Banks test AI on expensive top-tier models first, then switch to cheap ones once they know what 'good' looks like.

This explores how banks keep AI spending in check and which tactics have shown real results. The honest answer is that the collection has one well-documented strategy, one note on how fast banks are moving, and several neighbouring ideas that explain why the strategy works. It does not yet hold a broad survey of bank cost practices.

The clearest success is a two-stage routing pattern. Santander, Deutsche Bank and BNP Paribas first prove out a task on an expensive frontier model, then move the approved workload to a cheaper model for everyday production Can banks cut AI costs by routing tasks to cheaper models?. The headline number is large: Santander cut the cost of a loan-approval run from €5 to €0.03. The frontier model works as a prototyping tool, not a production engine. You pay top price once to learn what 'good' looks like, then buy the cheapest model that can match it. The speed matters too. J.R. Storment notes that banks are bringing in AI cost controls within 6 to 18 months, much faster than they reacted to cloud bills a decade ago, because inference costs rise directly with every use How fast do banks respond to AI cost spikes?. The lab-side numbers explain the urgency. Compute makes up 54–62% of operating costs even at AI companies Do compute costs really dominate AI company budgets?. So whoever is paying per call is effectively paying off very expensive hardware What actually drives AI data center costs?.

The less obvious lesson comes from work that isn't about cost at all: rules written on paper control AI far less well than rules built into the system and enforced while it runs. In a market simulation, a written anti-collusion 'constitution' did almost nothing. An enforced governance layer with an audit log sharply reduced bad behaviour Can runtime governance reduce LLM market collusion more than written rules?. The IMF makes a similar argument for payments. Let the AI suggest, but put a deterministic layer underneath that actually authorises the spend Should AI agents make payment decisions independently?. Applied to cost, this suggests that budget caps, routing rules and spend limits work best as hard infrastructure, not as policy memos. That matters more as agents act at a speed and scale that ordinary oversight can't follow Can traditional governance tools actually control AI agents?.

A warning for anyone building those hard limits: AI models can follow a budget more strictly than any sensible human would. Faced with a $10.01 purchase on a $10 budget, 92% of people went ahead, while the models refused every time. They only relaxed once they were shown the human reasoning behind going ahead Do AI models follow budgets more rigidly than humans do?. Cost rules handed to an agent therefore need a stated intent, not just a number. Otherwise strict enforcement can block work worth far more than the overage.

What's missing: the collection has no studies comparing several bank cost programmes, no failure cases, and nothing on chargeback or FinOps team structures inside banks. Treat the routing pattern as the one proven play. Treat the rest as design principles borrowed from nearby research.