Banks spent the last couple years talking about AI the way teenagers talk about crypto: like it was going to make them rich, fast, and effortlessly. Now reality’s arrived—with an invoice.
Roughly 70% of banks are cutting back their AI investments, spooked by the ugly operational costs that come with running generative AI at scale. The pitch was simple: automate work, speed up decisions, trim headcount. The fine print was brutal: every call to an AI API costs money, the computing infrastructure guzzles electricity, and software licensing fees have a way of blowing past “initial estimates.”
The quiet rationing nobody wants to brag about
Instead of holding a press conference that says “we overspent,” banks are doing what banks do: tightening the screws quietly.
Teams are getting usage quotas. Some AI projects are frozen. Others get stripped down to the bare minimum—less experimentation, fewer prototypes, fewer big rollouts. In some shops, the cost of operating large language models is starting to look like a line item with the same political sensitivity as payroll.
And the practical fallout is already showing. Customer-service groups that wanted 24/7 AI chatbots are settling for hybrid setups—part automation, part human. Back-office departments that dreamed of automating thousands of tasks are discovering an inconvenient truth: for certain workflows, the “AI worker” can cost more than a human employee once you factor in compute, integration, and ongoing usage fees.
Innovation meets the budget committee—and loses
This isn’t happening in a vacuum. Banks already shelled out serious money for data, cloud contracts, and the talent arms race—data scientists don’t come cheap. Piling a generative AI operating bill on top of that can look less like “investment” and more like “financial self-harm.”
The timing couldn’t be worse. While traditional banks tap the brakes, fintech rivals and Big Tech keep pushing forward with AI budgets that, for all practical purposes, don’t have a ceiling. If banks fall behind now, catching up later won’t be cheap—and it won’t be quick.
A forced diet that might actually help
Here’s the twist: this pullback could make bank AI efforts less dumb.
When money gets tight, the projects that survive tend to be the ones with a real return—not the flashy demos built to impress executives at quarterly meetings. Expect more pressure to cut wasted queries, tune models, and pick smaller, more efficient systems instead of brute-forcing everything with the biggest model money can rent.
But let’s not pretend this is painless. Banks are stuck in a three-way squeeze: customers want slick digital service, regulators want tighter controls, and competitors want their lunch. Slowing down on AI means risking ground on all three fronts at once.
Sources
Reporting and context drawn from: Le Temps (“En 2026, on verra les gagnants et les perdants de l’IA”), TDG (markets and AI shifts projected for 2026), Finance et Investissement (AI as a potential macroeconomic shock), Yahoo Finance France (record AI funding figures), and a YouTube segment on AI investing realities heading into 2026.
