UBS used its third quarter results(opens in a new tab) to publish one of the most detailed AI scorecards any large bank has put in an earnings release. It has 608 live AI use cases in production, up 81% on a year ago, with more than 1,090 more in development. Employees run over 12,700 personal AI agents in their daily work. More than 18,000 of its software developers use AI tools, which the bank says has led to measurable productivity gains. Its STAAT tool for advisor teams has generated over 30 million client insights this year.
Read the same release for money and the numbers change character. UBS has delivered $13.2 billion of gross cost savings and now expects about $13.7 billion by the end of 2026, above its earlier target of about $13.5 billion. Internal staff, counted in full-time equivalents, fell from 99,085 to 97,593 in the quarter. The release ties that progress to the Credit Suisse integration, including retiring nearly all the legacy IT applications in scope. The release does not say how much of any of it came from AI.
That gap is the point. Use cases, agents and developers are measures of activity. A CFO cannot bank them. A bank that can count 12,700 agents can presumably count the hours they return and the cost lines they touch, and it has not published that. Most enterprises are in the same place with far less to show.
It also says where to look for the return. PYMNTS Intelligence research(opens in a new tab) finds the most adopted AI uses in financial services and insurance sit in structured, auditable back-office work: revenue recognition and accounting close at 65% of firms, and credit risk scoring and sales forecasting at 60% each. Those are processes with existing cost baselines, which makes them the easiest place to prove a saving.
For the CAIO and CFO this week. Take your own AI scorecard and mark every metric as activity or outcome. For each of your five largest deployments, write the cost line or revenue line it is supposed to move and the baseline you will measure it against. If you cannot name one, that deployment belongs in the count of experiments, not in the business case.
Action items
UBS published one of the most detailed AI scorecards in any bank earnings release: 608 live use cases and 12,700 personal agents. It put no dollar figure on what that AI saves, which leaves even a leader reporting activity rather than return.
For the CAIO and CFO
Tie each of your largest AI deployments to a named cost or revenue line and a baseline before the next board report.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
Also worth knowing
- UBSUBS's third-quarter 2026 results(opens in a new tab)
UBS reports 608 live AI use cases, 12,700 personal agents and 18,000 developers using AI tools, alongside $13.2 billion of gross cost savings. None of those savings is attributed to AI, so even leaders still report AI as activity.
- PYMNTSUBS Employees Run 12,700 AI Agents as Bank Automation Accelerates(opens in a new tab)
Financial firms adopt AI most in auditable back-office work: revenue recognition and accounting close at 65%, credit risk and forecasting at 60%, and 85% plan bigger AI budgets. Start ROI proof where a cost baseline already exists.
- CloudflareCross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang(opens in a new tab)
Between August 27 and September 9, 155 responses from a hosted model contained data from other customers' requests, caused by a cache sizing mismatch and an open-source serving bug. Ask every inference provider how they isolate tenants on shared GPUs.
- Associated PressMicrosoft offers 60,000 employees voluntary buyouts as AI spending accelerates(opens in a new tab)
Microsoft offered voluntary exits to about 60,000 staff, roughly 27% of its workforce, as it spends heavily on AI. Expect your board to ask how your own AI budget changes your headcount plan.
- Liquid AIIntroducing LFM3: Every Modality for Specialized Intelligence(opens in a new tab)
Liquid AI released six open-weight 2B-class models for text, vision, audio, omni and decision tasks that run on local devices, with a 256K context for text. Small on-device models can take routine work off metered cloud token bills.