Banks Spent Two Years Talking About AI. In 2026 They Started Reporting Numbers.

Every bank on earth announced an AI strategy in 2024. In 2026, a handful of them quietly started reporting numbers — and the numbers tell a different story than the strategies did.

I’ve been reading these deployments the way a product manager reads a competitor’s release notes: not for the headline, but for what the scope reveals about the thinking. And what struck me is that the banks getting results are not the ones that were most ambitious. They’re the ones that were most willing to be boring.

That’s an uncomfortable lesson, because ambition is what gets AI projects funded. Here are four deployments, what each actually shipped, and the product principle sitting underneath it.

1. JPMorgan: the agent prepares, the human decides

JPMorgan moved from AI agents that ran for two or three minutes to agents that run autonomously for an hour or two. In private banking, they screen market activity, client positions and research ahead of the banker’s day. The reported outcome: a 20% increase in private banking gross sales, with internal projections that bankers could expand client coverage by up to 50%.

“We’ve entered now the era of long-running autonomous agents.” — Derek Waldron, Chief Analytics Officer, JPMorgan

The underlying principle: separate preparation from judgment

Look closely at the scope. The agent does not advise the client. It does not recommend an allocation. It assembles the ground the banker stands on, and then it stops.

Most failed AI features I’ve seen fail because they were scoped across that line. Someone builds a feature that makes a recommendation, the recommendation is wrong 8% of the time, and users stop trusting the whole surface — including the 92% that was useful. The value was never in the recommendation; it was in the forty minutes of gathering that preceded it.

The question worth asking about your own product: in the workflow you’re about to automate, where exactly does preparation end and judgment begin? Most teams have never drawn that line explicitly. Drawing it is often the entire design decision.

2. Goldman Sachs: they gave the AI the work nobody wanted

Goldman deployed Devin, an autonomous AI software engineer from Cognition Labs, across a 12,000-person engineering organisation, reporting 3–4x productivity against earlier AI coding tools. CEO David Solomon framed the goal as simplifying and modernising the technology stack.

The assignment list is telling: updating legacy code, migrating systems, cleaning up old frameworks, generating boilerplate. Goldman calls it a “hybrid workforce” — the AI works, humans supervise continuously.

The underlying principle: automate where correctness is checkable

Legacy migration is unglamorous, but it has a property that matters more than glamour: you can tell whether it worked. The tests pass or they don’t. The system runs or it doesn’t.

This is the constraint most AI roadmaps ignore. Teams reach for the workflow with the biggest headline value, which is almost always the one where quality is subjective — content, strategy, recommendations. Then they discover they have no way to measure whether the AI is doing well, which means no way to improve it, which means the project stalls at ‘seems fine, I guess.’

The harder truth: if you can’t define what a correct output looks like, you’re not ready to automate that workflow. You’re ready to assist it. Those are different products with different risk profiles, and conflating them is how AI features quietly die.

3. Wealth management: four competitors, one workflow

Through 2026, a cluster of independent launches converged on exactly the same target — the advisor meeting.

Merrill and Bank of America Private Bank launched an AI meeting journey covering prep, summaries and next steps
Zeplyn shipped Advisor Coaching in July 2026, scoring every client meeting and returning feedback
Savvy Wealth released an agentic platform for independent advisors
Flanks launched a co-pilot for interrogating multi-custodian portfolios

The underlying principle: convergence is the cheapest market research there is

When four companies who are not talking to each other ship variations of the same thing in the same year, that is not a trend piece. That is a market telling you, loudly, where the pain is concentrated.

Most PMs spend enormous energy trying to find unmet needs and comparatively little watching what rivals independently agree on. But independent convergence is a stronger signal than any single competitor’s roadmap, because it can’t be explained by one company’s bad strategy. Four teams, four different sets of customers, one conclusion: the meeting is where the value and the waste both live.

Worth doing this quarter: list what three competitors shipped in the last six months. Circle anything that appears more than once. That overlap is where your users are hurting too, whether or not they’ve told you.

4. KYC and onboarding: the queue nobody complains about anymore

A large Dutch financial institution applied AI to KYC and compliance onboarding and reported a 90% reduction in onboarding time alongside a 30% cut in staff workload. Sector-wide, banks using agents for credit risk memos report 20–60% productivity gains and roughly 30% faster credit turnaround.

The underlying principle: look for normalised pain, not reported pain

No customer ever files a feature request that says “your onboarding takes eleven days.” They just take eleven days, or they leave. The pain has been there so long it stopped registering as a problem and became a fact of the business.

This is where I think most AI opportunity actually sits — not in new capability, but in the queues that have been slow for so long that everyone plans around them. Every organisation has three or four. They rarely appear in user research because nobody thinks to mention water being wet.

How to surface them: stop asking users what they want. Ask them what they schedule around. The answer is almost always a queue.

What connects all four

Strip away the industry specifics and the same four decisions appear in every one of these deployments:

Narrow scope, deliberately. Not ‘AI for the bank’ — one workflow, bounded
Humans supervising by design. In regulated industries this isn’t optional, and it turns out it isn’t a limitation either; it’s what makes deployment possible at all
A metric chosen before the build. Sales lift, turnaround time, onboarding duration. Each project had a number it was answerable to
Checkable outputs. Every workflow selected was one where wrong is visibly wrong

None of these are AI insights. They’re product management fundamentals — scope discipline, measurable outcomes, honest problem selection. What AI changed is the cost of ignoring them. A vague scope used to produce a mediocre feature. Now it produces an expensive, confidently wrong system that users learn to distrust.

A practical exercise for this week

Map three workflows in your product. For each, mark where preparation ends and judgment begins
For the preparation half, ask: could I tell if the output were wrong? If yes, it’s an automation candidate. If no, it’s an assistance candidate
Pick the one with an existing backlog or queue behind it — normalised pain beats requested pain
Write the success metric before the PRD. Not ‘faster.’ A number, a threshold, a date

If you want a structure for that, the PM Docs Toolkit has a PRD template where the metric and non-goals sections come first — the two places AI projects most reliably go vague.

Sources

A note on the numbers: these are company-reported figures, not independently audited results. Firms have every incentive to publish flattering metrics and quietly shelve disappointing ones. I’d treat the percentages as directional and the structural patterns — narrow scope, human supervision, checkable outputs — as the durable takeaway.