Skip to main content
Back to blog
  • Fintech

Updated 14 min read

AI agents in banking. From pilots to real work in 2027

Only 4 of 50 big banks could show an AI return in 2025. What the few with agents in production do differently, from 19 reports

Daniel Mercer

AI product and design workflows

ShareLinkedInXEmail

In 2025, Evident looked at 50 of the world’s largest banks. Only 4 could show a realized return from AI.

Yet by March 2026, BNY had about 140 AI “digital employees” at work and about 220 AI tools in production.

Of the 65 reports we read from McKinsey, BCG, Deloitte, Accenture, PwC, Visa and Mastercard, more carry this theme than any other: agents leaving the demo room for everyday bank work such as KYC checks, servicing, collections and code.

For you, the stakes are your cost per customer and the size of the operations team you budget for in 2027.

The few banks that already get results do four things differently, and none of them is buying a better model.

The first difference is about how much of a process a bank is willing to rebuild.

Sign up with email

The short version

  • Returns are still rare. Only 4 of 50 large banks tracked by Evident reported realized ROI from AI in 2025, Deloitte notes.
  • Agents work where the job is paper. BNY says AI now reviews about 70% of its restricted party screening for payments, and those cases close more than 30% faster.
  • Production numbers are large. BCG reports about 50% lower collections operating costs and up to 50% lower financial crime costs at retail banks that run agents in production.
  • The change is mostly people. McKinsey’s Klemens Hjartar puts it at 85% people and culture, 15% technology.
  • The cost side is open. 56% of financial services finance leaders expect AI costs and complexity to rise substantially through 2027, Deloitte found.

1. Agents already do real work, in the least glamorous parts of the bank

BNY is the clearest case I found. In its first-quarter 2026 update, filed with the SEC, the bank reported about 220 enterprise AI solutions in production and about 140 “digital employees”, which it describes as multi-agent systems that work alongside staff.

The numbers under that headline matter more. AI reviews about 70% of restricted party screening for payments, and those reviews resolve more than 30% faster. About 25% of new client onboardings in the quarter ran through an AI-enabled process. In the same quarter, AI wrote more than 40% of BNY’s code. By October 2026, the bank said it had more than 300 AI solutions in production.

JPMorgan Chase reports in less detail. Its 2025 letters to shareholders say AI lets the commercial and investment bank review more than double the transaction screening volume with half the manual operator checks. In that division, over 90% of engineers use AI code assistants, and more than 65,000 people actively use LLM Suite, the bank’s own AI platform.

Bank of America shows the servicing version. More than 18,000 customer service staff use EricaAssist, which the bank calls a human-assisted AI agent. It cuts average call time by nearly one minute per interaction, according to the bank’s July 2026 release.

Goldman Sachs shows how thin the line between pilot and work still is. In February 2026, CIO Marco Argenti told CNBC that Goldman had spent six months building Claude-based agents with Anthropic engineers for trade accounting and client onboarding. He described the work as early stage, with launch expected soon.

The reports add their own examples:

  • McKinsey’s Global Banking Annual Review 2025 cites a US bank that rebuilt credit risk memos around agents. Productivity rose 20% to 60%, and credit turnaround improved by 30%.
  • BCG’s executive perspective on AI-first retail banks counts only production-grade results. Voice bots handle over 70% of human call volume at about one fifth of the cost. Operations turnaround falls by 70%.
  • BCG’s retail banking report describes an Asian bank whose collections agents cut costs by 30% to 40%.

These jobs have the same shape: read a document, check it against a rule, write a summary, route the exception. That describes most of a bank’s back office. McKinsey’s David Deninzon estimates that 50% to 60% of a bank’s full-time staff are tied to operations in some way.

Figure 1. What agents delivered in production at retail banks (%)
ResultValue
Lower operations turnaround time70
Human call volume handled by voice bots70
Lower collections operating costs50
Lower financial crime costs (up to)50

Source: BCG, “How AI-First Banks Are Rewriting the Rules of Retail Banking”, June 2026. Results from production-grade implementations.

What to do in 2027: list your exception-heavy, document-heavy processes. Those are the first candidates, not the chatbot.

2. Most banks are still in pilots, and the gap is widening

Deloitte’s 2026 banking outlook, citing Evident, says only 4 of 50 banks reported realized ROI from AI use cases in 2025. Counting projected returns too, Evident’s 2026 index, published this month, finds 12 banks, up from 8 a year earlier. That is better, and still a minority of 50. McKinsey’s Abhilash Sridharan describes the paradox in Asia: nearly 80% of the financial institutions McKinsey works with there use AI-led applications, yet a similar share globally reports no significant bottom-line impact.

BCG’s Build for the Future study, quoted in its retail banking report, finds the same split across all industries. 5% of companies get value from AI at scale. 60% get none.

Visa surveyed 325 AI decision-makers at banks and neobanks in 17 European markets. 86% think AI will fundamentally change retail banking economics by 2030. Yet only a minority claim transformational gains, and the pattern repeats across countries, bank types and functions.

The largest US bank admits as much. On the October 13, 2026 earnings call, JPMorgan’s CFO Jeremy Barnum agreed that AI benefits are not yet visible in quarterly results. He put the bank’s AI spend at something like $700 million a year, and said benefits will be partly offset by token costs.

Meanwhile the leaders pull away. BCG finds that companies generating value from AI at scale report five times the revenue increases and three times the cost reductions of the rest. McKinsey’s Saptarshi Ganguly estimates up to a six-point gap in return on equity between first movers and fast followers.

What to do in 2027: decide which group you are in, using money, not the count of pilots.

3. Difference 1: they rebuild a whole process, not a use case

McKinsey’s David Deninzon names the trap the “fingers and toes problem”. Banks automate a part of what one analyst does, and the rest of the job stays manual. He says this is why the robotic process automation wave of a decade ago fell short.

The fix is to pick fewer things and go end to end. Sridharan describes an Asian bank that broke its operations into about 600 processes and subprocesses. The COO’s team then chose the ten with the biggest payoff. BCG’s advice to retail banks is similar: pick 3 or 4 areas, and look at the nature of each job instead of the number of people in it.

When banks do this, the numbers move. In McKinsey’s UK banking report, relationship teams that rewired a single workflow end to end saw 20% to 40% lower cost to serve, and got back 10 to 12 hours a week per banker.

The goal matters too. Visa found that 30% of European banks name cost reduction as their main AI driver, nearly twice as many as those who put customer experience first. Yet the customer-first banks were 40% more likely to reach transformational gains.

I think this is a design question as much as a technology one. An agent that only fills a gap inside an old process inherits all its waiting and handoffs. When we map a bank journey end to end, the delays rarely sit in one screen. They sit in the handoffs between teams, and that is where an agent can take over.

What to do in 2027: choose three or four journeys, from customer request to closed case, and redesign each one with agents in mind.

4. Difference 2: they build the control layer before the agents

The banks that scale agents treat control as infrastructure.

BCG and OpenAI, in their joint report on retail banks, describe two pieces. The first is evaluation-driven development: testing each agent on the same document-heavy tasks people do, and tracking quality as models change. Their example is an underwriting agent that makes every system call correctly, yet checks income and skips credit history. Every call succeeds, and the decision is still wrong.

The second is a single middleware layer, a “front door” that every AI workload passes through. It holds identity, permissions, logs and kill switches in one place, so every AI application runs under the same governance.

BNY says all its AI tools pass a governance process before going live. In fund operations, AI-driven oversight cut false positives in net asset value checks from 90% to under 20%.

Regulators are watching control too. McKinsey notes that the UK’s Prudential Regulation Authority expects models used in credit risk and KYC decisions to be explainable, traceable and continuously validated, whether AI-based or not. McKinsey’s Andrea Del Miglio adds a useful point: there is no market standard yet for identifying an agent, but agents leave a full audit trail of every action.

We’ve seen more than one big fintech restrict Claude Code’s auto mode over data concerns, so every request needs a manual “Allow”. That is the control problem in small. Without a governed layer, the only safe setting is to slow everyone down.

What to do in 2027: fund the evaluation and control platform before the tenth agent, not after it.

5. Difference 3: they treat it as a people change

The reports agree on this more than on anything else. Hjartar tells CEOs the change is 85% people and culture, 15% technology. Ganguly says the agents themselves account for no more than 20% to 25% of the value. The rest comes from the operating model, data, talent, risk and governance.

The work itself changes shape. Sridharan expects employees to move from spending 80% of their time on coordination and rule-based tasks to spending 80% on customers, decisions and new ideas. McKinsey expects one person to supervise roughly 20 to 30 agents in some cases.

BNY shows what that preparation looks like. About half its employees use AI daily, and over half are building AI agents. It delivered 171,000 hours of AI learning in 2025.

Jobs do not simply vanish. Accenture’s survey of banking IT executives found only 6% plan to reduce traditional IT roles. Ganguly expects some roles to go (doers, checkers, hands-on coders), others to change, and new ones to appear, such as AI trust and safety specialists.

In our own work at Humbleteam, AI skills now let us design apps and web projects 20-40% faster. Most of that came from handing repetitive steps, such as first layout variations and handoff notes for engineers, to agents, while designers keep the decisions.

What to do in 2027: name who supervises each agent, what they are measured on, and what the people it frees up will do next.

6. Difference 4: they count money, not activity

Deloitte’s 2026 outlook warns that many AI benefits rest on user claims, with no baseline and no control group. Its advice: tie every claim to a financial or risk metric, and set a baseline before launch.

Finance leaders are already preparing for this. Deloitte’s Finance Trends 2027 survey of 205 financial services finance leaders found:

  • 56% expect AI costs and complexity to rise substantially through 2027.
  • 31% approve AI money through a stage gate: pilot first, then more funding if it works.
  • 23% require a formal business case with quantified ROI.
  • 76% are comfortable with AI moving into some form of autonomous decision-making in finance.

McKinsey adds a warning about where savings go. In its central scenario, banks’ net unit costs fall about 15%. But competition is expected to pass much of that saving to customers, so it will not lift profits for most banks in the long run.

What to do in 2027: report agent results the way you report a branch: cost, volume, error rate and customer outcome, against a baseline.

The 2027 checklist

  1. Processes. Three or four end-to-end journeys chosen, each with an owner.
  2. Control. One evaluation and middleware layer that every agent passes through.
  3. Data. Agents can reach the documents and systems a human analyst uses, with the same permissions.
  4. Roles. Every agent has a named human supervisor and a defined escalation path.
  5. Metrics. A baseline for every process before the agent goes live, and token costs counted in the result.

Where the reports disagree

The size of the prize. BCG, with OpenAI, says agents could cut bank costs by 30% to 40% by 2030. Accenture estimates scaled gen AI across the top 200 banks could lift revenue 5%, cut operating costs 8% and cut loan-loss provisions 16%, worth $289 billion. McKinsey’s central case is a 15% fall in net unit costs. It sees a 40% fall only in a scenario it considers unlikely in the medium term, and expects most savings to be competed away.

The productivity ceiling. McKinsey’s UK report says a copilot beside a banker lifts productivity about 1.2 times, while supervised multi-agent workflows could reach 20 times. In the same passage it says no bank has achieved that yet.

The proof. BCG and Accenture show strong case results. JPMorgan, with a 2026 technology budget of about $19.8 billion, says the benefits do not yet show in its quarterly earnings. I don’t see a contradiction. Process gains tend to appear years before they move a bank’s income statement.

What would change the picture is regulation. McKinsey’s most radical scenario depends on regulators accepting agents that execute transactions on their own. Until then, a human stays in the loop, and the 2027 gains come from back-office work.

What I’m watching

In 2027 I’ll be watching whether more banks publish counts of agents in production, the way BNY does, and whether JPMorgan’s results start to show the benefit. I also want to see Goldman’s onboarding agents move from early stage to daily work.

What separates the few from the rest is mostly unglamorous work: choosing the process, building the controls, bringing people along. If I ran a bank’s 2027 plan, I’d start there before approving another pilot.

Sources

Daniel Mercer

AI product and design workflows

Back to top

Read next

Let's talk

Write Sergey on LinkedIn
Write Alexandra on LinkedIn
Write Emin on LinkedIn
Have questions? Ask AI
ChatGPT
Claude
Perplexity
Opens a new chat with context about us pre-loaded — ask anything

We’ll reply within 24 hours with case studies, a timeline, and an estimate.

Prefer email? Write to hi@humbleteam.com
All set – our team’s on it. Expect a reply soon.
Send another one
Oops! Something went wrong while submitting the form.
Have questions? Ask AI
ChatGPT
Claude
Perplexity
Opens a new chat with context about us pre-loaded — ask anything
Europe
Národní 135/14, Prague
Middle East
UAE, Dubai, Internet City Offices
Services by industryCompare usHealthtechSaaSDubaiLondonBlogFor AI SaaS
We use cookies to enhance your browsing experience,
serve personalised ads or content, and analyse our traffic.
By clicking "Accept All", you consent to our use of cookies.
Accept all
Decline
Privacy policy