Banks are racing to count their AI agents. Few can say what those agents earn

In little more than a week, three banks in Europe and Asia put fresh AI agent tallies in front of the market. UBS said in its October 28 third-quarter results that employees are operating more than 12,700 personal AI agents. BBVA's new framework for building and governing agents, first shown to analysts and investors on October 6, was detailed again in November 3 coverage that attached production-pilot numbers. And on November 5, South Korea's NH NongHyup Bank unveiled ten back-office agents picked from 39 staff proposals. The tallies are getting bigger. What the agents actually return is still mostly missing from the disclosures.

That gap matters because agent counts are turning into the banking industry's preferred proof of AI progress, and the independent index UBS itself cites as a credential does not yet measure outcomes at all.

What the banks actually disclosed

UBS gave the longest list. Beyond the 12,700 personal agents, it reported 608 AI use cases live in production, up 81% from a year earlier, more than 1,090 further applications in development, over 18,000 developers using AI coding tools, and more than 30 million year-to-date insights from STAAT, its in-house tool for advisor teams. It also described nine large-scale AI transformation initiatives. What the release does not do is connect any of this to money. UBS reported $13.2 billion in cumulative gross cost savings and expects roughly $13.7 billion by the end of 2026, but the release ties those savings to the Credit Suisse integration, including retiring legacy IT applications, and never says how much of the total, if any, it credits to AI.

BBVA offered more operating detail. As reported by QA Financial, the Spanish bank says voice agents resolve close to 90% of interactions in production pilots in Peru and Mexico, and that assistants for relationship managers in Spain now screen more than half of the app conversations those staff receive. It puts developer productivity gains at about 28%, which is its own measure; the coverage gives no figure on software quality. Its customer-facing assistant, Blue, won't reach individual and business customers until a gradual 2027 rollout. Antonio Bravo, BBVA's global head of AI transformation, framed the bet as an integration race: "Everyone has access to AI models. The difference will lie in who is able to integrate them into their systems at scale."

NongHyup's announcement was the smallest and, in some ways, the most specific. Seoul Economic Daily reports that the ten agents, built with the bank's AI subsidiary AgileSoDA, include one that extracts 318 data points from the 500 to 1,000 pages behind each project-finance deal, and a loan-checking agent that reads the roughly 600 pages attached to a single loan. A staff member who helped build the project-finance tool expects drafting that took at least five business days to take about four hours. That is a projection, not a measured result. Demonstrations run through November 6, and the deployment versions are still being finalized after front-line feedback. AgileSoDA's chief executive, Choi Dae-woo, was reported as saying that while the number of agents matters, the focus will be on practical use in real work and on cutting errors. Coming from the bank's own AI unit on launch day, that is a telling caveat.

Why counts are a weak signal

An agent count tells you about supply, not demand. A personal agent that an employee builds once and forgets still counts toward UBS's 12,700. A use case moved into production may serve a handful of users or thousands. None of the three announcements reports usage depth, error rates, cost per task, or the headcount and spending each workflow replaced. The closest thing to an outcome, BBVA's 90% resolution rate, comes from pilots in two markets and doesn't say how resolution is defined or how often customers called back.

Evident, the research firm whose annual AI Index has become the industry's main league table, has been documenting this pattern for a while. Its August 2025 outcomes study looked at 173 public AI use cases across 50 large banks and found that only about 30% disclosed any outcome metric. It also noted that ten banks had rolled out generative AI assistants to more than 800,000 employees between them, yet few had firm-wide productivity figures. Co-founder Alexandra Mousavizadeh said at the time that many banks were still early in working out AI returns.

The 2026 index, published October 6, shows slow movement. According to coverage of the release, the number of banks putting any figure on AI returns, achieved or forecast, rose from eight to just twelve out of 50. Lloyds was one of them, having said in January that its generative AI work was worth around £50 million last year and projecting more than £100 million this year. The index itself ranks banks on public information across four pillars, with AI talent weighted most heavily, alongside innovation, leadership attention, and transparency on responsible AI. UBS, which ranked sixth and touts its standing as the index's leading European bank, is being rewarded for capability and disclosure, not proven returns. Evident says it will add a fifth pillar measuring AI outcomes in 2027. The leader shows why that matters: as we reported on October 19, JPMorgan Chase's CFO said he had stopped trying to isolate AI's return on investment even as the bank held the top spot.

The cost side nobody counts in agents

Banks are also spending more to defend against AI than their agent tallies suggest. At Regions Financial's October 16 earnings call, PYMNTS reported, CFO Anil Chadha tied an expected 1.5% to 3.5% rise in adjusted noninterest expense for 2026 to cybersecurity investment, pointing to "frontier AI" and calling the spending "the new reality that we live in today." That spending doesn't show up as an agent or a use case, but it is part of what AI costs a bank.

There are some harder numbers out there. PYMNTS reported in October that a Customers Bank pilot closed selected commercial loans in seven days, against a typical 30 to 60. That is the kind of before-and-after figure that lets an outsider judge a deployment, though it, like NongHyup's drafting estimate and BBVA's resolution rate, comes from a pilot or a projection rather than full-scale operation.

What to watch

For anyone tracking bank AI, the useful questions are simple. How many of those 12,700 agents are used every week? What share of the 608 production use cases has a measured cost or revenue effect, and is any of UBS's $13 billion-plus in savings attributable to them? Does BBVA's pilot resolution rate survive scale and outside audit? Does NongHyup's four-hour drafting estimate hold once real deal files hit the system? Evident's outcomes pillar in 2027 will push banks to answer some of this in public. Until then, a bigger agent count mainly shows that a bank has been building. It doesn't show what the building has paid for.