Can banks cut AI costs by routing tasks to cheaper models?

Bloomberg reports that as AI bills have begun surprising bank CFOs, several banks have converged on a two-step cost-control pattern: validate a task on a large, expensive frontier model, then move the proven workload to a smaller, cheaper model for everyday production use. At Banco Santander, this shift "cut execution costs from about €5 ($5.61) to €0.03 per loan approval process," according to Ignacio Bernal, the bank's chief artificial intelligence officer, quoted by Bloomberg. Deutsche Bank's Christoph Rabenseifner described testing "one task on a smaller, specially trained model instead of a large frontier model, cutting costs several times over." At a large European bank, an engineer identified only as Alex reworked a chatbot so that "simpler requests went to open-source models" while "only token-intensive prompts ... went to premium models," cutting the bill "by roughly a third," Bloomberg reports.

Bernal frames the logic as follows: "when you move AI from pilots to production across a global bank, inference cost stops being a rounding error and becomes a real operational cost line," since some processes "run millions of times a year." The routing decision is made case by case: a bank first confirms a frontier model clears its quality bar on a task, then substitutes a cheaper model for the repeated, high-volume execution, banking the savings on scale rather than on any single run. BNP Paribas applies an upfront version of the same calculus, per Bloomberg: the bank "approves every AI project after its running cost is calculated," multiplying an internal token price by expected volume and weighing the total against expected business value, said group chief data officer Hugues Even. Other banks pair routing with hard usage caps — CIBC restricts its priciest models (including Anthropic's Opus, which it says can cost C$3,000 per employee a month) to a few hundred staff, and JPMorgan has set $2,000 monthly spending limits on Anthropic's Claude Code for certain engineers, Bloomberg reports, citing Business Insider.

The pattern qualifies Should benchmarks report token counts instead of frozen costs?: Bain's Bhavi Mehta, quoted by Bloomberg, makes a parallel point from inside a bank — "the tokens required for a given task can vary significantly" as workflows turn agentic — but Bloomberg's bank examples manage that variance by actively re-routing workloads to cheaper models rather than by waiting for prices to fall. It also sits alongside Do AI customer service agents actually reduce total support costs?: both report a gap between AI's advertised savings and operational reality, though the banks here respond with engineering fixes — routing, caching, prompt-trimming — rather than Klarna's labor-substitution framing. And it extends the banking-AI-economics territory of Will AI agents break the deposit and credit card inertia? from the revenue side of the ledger to the cost side.

Bloomberg's reporting does not say how widely this validate-then-downshift pattern has spread beyond the handful of named banks, nor does it net the reported savings against the engineering cost of building the routing and caching layers themselves. Held to that limit, the implication is that cost-control is becoming a design requirement for production AI deployments at scale rather than an afterthought: Evident's count of only 12 of the 50 largest global banks publicly tying AI adoption to measured returns suggests most of the industry still cannot say whether these routing strategies pay off — only that enough CFOs are worried enough to try them anyway.