UBS Staff Now Run 12,700 Personal AI Agents. What Agent Work Looks Like Inside a Bank, and What UBS Doesn’t Say

By Thorsten Meyer | November 1, 2026 | ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Most companies still talk about AI agents as a plan. On 28 October, one of the most tightly regulated banks in the world reported them as a number. In its third-quarter results, UBS wrote that its employees “are running over 12,700 AI personal agents to support them in their daily activities.” In the same paragraph: more than 18,000 software developers working with AI tools, 608 AI applications live in production, and more than 1,090 more in development.

The sentence sits on page three of a 16-page results release that is mostly about profit, capital rules and the end of the Credit Suisse integration. On the analyst call that morning, the CEO gave AI two sentences, and no analyst asked about it. That is exactly why it is worth a closer look. Here is what UBS actually reported, what an “agent” at a bank can and cannot do today, what the bank does not disclose, and what it means for companies and for the people who work in them.

Key takeaways

• The numbers are UBS’s own. In its 3Q26 release of 28 October 2026, UBS reports over 12,700 personal AI agents run by employees, over 18,000 developers using AI tools, 608 live AI use cases (up 81% on the year) and 1,090+ in development. Reuters and PYMNTS reported the same figures.
• It grew fast. Three months earlier, on 29 July, UBS said “all UBS employees now have the ability to create their own AI agents.” Live use cases rose from 560 to 608 in one quarter; applications in development from 920+ to 1,090+.
• The headcount drop is the integration, per UBS. Internal staff fell by about 1,500 full-time equivalents in the quarter, to 97,593. UBS and Reuters tie it to the Credit Suisse integration. UBS does not link it to AI, in either direction.
• The bank keeps outside agents out. A day before the results, CEO Sergio Ermotti said UBS is “not allowing models to enter into the bank” to act for clients, as reported by Reuters.
• What’s missing is governance detail. UBS does not say what a “personal agent” can access, how many people run them, how they are reviewed, or how many of the 608 use cases face clients. The Swiss regulator FINMA expects banks to keep a complete AI inventory with risk classes, and says that has become harder since “applications are accessible to everyone.”
• For businesses, this is the template question of 2027: not “should staff build agents?” but “who checks 12,700 of them?”

Products Worth Considering

Apple Watch Series 12 GPS Smartwatch, 46 mm

AmazonView Latest Price

Bose QuietComfort Wireless Earbuds with Active Noise Cancellation, Dewdrop Mint

AmazonView Latest Price

Bose SoundLink Max Bluetooth Speaker

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

What UBS reported on 28 October

The paragraph comes under the heading “Strategically investing in our franchise and footprint.” Here is what it says, figure by figure, all from the UBS release:

• 9 “large-scale, end-to-end transformational AI initiatives.”
• 608 live AI use cases in production, an 81% increase on a year earlier.
• 1,090+ further AI applications in development.
• 12,700+ “AI personal agents” run by employees “to support them in their daily activities.”
• 18,000+ software developers “leveraging AI tools in the software development cycles, leading to measurable productivity gains.” The release does not put a number on those gains.
• 30 million+ insights generated so far this year by STAAT, a tool that feeds suggestions to the bank’s US financial advisor teams.

UBS also points to the 2026 Evident AI Index, an annual ranking of 50 large banks, where it placed sixth overall and as the only European bank in the top ten. UBS says it was recognised as the global leader for talent development and responsible AI. JPMorganChase came first for the fifth year in a row.

The figures check out against a second source: Reuters, in its results story from Zurich on 28 October, and PYMNTS the next day both report the 12,700 agents. Neither adds detail beyond the release.

So what is a “personal agent” at a bank?

UBS does not define the term in the release. But its earlier statements give a picture. In its second-quarter release of 29 July, the bank wrote that “all UBS employees now have the ability to create their own AI agents,” which it said would accelerate “our ability to deploy AI safely and at scale across the firm while enabling our people to spend more time on judgement, collaboration, and delivering value for clients.” Its technology page describes an in-house assistant called Red, rolled out to around 100,000 employees, and an AI-powered HR assistant that lets staff book holidays and look up policies in Microsoft Teams.

So a “personal agent” is most likely a small helper an employee sets up for their own work: one that drafts a summary of a long document, prepares a client meeting from internal notes, sorts a mailbox, or runs a fixed checklist every morning. Think of it less as a robot employee and more as a well-briefed intern who sits on one person’s desk. That is my reading, not UBS’s definition. The bank does not say which tools the agents are built on, what systems and data they can reach, or whether any of them can take actions on their own rather than only prepare work for a human.

That last point is the one that matters. An agent that writes a draft is a productivity tool. An agent that sends, books, moves or approves something is a new kind of risk.

The line UBS draws: no outside agents acting for clients

Here UBS has been unusually clear, just not in the results. Speaking at an event at ETH Zurich on 27 October, Ermotti said the bank is not allowing outside AI models to act on behalf of clients, for example to manage accounts or move deposits. “We are not allowing models to enter into the bank and do that,” he said, as reported by Reuters. The priority, he said, is to “make sure that we don’t get AI models that drive outcomes for clients in a black box.”

Put the two statements side by side and you get the shape of agent work in a regulated bank in late 2026:

• Inside, for employees: many agents, built by staff, supporting people “in their daily activities.”
• At the client’s door: no third-party AI agents acting on accounts.
• In between: tools like STAAT that suggest, while a human advisor decides.

If you have read my guide to letting an AI agent handle money, this is the bank-side version of the same rule: an agent can prepare, but the moves that cost money stay with a person, or stay out.

What about the 1,500 jobs?

Some coverage this week put the agents and the headcount in the same sentence. They do belong to the same quarter. But the numbers do not say what such a sentence suggests.

From UBS’s own key-figures table:

• Internal staff (full-time equivalents): 97,593 on 30 September 2026, down from 99,085 on 30 June. That is a drop of 1,492. Reuters described it as “almost 1,500 full-time employees in the quarter as integration progressed.”
• Internal plus external staff: 109,990, down from 112,388. The finance chief told analysts total headcount was “110 thousand, 2% lower sequentially and 30% below our 2022 baseline,” the year before UBS took over Credit Suisse.
• A year earlier: 104,427 internal full-time equivalents on 30 September 2025.

UBS frames the decline as part of the integration, which it expects to complete by the end of 2026: the bank has delivered $13.2 billion in gross cost savings so far and now expects about $13.7 billion, and around 90% of the legacy Credit Suisse applications slated for shutdown are fully decommissioned. Merging two big banks removes duplicate jobs. That is what the bank says is happening.

What UBS does not say is just as important. It does not claim AI caused any of the reduction. It also does not say AI played no role, or how much work the agents and the 18,000 AI-assisted developers took over. “Measurable productivity gains” is the only description, without a number. So the honest summary is: headcount fell by about 1,500 in the quarter, UBS attributes it to the integration, and nothing published so far allows anyone to say AI replaced those people. I wrote about the wider habit of labelling cuts as AI, in both directions, in Is AI Really Behind the Layoffs?

What UBS does not disclose

For a bank that ranks sixth in an AI index that also scores transparency, the agent figure comes with very little context. These are the gaps I could not close from the release, the earnings call remarks, the second-quarter release or the bank’s technology page:

• What counts as an agent. Is it a saved prompt, an automated workflow, or software that can act in other systems? Are 12,700 agents created or actively used?
• How many people run them. 12,700 agents is not 12,700 people. UBS has about 97,600 internal full-time staff, so the figure works out to roughly 13 agents per 100 employees. One power user may run ten; most may run none.
• Who checks them. There is no statement on how personal agents are registered, risk-rated, tested or switched off, or who is accountable when one gets something wrong. The bank says all employees take mandatory annual training in responsible AI use, which is a start, not a control.
• How many face clients. The release does not split the 608 use cases into internal and client-facing ones.
• What they deliver. No hours saved, no cost figure, no error rates. In the Evident index, only twelve of fifty banks report AI’s actual or targeted value at all. UBS’s results release puts no number on it.

None of this means something is wrong. It means outsiders, including clients and investors, have to take the governance on trust.

Why the governance question is not theoretical

Switzerland has no AI-specific law for banks. But the supervisor has spelled out what it looks for. In its Guidance 08/2024 of December 2024, FINMA wrote that institutions with many or significant AI applications should have “a centrally managed inventory with a risk classification and resulting measures,” clear responsibilities, testing requirements, documentation and broad training.

Then comes the sentence that describes the 12,700-agent problem almost exactly. FINMA found that ensuring complete inventories was a challenge “as AI development and use is often widely spread across the organisation and, since the advent of generative AI, applications are accessible to everyone.” It also lists a “degree of autonomy and process integration” as a factor that raises risk, and says risks are higher “when customers or employees are strongly affected” by an application’s results.

In plain words: the regulator’s model assumes the bank knows every AI application, knows how risky each one is, and knows who owns it. A firm where every employee can build their own agents has to do that bookkeeping at a scale no bank has done before. UBS may well be doing it. It just does not say how.

What this means for companies

• Count your agents before someone else does. If staff can build their own, you need a register: who built it, what it touches, what it may do on its own. FINMA’s inventory expectation is a good template even outside Switzerland.
• Separate “prepares” from “acts.” Most of the value of personal agents comes from drafting, summarising and checking. Keep actions that move money, data or commitments behind a human approval step, and write that rule down.
• Watch the corrections. FINMA suggests monitoring cases where users ignore or change an AI’s output. For personal agents, that is the cheapest early-warning signal you have.
• Be careful with the headcount story. If you cut jobs during an AI rollout, say clearly what drove the cut. UBS ties its reduction to the merger. Mixing the two in one message invites the reading that AI replaced people, whether or not it did.
• Publish one number that means something. “12,700 agents” says adoption. Hours saved, error rates or the share of agents that can take actions would say whether it works.
• Training is the foundation, not the fence. UBS pairs its agents with weekly “AI Power Hour” learning time and mandatory responsible-AI training. That helps people use agents well. It does not replace an inventory and controls.

What this means for employees and bank clients

This is general information, not financial or career advice.

• For employees: being allowed to build your own agent is an opportunity to automate the dull parts of the job. Keep a short note of what your agent does and what data it uses. If something goes wrong, you will want to show that you checked its work.
• Your judgement is the part that counts. UBS itself describes the goal as more time “on judgement, collaboration, and delivering value for clients.” Agents that draft and sort make that judgement more visible, not less needed.
• For clients: the agent you deal with today is almost certainly a human advisor using AI-prepared material. UBS says outside AI agents cannot act on your accounts. If you use a third-party AI assistant for your finances, do not hand it your banking password.
• Ask your bank directly. You can ask whether AI was used in advice you receive and who is responsible for it. A good bank will have an answer.

What we don’t know yet

• The definition and the count. UBS has not published how it counts personal agents, or whether the figure includes inactive ones.
• The controls. No agent-specific governance description has been published. FINMA’s expectations are public; how UBS meets them for personal agents is not.
• The client share. How many of the 608 live use cases touch clients directly is not disclosed.
• The effect on jobs. Whether AI will reduce staff needs once the integration is finished is an open question. UBS says its focus will shift to “sustaining both a structurally efficient cost base and disciplined resource allocation.” That sentence allows both readings.
• The comparison. Other banks report agents differently, or not at all. A count of “agents” is not yet a standard unit, so comparing banks by it would mislead.

The bottom line

UBS has done something few companies have: it put a hard number on how many AI agents its employees run, and the number is large. It also drew a clear line at the client’s door. What it has not shown is the part that makes 12,700 agents safe and useful: who builds them, what they can touch, who checks them and what they deliver. The staff reduction in the same quarter belongs to the Credit Suisse merger, by UBS’s own account, not to the agents. For every other company, the lesson is simple. Letting everyone build agents is the easy part. Knowing what all of them do is the job.

Sources

• UBS, “UBS’s third-quarter 2026 results” (media release, ad hoc announcement), Zurich, 28 October 2026, including the key figures table. ubs.com
• UBS, third quarter 2026 results, earnings call remarks, 28 October 2026. ubs.com (PDF)
• UBS, second-quarter 2026 results, media release, 29 July 2026. ubs.com (PDF)
• UBS, “Innovation and AI” (technology page, accessed 1 November 2026). ubs.com
• Reuters, Ariane Luthi: “UBS meets expectations with $2.7 billion Q3 profit”, Zurich, 28 October 2026 (via Investing.com). investing.com
• Reuters, Ariane Luthi: “UBS integration of Credit Suisse slowed AI deployment, Ermotti says”, Zurich, 27 October 2026 (syndicated). srnnews.com
• PYMNTS: “UBS Employees Run 12,700 AI Agents as Bank Automation Accelerates”, 29 October 2026. pymnts.com
• FINMA, Guidance 08/2024 “Governance and risk management when using artificial intelligence”, 18 December 2024. finma.ch
• Evident, “Inside the 2026 AI Index”, 8 October 2026. evidentinsights.com

This article is for information only. It is not investment, financial, legal or career advice. Mentions of UBS refer to its published statements and reported remarks, not to its shares.