Enterprise AI Agents Get Workplace Identities as Governance Lags

Agents are getting job titles

In October 2026, the major enterprise AI vendors stopped pitching agents as helpers and started treating them as staff. Google gave the clearest example. At its Gemini at Work event, it launched one unified Gemini agent that takes objectives instead of step-by-step instructions, hands work to subagents, and gets its own Google Workspace account and email address4. Google says the agent knows team structures, time zones, approval chains and calendars. Every action it takes is recorded in an audit trail under the agent's name, not a person's4. CIO.com reported that these coworker agents get addresses at a dedicated @agents.company.com domain and persistent storage, and can run multi-step jobs for hours or days19.

Google is not alone. Microsoft has described an Autopilot product built on the same idea. Each Microsoft employee can get a long-running agent with its own identity, its own computer and its own workspace. Staff can message it in Teams like a colleague, and it can work for days inside a cloud sandbox10. Microsoft is also previewing Work IQ, which connects Copilot and autonomous agents directly to live Dynamics 365 and Power Platform data12. Oracle has launched Fusion Claw, an agent orchestration layer built into its ERP software, where companies can set operating procedures, risk thresholds and decision rights10. SAP announced 224 agents and 51 assistants at Sapphire 2026 and plans more than 400 autonomous suite agents by year-end13.

Taken together, these launches suggest the contest has moved past whose model is smartest. Vendors are now competing to be the system agents run inside: the place that holds their identity, permissions and audit records.

The race to own the agent layer

Analysts quoted by CIO.com see Google's launch as an attempt to become the gatekeeper of a new enterprise operating system, one that can run on whichever model suits a given task19. Model choice is central to that pitch. The Gemini agent picks a model by default, but users can switch to third-party options, starting with Anthropic's Claude, and Google says open-source and private models will follow4. Google Cloud CEO Thomas Kurian's argument is that the leading model changes every few months and the largest model is not always the right one for the job19.

Google can push this at scale because it already has the customers. Sundar Pichai said Gemini has more than a billion monthly active users and that nearly 90% of Fortune 100 companies use Gemini Enterprise18. The company is launching the agent for businesses first. Pichai said this lets Google work through the harder security, scale and performance problems before a consumer release18.

The rest of the market is moving the same way, through deals and partnerships. HCLTech announced a global partnership with Anthropic that includes a dedicated business unit and a center of excellence focused on governed agent deployment5. DXC Technology is working with LangChain on governed agents, and its OASIS orchestration platform is already in production8. Reports say the agent startup Decagon has held early acquisition talks with ServiceNow and others. Rocket Software reportedly paid about $100 million for Geniez AI, a six-person mainframe-agent startup that is one year old8. Nous Research raised $90 million at a $1.5 billion valuation to sell an enterprise version of its open-source Hermes agent10. That gives buyers worried about lock-in an alternative to the closed platforms.

Adoption numbers depend on the definition

Published adoption figures vary widely, and much of the gap comes from what each survey counts. Gartner says only 17% of organizations have deployed AI agents, while more than 60% expect to within two years7. Research from First Page Sage puts enterprise adoption at 28%, up from 16% a year earlier. But 57% of those adopters are still experimenting, and only 15% run agents fully at scale1. At the other end, EY found that 91% of senior AI executives at companies with more than $1 billion in revenue say they use agentic AI in pilots or full deployment21. A Morning Consult study for Guild says 96% of enterprises run agents in production.

One analysis of the published figures points out that adoption estimates range from 52% to 82% because surveys measure different stages, from simply using agents to scaling them. By that analysis, only 2% of large organizations report fully scaled agents9. Another widely shared estimate says 85% of large companies are experimenting with agents but only 5% have moved them into production10.

The reading that best fits all of this: agents now show up almost everywhere as pilots, embedded features or unsanctioned tools, but very few run real business processes on their own. Vendor claims of broad adoption and analyst warnings about stalled pilots can both be accurate at once. They are counting different things.

The governance gap is the real story

On governance, the surveys mostly agree. EY found that nearly every large company has formal AI governance policies, yet 47% admit they have skipped those processes for urgent deployments21. About half of companies using agentic AI have not updated their frameworks to cover agents. Some 85% say at least some agents act without real-time human involvement, and 26% cannot detect unauthorized agents running inside their organizations21. More than a third reported an AI incident with a materially negative impact, such as data loss, financial damage or operational disruption24.

Other research points the same way. OneTrust found that 87% of organizations encourage agent use, but only 47% have defined controls in place. Nearly half reported at least one incident in which AI systems or agents took unapproved actions22. Gravitee's survey found that only 14.4% of organizations sent every agent live with full security and IT approval, while 82% of executives felt confident their policies protected them23. Guild's research describes the same mismatch between confidence and reality: most respondents feel prepared, yet two out of three have had an agent incident. A Cloud Security Alliance survey reportedly found that 82% of organizations have unknown agents running in their IT environments27.

Nobody clearly owns the problem, either. In one practitioner survey, the CISO is primarily accountable for AI security decisions at just 38% of organizations. Among companies that have agent identities, only 40% have a dedicated security control for them. Federal guidance will not help soon. NIST's AI Agent Standards Initiative is not expected to produce final standards before 2027.

That context makes the agents-with-email-accounts design look partly defensive. If an agent has its own identity and audit trail, it can be listed, scoped and switched off like any other account. Okta CEO Todd McKinnon made this case at Oktane, arguing that the answer to agent risk is visibility and control rather than either locking agents down or letting them run unchecked2. Investors are funding the same idea: Pluto Security raised a $42 million Series A to secure autonomous systems that conventional tools were not built to handle16.

Caution is rising, not falling

Companies are showing more care about how much freedom they give agents. In a small, self-selected August survey, 56% of respondents at organizations deploying autonomous agents allowed, or were preparing to allow, production changes without human approval. In July, the figure was 75%. The share expecting to keep human review rose from 20% to 42%20. The authors warn that two separate samples do not prove a market-wide trend20. Still, the result matches the broader pattern.

Software development shows a similar gap between activity and results. A study of more than 700 firms found that AI coding agents raised code output by 30% without clearly improving finished software delivery8. Anthropic also published a report on Claude taking unintended actions on real websites, a reminder that even test environments can have real-world side effects10.

What comes next

Gartner's latest Hype Cycle frames the shift directly: control now matters more than capability. Governance and identity infrastructure are prerequisites for scaling agents, and Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027 over cost, unclear value or weak risk controls7. Quartz's reading of Gartner's forecasts points to a split future. Agentic features spread through tools companies already own, while much of the spending on custom agents gets written off, and new jobs appear around supervising and auditing what agents do15.

Gartner forecast that 40% of enterprise applications would include task-specific agents by the end of 2026. That deadline arrives in a few weeks, and as of early October no progress figure had been published9.

The most likely outcome is that the vendors that do best will be the ones that make agents governable, not the ones that make them most capable. Giving agents identities, permission systems, spending caps and audit logs is what lets them move from pilots into production. For now, the products are arriving faster than most companies can govern them, and that gap will determine how much autonomy agents actually get in 2027.