves them with an urgent strategic question: if Search is now built around the individual, what does it take to be found by one?

Don’t Give Up on SEO

For Yu and Ahola Ward, the answer starts with not abandoning what already works. Both described this moment as an expansion of SEO rather than a replacement.

“The Ten Blue Links aren’t dead, but they’re no longer everything,” Yu argued. Think of SEO as being the foundation and AEO as the growth layer, he suggested. “You need to be discoverable in traditional search as well as understand how AI answers are being constructed to see if your brand is included in those answers.”

“What changes is what you’re optimizing for,” Ahola Ward said. “You’re no longer just trying to rank. You’re trying to be the source that AI reaches for in your category.” She noted that while traditional SEO still works best for branded and navigational queries, for informational and comparison queries, “the answer is increasingly happening inside Google before anyone clicks.”

Notably, Google itself has pushed back against any notion that SEO is obsolete. The company has published guidance affirming that SEO continues to “be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems.” In other words, the AI is built on top of the same web that SEO has always served.

What Do Marketers Do Now?

The scale of Google’s Search transformation can make the impulse to act immediately feel urgent, but both experts caution against rash decisions or racing to integrate new AEO tools before the fundamentals are in place.

“Marketing first needs to understand where brands show up in AI searches today,” Yu said. “You need visibility into how you are being described, what sources the engines use, and where competitors are being favored.” He advises companies to strengthen their SEO foundation and create content that answers the questions their customers are asking, with attention to depth and clarity. And websites must be accessible to AI crawlers. “Brands need to work backward from the sources AI engines rely on, pinpointing the sources that matter for your category and audience,” Yu added. “The strongest AEO strategies will be built on strong SEO.”

Ahola Ward encourages companies to work with a developer to audit their structured data. By doing so, she said, marketers will know whether AI systems can truly understand what the business is and what makes it credible. She also recommends that content be written by a human and easily summarized without sacrificing meaning. “Everything we create has two audiences—human and our robot friends, the agents,” she said.

Underscoring this, The Economist revealed this week that it’s creating two versions of its site, tailored for humans and for agents. “Every brand should be asking whether its content architecture serves both humans navigating a page and machines parsing it for synthesis, as those two things tend to conflict,” Ahola Ward added. “The underlying principle here is that AI systems are now a genuine audience with different consumption patterns than humans, which is key to remaining discoverable in the next few years.”

She also warned marketers to be skeptical of agencies suddenly rebranding themselves as “AEO experts,” saying, “People have been wanting to rename and rebrand SEO as long as I’ve been one (since the 2000s). The shops thinking about agent tooling, structured data, and machine-readable content for the last few years are the ones worth listening to. Everyone else is just branding themselves and their reports.”

To Yu, this marks the start of search’s next chapter, created by AI because “it changes how information is synthesized and how decisions are made.” For companies to remain visible and protect their search presence, he contends they need to adapt early by building on their SEO fundamentals and extending into AEO. “They will shape how their brands are understood in the AI era.”

Docusign made its intentions clear two years ago when it pivoted beyond e-signatures into the Intelligent Agreement Management space. Now, it’s ready to show what that actually looks like in an AI agent world. At its Momentum conference on Thursday, the company announced that its AI engine for agreements, Iris, is expanding into agents along with a broader ecosystem of integrations designed to keep deals moving without waiting on humans.

“Agreements are very inefficient,” Sagnik Nandy, Docusign’s chief technology officer, told The AI Economy in an interview. The company built IAM to address that problem, injecting AI across every stage of the agreement journey, from creation to signing to management. Now it’s moving beyond a reactive platform — its new Iris assistant and agents are designed to be proactive, advancing agreements. “Iris is going beyond its intelligence abilities to the agentic world,” he said.

The Iris assistant is purpose-built for agreements, trained on Docusign’s proprietary data rather than generic large language models alone. That’s important because a model trained on public data doesn’t have the necessary context to understand how deals, such as commercial agreements, are actually structured.

Iris also powers a set of new AI agents that Docusign built to tackle specific end-to-end workflows that have historically required the most back-and-forth between teams. “You can imagine redlining, renewal—entire flows we have agents for that,” Nandy said. The goal is to cut the time agreements spend waiting on humans before they can be pushed to the next stage.

To further its agentic ecosystem, Docusign is unveiling Agent Studio, a destination for creating and deploying custom agents tailored to how an organization manages deals, renewals, and approvals. “The beauty, going beyond standard tools, which we leverage, is that we also infuse your entire agreement context, our domain knowledge, our proprietary consented data advantage—all infused into that process,” Nandy said.

Truthfully, none of this is particularly new on its own—nearly every tech company has announced some version of the technology. However, what Docusign is doing with it is worth paying attention to. Nandy points to three areas where the company believes it stands apart: its data advantage from processing hundreds of millions of agreements from 1.8 million customers, its deep integrations that surface agreement capabilities inside the tools teams already use, and its compliance infrastructure built to satisfy the demands of 95 percent of Fortune 500 firms.

To illustrate his point, Nandy offers a scenario: a sales contract that accumulates amendments over time—new SKUs, additional clauses, and revised terms—each saved as a separate PDF. For a system uneducated in the world of agreements, it would see ten files and treat them as ten agreements, producing incorrect answers to basic questions such as how much a company is spending with a given customer. Docusign’s argument is that it knows better, recognizing that those files are one contract’s history.

“If you don’t take this notion that an agreement is actually a living set of documents that come together to create an agreement, it will not work the same way,” Nandy said.

He noted that Docusign has worked to transform agreements into living schemas, extracting key information automatically while letting customers define additional constructs specific to their business. “That semantic layer becomes a key artifact that now you can query, [and] other systems can plug into,” he said.

Along with its Iris announcements, the company is also introducing specialized IAM offerings for human resources and sales. With IAM for HR, Docusign’s platform will automatically collect key information from employee agreements, reducing repetitive tasks, and ensuring systems are in sync. It comes with mobile I-9 verification and integrates with Workday and Greenhouse.

There’s also IAM for Sales that targets a familiar frustration: sales teams spending more time chasing approvals and re-entering data than actually closing deals. By embedding directly into Salesforce and Microsoft Dynamics 365, it enables sellers to generate agreements, route approvals, track negotiations, and manage renewals without leaving their CRM. And the addition of three native experiences—Agreement Desk, Agreement Prep, and Agreement Manager—keeps the entire workflow connected in one place.

“Every business runs on agreements, but until now, they’ve been static records of work that already happened,” Allan Thygesen, Docusign’s chief executive, said in a statement. “Docusign is the only platform with the full context of your agreement history and relationships, and that’s what allows us to turn agreements into something more powerful: systems that can guide business decisions and move work forward.”

Docusign said that its Iris AI assistant, agents, and Agent Studio are available today in beta for select customers. It’s expected to roll out in the U.S. starting in July. Additionally, IAM for HR will be available for select customers in beta next month. The company’s MCP server is now in beta globally in English.

AI agents are getting more capable by the day, pushed forward by computer-use frameworks like OpenClaw and its variants. Enterprise adoption should be accelerating—and it would be if not for the security and governance issues: No one wants an agent wiping out an entire database in nine seconds without permission.

On Tuesday, Automation Anywhere unveiled EnterpriseClaw, a new offering designed to give enterprises the control they need to actually trust claw-style agents with real work. The announcement, made in collaboration with Cisco, Nvidia, Okta, and OpenAI, is the latest sign that the autonomous enterprise is no longer a concept. It’s being built right now.

“For AI to have a transformational impact on business, it needs to be able to do work where the work actually happens,” Mihir Shukla, Automation Anywhere’s chief executive, said in a statement. “Many claw-style AI agents are incredibly powerful, but most were designed for isolated environments or individual users—while enterprise operations span teams, cloud platforms, desktops, on-premises systems, and highly regulated industries.”

At its core, EnterpriseClaw runs on Automation Anywhere’s Process Reasoning Engine, which controls the automation, and its Contextual Intelligence Graph. On top of that sits a partner stack handling the enterprise hardening: Cisco contributes AI Defense and DefenseClaw for agentic security, while Nvidia brings OpenShell, NIM microservices, and its Nemotron open models for on-premises deployments; Okta handles identity management and authentication; and OpenAI’s models, including GPT-5.5, power agentic development.

The claw trend traces back to OpenClaw, the open-source project Austrian developer Peter Steinberger first published in late 2025. It became one of the fastest-growing agentic projects in GitHub history: a local assistant that handles tasks on your behalf across messaging apps, files, and services. But its power came with a documented liability: Researchers discovered third-party OpenClaw skills could perform data exfiltration and prompt injection without user awareness.

That gap invited competition from alternative frameworks—NanoClaw, PicoClaw, and Nanobot—each emerging to fill different niches. Out of all of them, NanoClaw made the strongest claim to security consciousness. None seems to have matched OpenClaw’s dominance, though, as it remains the most popular project in the space.

Its popularity hasn’t gone unnoticed in enterprise software. Companies like Salesforce and Adobe have been tracking the trend closely—claw references have become a recent fixture in some keynote addresses at major industry conferences. But others have gone further. Microsoft has assembled what team lead Omar Shahine calls an “Ocean’s Eleven”-style group to build OpenClaw-enabled tools, including a Teams plug-in, a ClawPilot desktop experience, and agents that operate around the clock within Microsoft 365.

Nvidia has been building its own claw implementation as well. At GTC in March, it announced NemoClaw, a version of OpenClaw wrapped in Nvidia’s OpenShell runtime, which the company says is designed to prevent malicious skills from interfering with agents that have unconstrained file system and network access.

Where EnterpriseClaw differs from its peers is scope. It’s not a single security layer or a model wrapper. Rather, it’s a coordinated stack of security, identity, infrastructure, and AI capabilities designed to give enterprises end-to-end control over claw-style agents, built on technology from established vendors like Cisco, Nvidia, Okta, and OpenAI. Until now, no comparable solution had gained meaningful traction in the enterprise.

There’s also a neutrality argument. EnterpriseClaw is designed to work across third-party agent frameworks and internally built agents alike. This means organizations aren’t locked into Automation Anywhere’s ecosystem to benefit from it.

Automation Anywhere said that EnterpriseClaw is currently in preview with general availability planned for later this year.

Zendesk spent years building software to manage customer service tickets. Now, its attention has turned to building an AI workforce that resolves them. It’s a “fundamental reset” of its platform, one where deflection-based bots are replaced by specialized AI agents that operate across all channels.

At this year’s Relate conference, Zendesk is unveiling autonomous AI agents for Employee Service; an expanded agent portfolio across all customer channels; a new copilot suite for human teams; a no-code Agent Builder; continuous quality measurement; deeper knowledge infrastructure; and support for the Model Context Protocol. To top it all off, the company is introducing an enhanced version of its Outcome-Based Pricing model, one that supports double verification.

“We are entering the age of the autonomous service workforce,” Tom Eggemeier, Zendesk’s chief executive, said in a statement. “We believe the day is coming when every great business will be powered by specialized AI agents that don’t just work for you, but work with you.” He added that the agents will behave like real “team members,” with the same level of accountability—it’s a pitch that sounds familiar from nearly any enterprise AI vendor right now, such as Asana, Salesforce, and Microsoft.

Despite Eggemeier’s pronouncement, today’s announcements lean heavily on infrastructure upgrades and human-assist tooling. But that reading misses the point. Zendesk laid the foundation last year with its Resolution Platform—the unified layer that brings together data, intelligence, knowledge, workflows, and governance—and what’s being added now are the capabilities that make autonomy operational. A few targeted additions carry more strategic weight than the full product list implies.

The autonomous service workforce only works if agents can operate without handoffs to humans. This happens when there’s a loss of context. Every time a customer moves from chat to email to a phone call and has to re-explain their problem, that’s a failure point. To address this, Zendesk reveals that its AI agents now operate across messaging, email, and voice with shared context across interactions. Agent amnesia is an industry-wide problem, one that companies like Twilio have also moved to address. For Zendesk, this is made possible through its acquisition of Forethought in March.

Cross-channel continuity is only part of the equation. Zendesk is also expanding multilingual support for its Voice AI Agents. These bots can now converse in more than 60 languages and switch mid-conversation without losing context, a capability that could go a long way toward making support calls less dreadful. The expanded language support benefits customers directly, but it also keeps Zendesk competitive against a growing field of voice-native challengers like Sierra, Dialpad, Fin (formerly Intercom), Google, Twilio, and Salesforce—all moving aggressively to reshape how businesses talk to their customers.

Zendesk’s ambitions for an autonomous service workforce don’t stop at the customer. It also extends inward. The CX platform is furthering its march into territory traditionally dominated by vendors like ServiceNow and Moveworks. It’s introducing autonomous AI agents as part of its Employee Service offering, a platform introduced at last year’s Relate conference. Powered by its Unleash acquisition, these agents operate inside Slack and Microsoft Teams, performing functions similar to those of Atlassian’s Rovo and Glean, searching across enterprise systems to provide workers with answers to their queries based on the systems they’re authorized to access.

And for organizations wanting to build their own agents, Zendesk has introduced Agent Builder, a no-code tool for building, testing, deploying, and optimizing custom agents. The company said it’s designed to help “automate more complex front, middle, and back office service work while maintaining governance and oversight from a single control plane.” That said, it’s not the only agent builder that Zendesk has—it unveiled a similar offering last year as part of its Resolution Platform. The status of that one remains uncertain. In addition, for developers wanting greater flexibility, Zendesk is unveiling an MCP client and server.

“Our mission is to put the power to create this workforce into the hands of every enterprise, on one elegant platform,” Eggemeier said. “Whether those agents are crafted by Zendesk, by our partners, or by your own teams, they will all speak with one voice. We are not just giving you a tool; we are giving you the future of work.”

Here’s a look at other announcements from Zendesk Relate:

The Context Graph

Zendesk is debuting its Context Graph, an operational memory layer that captures past analyses, performance context, and agentic reasoning to improve future recommendations.

And speaking of graphs, Zendesk is expanding its Knowledge Graph by adding connectors for SharePoint, Google Drive, Notion, Guru, Contentful, and Document360.

Enhanced Copilot Experiences

Not everything Zendesk announced Tuesday is about replacing human effort. The company also unveiled a suite of Copilots, tools designed to work alongside human agents rather than supplant them. The Agent Copilot connects to internal and external sources to generate procedures and take action on at least 30 percent of tickets from day one. The Admin Copilot helps administrators identify operational issues, recommend fixes, and apply workflow changes in real-time.

For knowledge teams, the Knowledge Copilot surfaces gaps, outdated content, and inconsistencies based on actual customer conversations. And the Analyst Copilot helps teams spot trends and determine root causes through an agent-powered analytics experience.

Measuring What Matters

Rounding out the human-assist portfolio is Quality Score, a new continuous measurement feature that analyzes 100 percent of human and AI interactions to give teams an objective view of service quality and surface opportunities for improvement in real time. It’s available exclusively on Suite Professional plans and above.

Doubling Down on Outcome-Based Pricing

Introduced in August 2024, Zendesk’s outcome-based pricing charges customers only for resolutions that are verifiably resolved. Tuesday’s update reinforces that model with a double-verification process. Every resolution is first confirmed by the AI agent, then independently by a separate evaluation model. It’s an approach that echoes what Microsoft is doing with its Microsoft 365 Copilot, and a signal that verification standards could become a competitive battleground in enterprise AI pricing.

All of this is Zendesk making good on a promise. In December, Eggemeier told The AI Economy it was time for Zendesk to be the disruptor, not the disrupted. The company generated $200 million in AI annual recurring revenue last year and is targeting up to $500 million this year. The autonomous service workforce is how it gets there—lower costs and better outcomes for customers translates directly into higher retention and expanded spending for Zendesk.

Not everything announced at Relate is available today. Agent Builder, the Context Graph, the MCP Client, and the Analyst and Knowledge Copilots are in early access now. The updated Voice AI Agents reach general availability later this quarter, while the Employee Service agents and MCP Server open to early access this summer. The Admin Copilot and expanded Knowledge Graph connectors are generally available today. Quality Score has no confirmed timeline beyond a vague “coming soon to early access.”

This story originally appeared in the May 15 issue of The AI Economy on LinkedIn. Read the full issue here.

The AI infrastructure buildout is running headlong into a wall of public resistance. As tech giants race to expand compute capacity to power their models, a new Gallup survey finds that seven in 10 Americans oppose building AI data centers in their backyard. Among those opposed, half believed the data centers would negatively impact natural resources. Less than a quarter (22 percent) were worried about the impact on the quality of life, 20 percent were afraid of rising costs, and 16 percent thought pollution levels would rise.

American supporters of data centers cite the economic benefits (66 percent) that would be created, from job growth, tax revenue, and community development. Seventeen percent touted the technology benefits, while ten percent believed it would help them personally.

Every community is in the middle of this conversation, and the mere hint of a tech company acquiring land in a rural area is often enough to ignite fierce local opposition. The financial stakes are significant: according to Data Center Watch, local activism led to $64 billion in U.S. data center projects being blocked or delayed between May 2024 and March 2025. Yet the industry shows no sign of slowing, as Big Tech companies are reportedly on pace to spend nearly $700 billion on AI infrastructure this year alone. That gap between community resistance and corporate momentum is what makes this fight so hard to resolve.

One of the most consequential battlegrounds has been Prince William County in Virginia, where developers proposed the Prince William Digital Gateway—a 23-million-square-foot data center complex planned across 2,100 acres near Manassas National Battlefield Park. The project would have become the largest data center campus in the world. Residents and historic preservation organizations pushed back hard, but in December 2023, the county board voted 4-3 to rezone the land. Lawsuits followed, and the courts sided with opponents. The Virginia Court of Appeals ruled that the Board of Supervisors had sped up the approval process without giving the public a fair chance to weigh in. The decision led the county and one of the project’s developers, Compass Datacenters, to not appeal. However, the other developer, QTS, has petitioned the state’s Supreme Court to consider the case.

There have been similar situations happening in states like Pennsylvania, Maine, Louisiana, and most recently in Utah. The project in Box Elder County in the Beehive State is backed by “Shark Tank’s” Kevin O’Leary. The 40,000 9GW data center received approval from county commissioners this week, much to the dismay of opponents. In some cases, like in Saline Township in Michigan, local leaders deny permits, but somehow construction begins anyway. And even if data centers get built, that’s not the end of the story—lawsuits are being filed against companies over the dangers being inflicted on communities.

Politically, there’s bipartisan anger towards new data center development. That said, Gallup’s data shows that Democrats are more opposed than Republicans, 56 percent vs. 39 percent. Nearly half of independents (48 percent) are strongly against. Women are also more likely than men to register strong opposition (55 percent vs. 43 percent).

Gallup finds that Americans have adopted a “not in my backyard” (NIMBY) attitude to data centers. It will certainly continue to be a hotbed issue in the coming years, especially as AI usage grows. As opposition hardens on the ground, some in the industry are already looking up. Google and SpaceX are reportedly in talks to put data centers into orbit—a sign that the search for compute capacity with fewer neighbors may eventually leave Earth altogether.

This story originally appeared in the May 15 issue of The AI Economy on LinkedIn. Read the full issue here.

AI’s promise was that employees would stop doing the boring stuff. Software vendor after software vendor has touted their offerings as the key to freeing workers from repetitive tasks so they could focus on strategic tasks. New research from Workday released this week suggests most organizations forgot to tell their systems. While nearly all employees believe AI can improve their work, more than 8 in 10 report spending significant time as the glue between tools—moving information, reconciling conflicting data, and bridging systems that were never designed to talk to each other.

Workday’s research surveyed 6,100 HR, finance, IT, and operations professionals, all active AI users in their roles. Nearly all (97 percent) report having a positive day-to-day work experience, countering a prevailing narrative that workers are disengaged, distrust AI, or are resentful following layoffs. Still, despite this cheerful outlook, in reality, less than half are using AI to eliminate friction (43 percent) or reduce their manual work (46 percent)—a closer look at the data explains why.

The villain, as it turns out, is the computer systems themselves. Employees are frustrated because they’re forced to play traffic cop between teams, tools, and data. System incompatibility is what’s wasting significant portions of an employee’s day, which only “creates friction, slows decision-making, and contributes to unnecessary stress.” And it turns out to be a far-reaching problem: Just 27 percent of organizations have embedded AI directly into core workflows. That means the vast majority of companies are applying AI as an add-on, something that operates at the edge of business and doesn’t touch the workflow directly.

“Organizations have the ingredients they need—engaged employees, trusted systems, and advanced AI capabilities—for workforce transformation,” Workday wrote in its report. “Now is the time to close the gap and build a trusted, dependable, deterministic, AI-powered enterprise.”

Employees Are Tired of Gluing Work Together

So, what exactly are employees wasting their time on? The answer is mundane, repetitive, and almost entirely avoidable. They’re not struggling with complex strategic problems. Instead, they’re losing their days to the kind of administrative busywork that AI was supposed to eliminate. Workday’s research shows that 1 in 5 employees lose more than seven hours a week moving information and reconciling data. For IT professionals, it’s even worse with a quarter of them saying this kind of grunt work is basically what their job has become.

Looking at the data more closely, the majority of workers report spending significant time coordinating or translating work between teams and systems (82 percent), constantly shuffling information between platforms that don’t communicate with each other (81 percent), and reconciling conflicting data from different business systems (77 percent). Another 70 percent are either wrestling with administrative friction or re-entering the same information into multiple tools repeatedly.

Organizations are failing to address the root cause. Nearly two thirds of employees (61 percent) say core systems are too rigid to keep pace when change happens, and the human cost of that rigidity is real. Four in five identify having to navigate those same processes and systems as a direct source of workplace stress. In practical terms, every time a company reorganizes, launches a new product, or shifts strategy, employees are left manually bridging the gap between where the business is going and what its systems can actually support.

Employees want better ways to do work, including using AI and also having systems communicate with one another. Eighty-three percent report AI having improved their day-to-day work experience, while 61 percent attribute task completion time to the technology. And 54 percent claim AI has accelerated work in a productive way.

The problem isn’t that employees don’t want or aren’t using AI—we’ve known that for a while. It’s that they’re too buried in coordination work today to unlock its full value.

How Organizations Are Responding

While employees push for better system integrations, companies aren’t being bold enough to deliver it. “They’re not asking, ‘How can AI work with the systems that run the business, instead of alongside it?’” Workday said. The result is a workforce using AI within the constraints of the same broken infrastructure, patching gaps rather than closing them.

The barriers to why organizations aren’t prioritizing embedding AI in workflows vary but are consistent. Uneven skills, training, and access to AI tools top the list (27 percent), tied with too many approval and governance hurdles. Rigid systems and workflows follow at 23 percent, with unclear guidance on when to use AI and poor data quality each cited by 22 percent. Rounding out the list: murky accountability on AI use and AI outputs that conflict with existing policies (both 21 percent), and a lack of trust in AI altogether (18 percent).

It’s worth noting that several of the cited reasons—rigid systems, poor data quality, and AI outputs misaligning with existing rules—are exactly why organizations need to better system integrations in the first place. The consequences of inaction are already showing up in the data. More than two-thirds of employees (68 percent) say missing or unclear information has delayed decisions, and nearly as many report (64 percent) that teams regularly disagree over whose numbers are actually correct.

While Workday’s report focuses on outdated systems, it also highlights that management shoulders some of the blame for inaction. Having too many approval and governance hurdles, unclear guidance on when to use AI, and murky decision-making and AI accountability are not technological problems—they’re leadership ones. The data echoes reports from other companies that cite executive paralysis for sluggish enterprise AI adoption.

When companies do embed AI, they’re nearly twice more likely to report “meaningful” time savings than those that haven’t yet made the switch.

“AI can reduce friction, but it cannot compensate for fragmented systems, weak governance, or poor operating design on its own,” Workday writes in the report. “This is a call to action to organizations to evolve their architecture, integrating AI in their end-to-end processes.”

Workday makes a compelling case for why interconnected systems are essential in the AI era, but the report is notably vague on how organizations should get there. Modernizing core infrastructure isn’t a software decision—it’s an organizational one, requiring executive commitment, governance reform, and a willingness to retire systems that employees have worked around for years. Buying a new system is the easy part. Redesigning how an entire organization operates around it is another matter entirely.

For years, enterprises have struggled to govern AI where it actually runs, across clouds, tools, and now, increasingly, the desktops where employees do their work. ServiceNow thinks it has an answer, and it’s building the solution with Nvidia. Project Arc is the result: an enterprise desktop agent designed to bring autonomous capabilities seen in tools like OpenClaw into the corporate space, but with the security guardrails and governance that IT leaders actually require.

Announced at ServiceNow’s Knowledge 2026 conference, Project Arc is powered by Nvidia’s OpenShell, a sandbox environment where “everything is denied by default,” and IT administrators can define policies on what’s allowed. Installed on an employee’s desktop or laptop, this container hosts the autonomous agent, which has access to whatever the employee grants it and can take actions on their behalf. Everything is overseen and managed by ServiceNow’s AI Control Tower, while also grounded in the company’s Configuration Management Database (CMDB).

“Long-running, autonomous agents are rapidly changing the game for enterprise AI, and delivering them securely at scale requires governance that spans models, software, and AI infrastructure,” Kari Briski, Nvidia’s vice president of generative AI for enterprise, said in a statement.

Employees can think of Project Arc as similar to AI-powered IT support for handling mundane, routine tasks. The autonomous agent is capable of thinking, code writing, executing tasks, and adapting when things don’t go as planned. ServiceNow claims it can handle multi-step work across enterprise tools and systems without requiring prebuilt workflows.

“What we’re trying to help customers with ultimately is automation,” Joe Davis, ServiceNow’s executive vice president of AI engineering and delivery, tells The AI Economy in an interview.

Project Arc also works in tandem with ServiceNow’s AI specialists, prebuilt agents designed for specific business domains like IT and HR. These agents handle cloud-based requests by default—a Zoom access request, for instance, is really an entitlement issue with nothing to do with the desktop. But when a task requires direct access to an employee’s machine, such as restarting a browser or creating an email cache, Davis said that’s where Project Arc steps in. “That’s where Arc can come in because it can take action on a user’s machine.”

He acknowledges that IT-specific desktop agents could be next on ServiceNow’s roadmap. “There [are] a lot of settings and applications that need to be configured or changed to fix things—you’ve got to change some settings and restart,” Davis said. Some self-evolving desktop agent will be really good at that, and it will teach itself how to fix these things, and then we can share that across users.”

“The other thing that we’re looking at here is it’s pretty hard to set this stuff up today,” he went on to say. “There’s a lot of software that you have to grind through to get these working. We think that we can make this turnkey and really easy for enterprises to deploy safely.”

To be clear, Project Arc isn’t ServiceNow’s answer to Microsoft’s Copilot+ PC. The goal isn’t local AI processing—most tasks will run in the cloud, with sensitive work handled on-device when possible. ServiceNow isn’t asking companies to refresh their hardware either, though Davis acknowledged a practical tradeoff: older machines will lean heavily on the cloud, and organizations “have to be comfortable with what’s flowing back and forth.”

This isn’t new territory for ServiceNow. The company already runs robotic process automation agents that handle tasks across Windows applications and web agents that operate inside browsers. But the competitive landscape is intensifying. Last year, Microsoft upgraded its Edge browser to let developers integrate AI features directly into web apps, and the company is also exploring OpenClaw technology for its own Microsoft 365 Copilot ambitions. Atlassian acquired the startup behind Dia, an AI-powered browser, last September. And AI model makers—Anthropic, OpenAI, and Perplexity among them—have staked their own claims in the space.

ServiceNow isn’t backing down. “There’s a lot of experimentation that customers and companies want to do right now, so I don’t think there’s a winner-take-all in the AI space anytime soon,” Davis remarked. “We’re seeing companies that want to place multiple bets and see who wins and what happens in each of these categories.”

Widespread enterprise adoption of autonomous desktop agents is still nascent, but the window may be closing faster than expected. The emergence of OpenClaw and others in the so-called “Claw” family, including Nvidia’s own NemoClaw, has reset expectations for what these agents can do and how quickly they can do it.

The harder challenge isn’t the technology. It’s compliance. Davis acknowledged that heavily regulated industries, like financial services and pharmaceuticals, face a longer road, noting that “there’s going to be a number of customers we have to work with to get through some of these things.” ServiceNow sees AI Control Tower as its advantage here, a cross-platform governance layer that abstracts away the underlying operating system, with Nvidia’s partnership helping solve the trickier problem of edge connectivity to local devices.

Project Arc is currently available as an early preview. There is no timeline for when it’ll publicly launch.

Disclosure: I attended ServiceNow’s Knowledge 26 as a guest of the company, with my flights and hotel stay paid for. The AI Economy’s coverage is editorially independent from those that it covers. These words are my own.

In software development, the costliest mistake rarely lives in the code. It lives in the requirements that came before it. Typically, the issue arose because a developer interpreted the specs differently from the author—and that’s human-to-human. Imagine how much more challenging this can be when vibe-coding. That’s why Amazon Web Services (AWS) is updating its Kiro integrated development environment (IDE) to address this issue and ensure that the code generated is trustworthy.

“Addressing these bugs in your requirements is so important because the errors or incompleteness gaps, sort of logical inconsistencies, just multiply as they get missed throughout the software development life cycle,” Mike Miller, AWS’s director of AI product management, said to The AI Economy in an interview. “What we want to try to do is address…one of the most expensive classes of bugs at the cheapest possible moment.”

Trust Through Neurosymbolic AI

What AWS is doing goes well beyond helping developers write better spec documents. It’s about mathematically verifying that what they’re asking AI to build is actually buildable—before a single line of code is generated. The approach relies on Neurosymbolic AI, a technique that combines the language fluency of large language models with the provable certainty of formal mathematical logic.

The concept isn’t new—its roots trace back to the earliest days of AI research. But AWS has quietly been applying it for the past decade under what it calls automated reasoning. Miller explained that the company has used it internally to verify cryptographic algorithms, validate access control policies, and mathematically guarantee that an S3 bucket marked as private can never be reached from the public internet.

The integration with Kiro marks the first time AWS has brought that same rigor directly to developers as a hands-on, consumer-facing tool.

“We’re really excited about this Neurosymbolic technology because we see applications of it to a number of different capabilities, both in Kiro and across other user-facing tools at AWS,” Miller said. “We really see this as a key capability to help our customers achieve more trustworthy AI.”

How It Works With Kiro

Software engineers often treat spec documents, no matter whether they’re written by hand or using an LLM, as a first draft. But what happens when an LLM comes up against something ambiguously written? Vague or contradictory specs don’t just slow things down; they get implemented, and it becomes someone else’s problem down the road.

Neurosymbolic AI becomes visible in Kiro’s new Requirements Analysis feature. First, the LLM reviews the acceptance criteria and rewrites anything too vague to be testable—removing ambiguous language and tightening the level of detail. Next, those refined requirements are translated into a formal mathematical representation. Kiro samples multiple translations of the same requirement and looks for divergence. If the translations cluster around a single consistent interpretation, the requirement is unambiguous. However, if they scatter, that’s a flag—the LLM was guessing at the meaning, which means a human would too. Kiro identifies exactly where the interpretations split and serves up a plain-language question: here’s the ambiguity, which did you mean: A or B?

In the final phase, an automated reasoning engine analyzes the full set of requirements together, checking for contradictions between rules that looked fine in isolation, finding gaps where certain situations have no defined behavior, and flagging what Miller called “vacuous requirements”—rules that don’t actually constrain anything and would only produce unnecessary code.

AWS says roughly 60 percent of draft requirements across 35 internal Kiro projects required refinement before they were ready to generate valid code.

Other Kiro Updates

Requirements Analysis isn’t the only thing changing in Kiro. AWS is also shipping two updates aimed at speed:

Run Tasks in Parallel

With “Run all Tasks,” Kiro can run multiple tasks simultaneously rather than sequentially. When you kick off a spec, Kiro maps out which tasks depend on each other and which ones don’t—and runs the independent ones concurrently. Each task operates in its own isolated context, so there’s no interference between them—if one fails, the others keep going. AWS reports implementation times for large specs have decreased by 75 percent from over an hour to about 15 minutes in some cases.

Quick Plan for Specs

A specification workflow typically has three phases: requirements, design, and tasks—each needing approval before moving on to the next stage. This is useful when working through something unfamiliar. But when it comes to building well-documented features, the process can be laborious.

To move things along, AWS has introduced “Quick Plan” in Kiro. Developers are first asked clarifying questions about their app’s scope and constraints. Then the AI will generate all three phases in a single pass, resulting in a task list ready to be built. The underlying documentation is still generated and saved, but developers aren’t required to stop and approve each piece before the next phase begins.

The Investment in Being Correct

Today’s updates arrive alongside a leadership change inside AWS’s Automated Reasoning Group (ARG). Shawn Bice, Splunk’s former president of products and technology, has been tapped to lead the company’s investment in Neurosymbolic AI. AWS’s Vice President of Agentic AI, Swami Sivasubramanian, described him in an internal employee memo as someone who brings “decades of experience building and operating cloud services at massive scale, deep customer obsession, and a track record of attracting and developing world-class talent.”

His predecessor, Scott Wiltamuth, who has made significant contributions to ARG, will remain on the team focused on what AWS describes as its “highest-leverage technical problems.” The transition signals that AWS is treating automated reasoning not as a research curiosity but as a product priority.

When asked about future plans for Neurosymbolic AI beyond Kiro, Miller said he couldn’t speak about specific features but was direct about the direction. Requirements Analysis, he said, is “one step in our continued investment” in the space—one he expects to expand as agentic AI becomes more embedded in everyday development workflows.

For developers weighing what that means in practice, Miller’s message is straightforward: speed alone is no longer enough. “As AI writes more code, the human role should be shifting upstream,” he said. “The highest value work is going to come from defining what to build with precision, and not necessarily getting stuck on how to build it.”

Updated on 5/12/2026: Corrected Scott Wiltamuth’s role at ARG—he did not found the ground, that was Byron Cook, AWS’s vice president and distinguished scientist. However, Wiltamuth has played an oversized role in scaling the team.

Enterprises have spent the past year building AI agents. Now they have to figure out what to do with them. To help, Glean has released its Enterprise Agent Development Lifecycle (ADLC), a framework that gives Chief Information Officers and IT leaders a repeatable path to scale agents across their organizations.

“Agents are software. They need a disciplined way to be defined, built, launched, governed, and improved over time,” Emrecan Dogan, Glean’s chief product officer, said in a statement. The ADLC delivers that, giving CIOs a structured approach for ensuring agents are useful, secure, and tied to measurable business outcomes.

The framework draws from Glean’s own experience. After releasing autonomous agents last December, the company found itself facing the same problem it now wants to help customers solve: thousands of agents built across teams, with no consistent way to govern or measure them. “It’s really important for us to define what success looks like, what do we want AI agents to be able to do, and create a product that allows our customers, as well as ourselves, to be able to operationalize the agents that are built,” said Selene Kim, a Glean product manager, in an interview with The AI Economy.

Glean’s Enterprise Agent Development Lifecycle features seven stages: Opportunity, Design, Performance, Input, Develop, Launch, and Monitor and Improve. “Just like the software development lifecycle, you need every single step of the flow,” Kim pointed out. It’s intentionally platform-agnostic—Kim described it as Glean’s “opinion piece” for the broader AI community, something any organization can adopt.

But Kim says the harder problem isn’t building—it’s what comes after. “The problem that we’re solving with the Agent Development Lifecycle is not just about building the agents,” she explained. “How do we make sure that those agents that are built…can you get it to work in production? How do you measure whether it is successful? How do you know when to say it’s not working as expected?”

That said, Glean acknowledged the framework is freely adoptable—but argues its advantage lies not only in its context engine but also in the platform capabilities it has built to execute every stage of the ADLC—all of which competitors would take time to replicate.

Alongside its ADLC announcement, Glean is introducing eight capabilities that address parts of the framework where most enterprises are stuck: building agents with the right context, launching them with the right governance, and measuring the value they generate over time. “They’re mapped to each part of the Agent Development Lifecycle because we know that there are many things that need to be done to have a perfect agent development life cycle,” Kim shared.

Here’s a look at the new tools and which phase they fall under:

• Auto Mode Agent Builder (Develop): Using natural language, this editor generates a Glean-powered agent capable of planning, reasoning, and executing across the enterprise graph without the need for pre-defined workflows or manual configuration.

• Debug and Trace Views (Develop): This provides full visibility into every action taken by the AI agent, cataloging inputs, tool calls, LLM decisions, and outputs. No longer must developers guess why their agents have failed.

• Sub-Agents (Develop): These are specialized agents focused on discrete tasks, coordinated by a parent agent at runtime.

• Expanded Agent Sandbox: Glean customers now have access to a secure file system and code execution within their virtual private cloud. In addition, it supports adding both apps and individual actions.

• Content and Scheduled Triggers (Context): This empowers agents to automatically react to content changes, scheduled runs, forms, and external events, enabling them to operate more within existing business processes.

• New Agent Library Controls (Launch): Glean has added verification badges, featured agents, departmental categories, and a soft-delete with an admin restore—all aimed at turning this marketplace into the “governed front door for agent distribution.”

• Agent Access Policies (Launch): AI agents now have consistent controls, such as blocking or flagging sensitive material, thanks to organization-wide guardrails.

• Updated Agent Insights Dashboard (Monitor and Improve): Glean has rebuilt this monitoring tool so it better tracks adoption, top use cases, estimated hours saved, and feedback trends over time. In doing so, CIOs will better understand which agents are delivering value, and which ones aren’t.

Glean revealed that Auto Mode Agent Builder, Debug and Trace Views, the Sub-Agents, Agent Sandbox, and the new Agent Library Controls are all generally available today. However, the Content and Scheduled Triggers and Agent Access Policies are in beta, and the Agent Insights Dashboard is coming soon.

Enterprises that have gone the build-your-own route with agentic AI—picking their own models and standing up their own agent infrastructure on platforms like Amazon Bedrock AgentCore—have had to figure out governance themselves. ServiceNow is offering them a shortcut. The company on Wednesday connected its AI Control Tower with AgentCore, giving customers a single layer to govern agents they’ve built across AWS, regardless of which models are running underneath.

ServiceNow is also pairing its AI Specialists with AWS agents to handle workflows across security, IT operations, and telecommunications, with humans approving key decisions. Separately, the company is bringing its SDK into Kiro, AWS’s agentic IDE, so developers can build ServiceNow applications and AI agents without leaving the editor.

“Organizations aren’t experimenting with AI anymore, they’re operationalizing it,” Chris Grusz, AWS’s managing director of technology partnerships, says in a statement. “ServiceNow and AWS are delivering the architecture to make that real: unified governance, trusted infrastructure, and developer tools that take AI from idea to impact…We’re making it easier than ever to deploy, govern, and scale AI agents across the enterprise.”

ServiceNow’s AI Control Tower debuted at Knowledge 2025 as a single pane of glass for the AI agents and models running across an enterprise. Since then, the company has added new capabilities for discovering, observing, governing, securing, and measuring those deployments. For enterprises building agents on AWS, the Bedrock AgentCore integration provides a consolidated offering rather than piecemeal assembly.

According to Joe Davis, ServiceNow’s executive vice president for platform engineering and AI, enterprise customers don’t put all their eggs in one basket—they have solutions spread across multiple hyperscalers and products. But in order to really secure all of the organization’s AI assets, it’s critical to first take inventory. “You need to have visibility of all of it to make sure that you can contain exposure,” he says in an interview with The AI Economy. “In order to do that, you have to be cross-platform, neutral, and integrate better.”

It’s why ServiceNow made AI Control Tower extensible, integrating it with Amazon Bedrock APIs to identify AgentCore-created agents. Once they’ve been discovered, the AI Control Tower can begin securing and governing them.

AI Specialists 🤝 AWS AI Agents

When it comes to AI Specialists, ServiceNow is connecting them with AWS’s AI agents so they can work more closely together. It’s a recognition that both platforms have limits in what their tools can do, so to help workers complete tasks, ServiceNow and AWS have worked on a way to ensure seamless system handoffs to minimize risk and reduce delays.

“It’s like an MCP and A2A integration, where there are parts of the workflow that our AI Specialists handle, and then they will work with other agentic systems, like AgentCore, to do some of those domain-specific tasks that maybe…our AI Specialist doesn’t do,” Davis says.

How this works is when an AWS service detects an issue, ServiceNow’s AI Specialist will orchestrate the response, a human approves the fix, and the workflow closes with a full audit trail. In security, a configuration change in the configuration management database (CMDB) would trigger ServiceNow’s Vulnerability Resolution AI Specialist to call the AWS Security Agent via MCP to run an on-demand penetration test. In doing so, it would include identity exposure data from ServiceNow’s Veza and Armis device data before remediation.

For IT operations, an anomaly in CloudWatch would be routed to ServiceNow’s AIOps and Site Reliability Engineering Specialists, which works with the AWS DevOps Agent to correlate events and execute fixes.

“It’s kind of how people work together,” Davis explained. “I have some general knowledge, but then there are subject matter experts that I might want to go ask for help. It’s kind of the same thing from an AI standpoint.”

Build ServiceNow Agents in Kiro

ServiceNow is also bringing its SDK into Kiro, AWS’s agentic IDE, where developers can vibe-code ServiceNow AI agents without switching tools. The company’s SDK with Build Agent skills can now be installed from the Kiro Power Marketplace with a single click. The promise is that developers will be able to “scaffold” applications, configure workflows, and create agents through natural-language prompting, without leaving their IDE.

The expanded integration lands as ServiceNow and AWS disclosed a commercial milestone: customers have now transacted more than $1 billion through the partnership on AWS Marketplace. ServiceNow described it as something more significant than “a commercial threshold,” framing it instead as a shift in the enterprise to consolidate infrastructure around platforms they trust. ServiceNow is positioning the billion dollars as market validation of the thesis that organizations prefer a layered tech stack over point solutions or building it all themselves.

Davis highlighted that ServiceNow is focused on delivering turnkey, outcome-oriented solutions—prebuilt AI that plugs directly into real enterprise workflows with permissioning, security, and governance baked in. Doing it yourself is “extremely complicated, and there are so many things that can go wrong,” he pointed out.

ServiceNow AI Control Tower with Amazon Bedrock AgentCore is available today in the AWS Marketplace. The company’s Vulnerability Resolution AI Specialist, AIOps AI Specialist, and Site Reliability Engineering AI Specialist integrations are expected to be available later this year. Finally, the ServiceNow SDK for Kiro is available now in the Kiro Power Marketplace.

Disclosure: I attended ServiceNow’s Knowledge 26 as a guest of the company, with my flights and hotel stay paid for. The AI Economy’s coverage is editorially independent from those that it covers. These words are my own.