No, Jamie Dimon didn’t relinquish the CEO crown at JPMC. The most important job announcement in banking: M&T Bank named a new head of AI engineering. Finextra mentioned it in a 111-word article. M&T’s job description itself is nearly that long:
My take: This is a lot more important than you think—it’s a sign of a bank making a broad move from AI experimentation to AI transformation. This next stage of bank AI is about building the environment that allows AI to effectively perform work.
Oversimplifying a bit (but not much), AI budgets fall into two buckets:
I don’t have the numbers, but my strong bet is that, at most financial institutions (large and small), the lion’s share of the AI budget goes to Bucket 1.
It’s easy to see why. Bucket 1 spending is easy to explain, easy to scope, easy to stop, and makes someone faster. A $40 per seat license that saves an analyst four to eight hours a week is a no-brainer.
Bucket 2 spending has none of those properties. You can’t demo it, the business case reads like a plumbing invoice, and the payoff is down the road.
AI does three kinds of work: 1) see (read, classify, extract, summarize); 2) say (draft, respond, explain); and 3) do (waive an overdraft fee, raise a credit limit, issue provisional credit, re-age a delinquent account, book/fund a loan).
PenFed Credit Union’s AI agents exemplify this:
PenFed worked through governance, change management and testing before it scaled anything, then shipped internally first.
Bucket 1 doesn’t necessarily mean “see and say,” and bucket 2 doesn’t necessarily mean “do.” Bucket 1 can include an agent. But Bucket 2 is what makes agents scalable, governable, and reusable.
Is your Bucket 2 money having any impact? Here are four tests to determine the answer:
If the answers are “six,” “it uses a service account somebody set up in 2019,” “no,” and “we’d have to check with the vendor,” your AI budget is likely going into Bucket 1, and if there is anything going to Bucket 2, it’s not having much of an impact.
The payoff shows up as speed. Olivia Boles, who runs platform development and engineering at PenFed, described the hard part as ironing out security, privacy, trust and governance ahead of that first member-facing agent.
Once that was done, her team could “freely add actions thereafter,” because every new one reuses the same secure framework. Ace is now the orchestrator for a set of secondary agents covering mortgages, consumer loans, IRAs and investments.
The expensive work happens disproportionately up front. Once the environment exists, each additional capability gets cheaper and faster.
Without it, here’s what happens: A vibe coding product manager stands up a working prototype in an afternoon, and the build is nearly free. But everything downstream isn’t: the data, permissions, security review, a path to production. The prototype sits in a demo environment for months and dies there.
When the environment exists, that vibe coded app has somewhere to go. The cost of trying something drops toward the cost of thinking of it, and the constraint on an institution stops being its technology budget and becomes the quality of its ideas.
OK, that’s a bit of hyperbole, but the constraint does shift from the cost of building something to the institution's ability to decide what’s worth building.
M&T’s chief data officer has talked about deploying AI tools to 17,000 employees, which is a Bucket 1 move. Hiring an engineering leader to own platforms, architecture and governance is a Bucket 2 signal.
The pay for the position is in the $300K range, with responsibility for budgets across multiple cost centers, AI-enabled SDLC transformation, guardrails, operational controls and audit readiness.
Advisory roles don’t come with cost center ownership. This is a hire with a checkbook.
Note where the governance sits in M&T’s job description: the engineering department.
Bank and credit union AI governance today is an approval body: a committee, an acceptable use policy, a model risk process, a form. Necessary, but backwards looking.
The missing piece is AI deployment governance: who can ship what, to which systems, with which entitlements, and how it gets rolled back at 2am on a Saturday.
Putting governance under the person who owns the platforms means the controls get built into the pipeline instead of enforced at a meeting.
PenFed’s version of this was augmenting an already mature change management process with AI-specific tools before scaling, which is why its agents could expand scope after launch instead of stalling in review.
Productivity gains don’t pay for transformation. They increase capacity. But they don’t—and won’t—deliver the promise and potential of AI to achieve radical compression of process cycle time and process cost reduction.
That begs a question: Who’s going to pay for Bucket 2? It can come from one of three places: new money, reallocated money or realized economic value.
Best of luck with the first two.
Where does realized economic value come from? It can come from revenue, losses, capacity, headcount or technology spend. But the most immediate cost-reduction opportunities are in work currently performed by people or vendors.
That produces a Catch-22: You can’t reduce staff or reap tech spend reductions until after you’ve made investments in Bucket 2.
But when you do make that investment, a slew of tech spend reduction opportunities presents itself: overflow and after-hours contact center, card dispute processing, document processing BPO, collections agency placements, standalone chatbots and outsourced development.
Every one of those is a check written to a third party that an AI environment can begin absorbing. PenFed retiring an internal support phone line is a small, but real example of that conversion.
The cost of execution doesn’t disappear when AI gets cheap. It moves out of headcount and vendor invoices and into infrastructure, and the institutions that prebuild the destination get to move it deliberately instead of scrambling later.
The money to build the AI environment is a massive new vendor opportunity—and that's why the vendors are scrambling now.
The pundits will say this will happen in a 3-to-5-year timeframe. Too aggressive, and the major reason why is contracts. Even if you could build the AI environment to move everything, you’re locked into existing contracts.
AI transformation doesn't require waiting for every contract to expire. Institutions can (should) build the environment now while existing systems and contracts remain in place, then redirect work as contracts roll off.
A “head of AI” can mean strategy, use cases, evangelism and a roadmap. A “head of AI engineering” signals something different: ownership of the operating environment that gets AI into production—identity, platforms, pipelines, entitlements, controls and deployment.
When the second title starts showing up across the industry, you’re watching balance sheets reallocate in real time.
Not every institution has to fund Bucket 2 right now. You can sit it out and play catch up in the 2030s. The choice is yours.
Ron Shevlin is managing director and chief research officer at Cornerstone Advisors. Tune in to Ron’s What’s Going On In Banking podcast and follow him on LinkedIn and X.