UBS now reports its agents alongside its profits. In its third-quarter 2026 results, covered by Fintech News Switzerland on November 2, 2026, the bank said employees run more than 12,700 AI personal agents. It also counted 608 AI use cases in production, 81% more than a year earlier, and more than 1,090 applications in development. The filing for the same quarter carries a different kind of count. UBS had 97,593 internal staff on September 30, 2026, a net decrease of 1,492 from June 30 and 6,834 from a year earlier.
Put those two counts side by side and a small, odd problem appears. People leave. The things they built keep running.
Most advice on designing an AI operating model starts with structure. You get a central team, a chief AI officer, a hub and some spokes, a review board that meets monthly. I think the design should start somewhere less flattering, at the exit. An operating model is a set of standing answers about who decides, who pays and who answers when something goes wrong. Those answers face their hardest test on the day the person holding them hands in a badge. An agent built by one employee, tuned to that employee's habits and running on that employee's say-so, can lose its boss and never find out.
Twenty-One to One
For illustration, divide 12,700 by 608 and you get about 21 personal agents for every production use case. The two figures measure different things, so the ratio proves nothing on its own. It does show the shape of what is happening inside large companies. There is a small population of AI systems that went through a formal process and a much larger population that a person simply made.
The second population is the interesting one, and it is the one we know least about. The coverage of the UBS figure gives no detail on what those agents do, what they touch or who reviews them. I have no reason to think UBS has mishandled any of it. A regulated bank that publishes the number may well have a tidy answer behind it. UBS is useful here for a simpler reason. It said the number out loud, and most companies with the same split have never counted.
The nearest thing to a market-wide reading comes from the Cloud Security Alliance, which surveyed 418 IT and security professionals in January 2026. In that survey, 82% said they have unknown AI agents in their environments, and 21% said they have formal processes for decommissioning agents. Token Security, which sells in this market, commissioned and paid for the work, so discount it as you see fit. Even after a heavy discount, the gap stands. Companies have learned to start agents far faster than they have learned to end them.
Attrition Is the Number Missing From the Deck
Every AI adoption dashboard I have seen counts upward. Agents built, users active, use cases live. A count of people leaving sits in a different system, owned by a different executive, and reviewed in a different meeting. The two numbers rarely meet.
Suppose UBS's personal agents sit evenly across its staff, which nobody outside the bank can check, and a net drop of 1,492 from a base of about 99,085 would strand about 1.5 percent of them, or roughly 190 agents in one quarter. Suppose the same even spread over twelve months, and a net drop of 6,834 from about 104,427 would strand roughly 830. Those are rough figures, and they lean low. Net numbers hide the leavers who were replaced by hires, and they miss every person who changed roles and stayed. A mover is the harder case, because the person is still around and the reason the agent had its access has gone.
I am claiming no link between agents and the falling headcount at UBS. The filing says people left and is silent on why. The point runs the other way. Whatever your reasons for changing your workforce, each change now carries a second, invisible transaction, in which a piece of working software loses the human who understood it.
This matters more as executives get ambitious. Gartner's 2027 CEO Survey, published in October 2026 and conducted in April and May 2026 among 203 global CEOs and senior business executives, found that 63% named radically redesigning the operating model as a top move for speeding up financial returns from AI through 2027. In the same survey, 52% named aggressively deploying AI agents as employees, and 47% named freeing up or reallocating capital through significant restructuring. Meanwhile, 44% of CEOs struggle to show AI-driven financial returns to stakeholders. Gartner's David Furlonger put it this way: "The AI value problem will not be solved by simply adding more AI to existing ways of working."
Read those three moves as one plan and you see the collision. Redesign the model, put agents on the staff and restructure the people. A company doing all three is moving humans out from under agents at the very moment it is giving those agents more to do.
Two Vendors, Two Answers
The clearest evidence that this has become a design problem is that the big platforms have started writing it into their products, with very different levels of finish.
Google gave the agent a desk
On October 8, 2026, at Gemini at Work, Google introduced coworker agents, as Rajesh Beri reported that day. A coworker agent receives its own Workspace account, with an address in the form @agents.company.com, and "lives in your directory like an employee." Its actions are attributed to the agent instead of a person. That is a clean idea with real value for anyone who has tried to read an audit log.
Beri also found the blank spot in the launch. "Nothing says an agent must name a human sponsor, or what happens to the account when its creator leaves." Google, he noted, "published no price, plan list or GA date" either, so some of this may be filled in before customers have the product. As of the launch, though, the agent had a mailbox and a directory listing, and the public account of it left out what happens to its supervisor.
Microsoft wrote the succession rule
Microsoft has gone further, and its documentation reads like a first draft of corporate law for software. In Entra Agent ID, owners are technical administrators and are optional. Sponsors are "business representatives accountable for the agent's purpose and lifecycle decisions," and at least one is required. Then comes the line that should catch an operator's eye. "For unpublished agents, creators often serve as sponsors." If a user creates an agent and names nobody else, "the calling user automatically becomes the sponsor."
So for personal agents, the accountable business representative is usually the person who made it for their own convenience. That works until that person goes. Microsoft's Zero Trust workshop guidance describes what happens next when no lifecycle workflow is set up. "A departing sponsor's agent identities silently transfer to their manager through automatic succession," it says, "but no one is alerted that the transfer occurred." It continues: "The agent continues to operate with whatever permissions it had." The same guidance adds that "no human actively knows they are now responsible for its lifecycle and access decisions."
I find that passage remarkable. The vendor's own guidance describes a state in which the system of record shows a responsible human and the human has no idea. Accountability exists on paper and nowhere else. Microsoft offers workflows that fix the silence by notifying managers and cosponsors, and they are worth switching on. They fix the notification. The manager still needs to know what to do with the email.
What the Heir Actually Receives
Say a manager has eight reports, each running three agents, and two of those reports leave in the same restructuring, which hands the manager six agents she never asked for and has never seen. Her new duties, on Microsoft's definition, cover renewal, extension and removal based on business need. She has to judge the business need of software she did not write, built around a colleague's private shortcuts, in the same weeks she is covering that colleague's actual work.
Her powers are strange too. Sponsors can disable an agent, but under the same Microsoft documentation, "Sponsors can't re-enable or restore agent blueprints or identities." If she does the sensible thing and switches an agent off to see what breaks, she needs a resource owner or an admin to bring it back. Owners are optional, and on a personal agent the owner, if there ever was one, may be the person who just left. The cheapest diagnostic in all of operations, turning the thing off to see who complains, is a one-way door for the person most likely to need it.
The estate she inherits also includes a diary. Microsoft's Copilot Studio documentation uses the plain word for all this, "orphaned agents," and warns that "there might not be a responsible party to manage the agent's settings and compliance requirements." It also explains that once an admin reassigns the agent, the old owner loses access, and if the new owner holds the right privilege, "they can access all the existing transcripts of the agent." Every question a departed colleague asked that agent, every draft and half-formed worry, can pass to whoever takes over. Few employees building a helpful little assistant in their second week imagined their boss reading the transcript in their last.
Finance teams have lived a mild version of this for decades, with the spreadsheet nobody dares to touch after its author retires. The agent version is worse in one specific way. A forgotten spreadsheet waits to be opened. A forgotten agent with a schedule and a set of credentials goes on acting, and it keeps doing so under assumptions that were true on the day its maker last paid attention.
The Accidental Process Map
There is a brighter side to this, and it is where the money is.
Companies have spent years trying to write down how work really gets done. The official process says one thing. The person who has done the job for eleven years does something else, and the difference between those two is often the reason the work comes out right. Nobody ever captured that difference, because nobody had a reason to write it down.
Personal agents give people a reason. To make an agent useful, an employee has to tell it the exceptions, which client wants what, which report is always late, which number to distrust. I think those instructions are the most honest process documentation most companies will ever own, produced by accident, by the people who know, for their own selfish benefit. That is why they are good.
So the exit is a risk event and a discovery event at once. When someone leaves, you can see exactly what that person found worth automating. You can see which of their agents other people came to depend on. You can see, in plain language, what a veteran knew. A company that deletes all of it loses knowledge it never knew it had. A company that keeps all of it unread keeps paying for something it does not understand, and carries the exposure too.
The ownership of that material deserves a lawyer's attention before your first senior departure, in your own jurisdiction, under your own employment agreements. Ask whether those agreements reach an agent's instructions, memory and skills, and ask what happens if the agent was built in a personal account. You want that answer on file before a departing rainmaker's counsel asks for it.
Designing an AI Operating Model From the Exit Backwards
Start with the leaver and work back to the day of creation. I would make six decisions, in this order. They read as policy, and they are the operating model, because they settle who decides, who pays and who answers.
Personal agents die with the job
Make expiry the default. When an employee leaves or changes role, their personal agents stop, and the agent's access stops with them. This will feel harsh to anyone who built something clever. It is the only default that scales, because it means inaction produces a safe result. Silent succession does the reverse, since inaction leaves an agent running with its old permissions.
Promotion becomes the intake gate
Anything that must outlive its maker has to be promoted, deliberately, from personal to team property. Promotion is where the review happens. This turns the 21 to one shape from a governance headache into a funnel. The large population of personal agents acts as your cheapest research lab, and the small population in production is what survived selection. Central teams stop trying to police everything people build and concentrate on the narrow moment when something personal becomes something the company relies on.
Sponsors are seats
A promoted agent answers to a role or a group, and a named human holds that seat for now. Entra already allows this, since groups can be sponsors, and when one is assigned, all its members hold sponsor rights. Groups can't be owners, though service principals can, so the technical side needs its own plan. Treat a personal name in the sponsor field of a promoted agent as a defect.
Split the estate in three
A handover moves three different things, and each deserves its own rule. Authority covers what the agent may touch, and it should reset to least privilege and be requested afresh by the new sponsor. Know-how covers the instructions and skills, and it should pass to the team. Transcripts cover the person's working conversations, and they should stay sealed by default, opened only for a stated reason. Tell employees this on the day they get access to the builder, because the rule changes what they will be willing to write.
HR holds the trigger
The leaver event lives in the HR system, so the agent process has to start there. Add one line to the offboarding checklist and one to the role-change checklist, each asking for a disposition of every agent the person sponsors, with a named reviewer. Fold the same step into severance planning and restructuring models. If your CFO has a spreadsheet of positions leaving next quarter, someone should be able to pull the list of agents attached to those positions before the announcement goes out.
Count dispositions
Add one number to the dashboard that counts upward. Report the share of leavers and movers whose agents had a recorded decision by their last day in the role. Report the count of agents whose sponsor has changed in the last 90 days and who have had no human review since. Those two figures tell a board more about the condition of an AI operating model than any count of agents built.
Run the Drill Before the Reorg
You can test all of this inside a quarter, and you do not need a platform purchase to do it. Say you pick one team of 20 people in a function that matters, and ask its most agent-heavy member to take a planned two-week absence with all their agents paused. Then watch. Write down who complains, which reports fail to arrive and which process quietly stops. Ask the manager to name every agent on the team, its purpose and its access, with no help from IT.
Most leadership teams will learn something uncomfortable from a morning of that. The drill costs a little disruption. The live version arrives during a restructuring, when the people who could explain the agents have signed their paperwork and the people who remain are too stretched to look.
Pricing tells you who the vendors think is responsible. Microsoft says "agents do not require their own licenses," with the license sitting with the owner, sponsor or manager, and Agent 365 at $15 per user per month standalone, per Beri's account. The agent's cost and its accountability both hang from a human. Your operating model has to say what happens when that human goes, because the invoice will keep arriving either way.
Architect the Exit
You can buy the sponsor field. You can buy the workflow, the registry and the notification email, and you should. No vendor can tell you which of your people's agents deserve to survive them, who in your business should answer for what those agents do, what a departing employee is owed in privacy, or how any of that fits your next restructuring. Those are decisions about how your company runs, and a company that leaves them to product defaults has let a settings page design its succession plan.
Act before your next reorganization. The agents are multiplying now, in your company, whether or not anyone has counted them, and a chunk of the people who built them will be gone by next year. Each departure without a plan leaves a working system with credentials, a stale purpose and a nominal boss who was never told. Each departure with a plan gives you a process map you could never have commissioned. That gap compounds, and it decides whether your AI operating model is a slide or a thing that holds up under pressure.
Agor AI Advisory works with leadership teams on this architecture, joining HR, risk, finance and technology around one set of decisions, then testing those decisions against the exits you already know are coming. Design it while the people who built your agents can still explain them.
Sources
- UBS Grows Live AI Use Cases to 608 as Staff Run 12,700 Agents, Fintech News Switzerland, November 2, 2026
- UBS Group AG Form 6-K, third quarter 2026, SEC
- Gartner Says 44% of CEOs Struggle to Prove AI Returns, TechEdgeAI, October 2026
- Google's Gemini Agent Gets Its Own Mailbox Before It Gets a Price, Rajesh Beri, October 8, 2026
- Administrative relationships in Microsoft Entra Agent ID, Microsoft Learn, 2026
- Zero Trust workshop guidance AI_014, Microsoft
- Use the Power Platform API to reassign ownership of orphaned agents, Microsoft Learn, 2026
- New Cloud Security Alliance Survey Reveals 82% of Enterprises Have Unknown AI Agents, April 21, 2026