Short answer. A bank controls agentic AI costs by setting hard per-task and per-session spend caps before go-live, not after the first invoice. An agent can consume 5 to 30 times more tokens per task than a single chatbot reply, and the same task can cost a different amount every time it runs. Treat each agent as a metered utility: give it a budget, a kill switch, and a dashboard the finance team reads daily. Banks that run AI in production govern its cost the way they govern credit risk, with limits set in advance and overruns that need sign-off.
A chatbot answers one question with one model call. An agent plans, retrieves documents, calls tools, checks its own work, and retries when it fails. Each of those steps is a separate model call, and the number of steps is not fixed. A model might close an anti-money-laundering alert in three calls on Monday and twelve calls on Tuesday for a case that looks almost identical. The cost follows the call count, so the cost is variable by design.
That variability breaks the budgeting method most banks still use. A per-seat or per-project estimate assumes a stable unit cost. Agentic work has no stable unit cost. One independent 2026 survey of enterprise AI budgets found that 85% of organizations missed their own cost forecast by more than 10%, and almost a quarter underestimated by 50% or more. The gap widens with usage, not with headcount.
A worked example makes the shape clear. A fraud-detection agent ran at roughly 5,000 EUR a month for 50 users. At 500 users, ten times the user count, it ran at about 15,000 EUR a month. Spend tripled while usage grew tenfold, because heavier users triggered longer agent runs. The controls that keep this in bounds are concrete:
In a bank, a cost overrun is not only a finance problem. Under the Digital Operational Resilience Act, in force since January 2025, a financial entity must understand and manage its dependency on ICT third parties, including the providers behind its AI. An agent whose spend is opaque is also an agent whose provider dependency is opaque. When cost forces a rushed switch of model or provider, that switch lands inside DORA’s expectations for a tested exit and continuity, not outside them.
Cost discipline also supports accountability. The EU AI Act requires record-keeping and human oversight for high-risk systems such as credit scoring and insurance pricing, with those obligations phasing in through 2 December 2027. A bank that already meters every agent call per task has most of the operational log those obligations expect. Cost control and audit readiness are the same instrumentation seen from two angles.
An agent can use 5 to 30 times more tokens per task than a chatbot, and the same task can cost a different amount each run.
The execution path is decided at run time, not at design time. A traditional application runs the same code for the same input, so its cost is stable. An agent chooses its next step based on what the previous step returned. Retries, extra retrieval, and self-checks all add calls that no one scheduled. Research on agentic cost estimation in 2026 shows that models are poor at predicting their own token use before a task runs, which is why forecasts built on an average task size break as soon as real work replaces demos. The honest planning number is a range with a hard ceiling, not a single figure.
Cost control is an architecture decision, not a monthly review. The controls that work are built into the agent before it reaches production: a token budget per task, a spend ceiling per session and per day, automatic termination when a limit is hit, and a queue that refuses new work once a daily cap is reached. Model tiering cuts the largest line item, because most steps in a banking workflow are routine and do not need the most expensive model. Caching repeated retrievals removes duplicate calls. None of this can be added cleanly once the agent is live, so a bank decides its cost limits in the same design review that decides its data access.
Run the agent against a representative sample of real cases and measure the full distribution of cost, not the average. The number that matters for a board approval is the 95th-percentile task cost multiplied by expected volume, plus a margin for growth in usage per user. Budget against that upper band, then set the hard caps below it so the system cannot exceed the approved figure. IDC projects that large organizations could face up to a 30% rise in underestimated AI infrastructure costs by 2027, driven by exactly this opaque consumption. Forecasting against the worst-case band, not the demo, is what keeps the production bill inside the business case.
Cloud FinOps governs resources whose cost scales with something finance already tracks, such as compute hours or storage. AI cost governance, sometimes called FinAIOps, governs a resource whose cost per task can swing by a large multiple depending on what the model decides to do. The levers are different. Cloud FinOps uses reserved capacity and right-sizing. AI cost governance uses per-task budgets, model tiering, and automatic termination. A bank that applies only its existing cloud cost tools to agents will see the spend but will not be able to stop it, because the control point is inside the agent run, not in the infrastructure bill.
Instrument before you optimize. Put per-task metering on every agent in the pilot so the real cost distribution is visible, then set the first hard caps at the 95th percentile of that distribution. This single step converts cost from a surprise on the invoice into a number the system enforces. Everything else, model tiering, caching, and budget-aware routing, follows from the data that metering produces.
Ableneo shipped 34 AI projects to production in 2025, 94% of them using large language models, with 4 of 5 projects reaching production rather than stalling in a pilot. Running AI in production is where cost governance stops being theory. We design cost limits into the agent at the same point we design its data access and its audit trail, so a bank approves a number it can hold. For the decisions behind getting an AI system to production and keeping it accountable, see the Ableneo AI transformation FAQ.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.