More than the model bill suggests, and most of it is people's time. Running AI agents in a bank costs model usage, integration with core systems, testing and validation, the staff who review held work, record storage, and revalidation each time something changes. The model bill is usually the smallest line. Oversight and evidence are the largest.
What are the cost lines?
A bank that budgets only for model usage will find the real cost in other departments' time. Break it down into lines you can track.
| Cost line | What it covers | What makes it rise |
|---|---|---|
| Model usage | Calls to the model for each task | Long documents, many steps per case, retries |
| Platform and hosting | Running the agents, the record, the controls | Number of agents and capabilities |
| Integration | Connecting to core banking, payments, case systems | Old systems, custom interfaces, security review |
| Testing and certification | Building test sets, replaying, scoring | Number of capabilities and case types |
| Validation | Independent review by model risk | Agents in the model inventory, frequency of change |
| Human review | Staff who approve held items and check samples | Low depth, high hold rate, poor drafts |
| Record keeping | Storing every verdict and its evidence | Retention rules, volume |
| Change | Revalidation after model, source or process changes | How often things change |
| Incidents | Investigation and remediation when something goes wrong | Weak controls, wide scope |
The general version of this breakdown is in how much it costs to run an AI agent, and the argument that the cost of an agent is not the model bill applies with more force in a bank, where review and evidence are required, not optional.
Why is human review the biggest variable?
Because depth decides who does the last step. An agent at depth 2 drafts and a person sends, so every case still costs a person's time, though less than before. An agent at depth 3 completes routine cases on its own, so people only see the exceptions.
That makes the hold rate the number to watch. If most of an agent's work is held for a person, the bank pays for the agent and still pays for the review. The fix is not to raise depth by configuration. It is to earn depth on evidence, one capability at a time, so the routine cases move to depth 3 and people keep the hard ones.
What does a lapse or failed test cost?
A failed test should drop the agent to read-only the same day. That is the right control, and it has a cost: work the agent was finishing goes back to people at short notice. Banks should budget for that capacity. The principle is set out in a failed run should cost depth the same day. The cost of not doing it, an agent acting without valid evidence, is far higher.
Certificates that expire carry a similar planned cost. Retesting on a cycle is a recurring line, not a one-off.
An illustrative example
Take a bank running an agent that prepares responses to card disputes. In its first months it works at depth 2: it drafts every response and a handler reviews and sends it. Handlers save time, but every case is still touched.
After testing on the bank's own disputes, the routine category, small amounts with clear merchant evidence, is certified for depth 3. Handlers now see only the complex and high-value disputes. Model usage barely changes, but the review line falls sharply because people stop touching routine cases. The bank's real saving came from evidence, not from a cheaper model.
Finance teams running their own agents see the same pattern; compare how much an AI finance agent costs.
How should a bank plan the budget?
- Cost per case, not per month. Divide the full cost, including review time, by cases completed. Compare with the cost of the same case handled by staff.
- Track each line separately. A simple AI agent cost tracking template keeps model usage, review time and change costs visible side by side.
- Budget for testing as a recurring cost. Every model update, new source or process change means retesting.
- Reserve capacity for step-down. Know who picks up the work if the agent drops to read-only.
- Measure depth gained. The share of cases completed at depth 3 is the clearest sign that cost is falling for the right reason.
What stays with people, whatever the cost?
Cost pressure should never push decisions to the agent that belong with people: credit decisions on individuals, suspicious activity reporting, payments above the bank's limit, and bank-initiated account closures. Those stay with named staff. The wider limits are in AI agents in banking: what they may and may not do.
Questions people ask
Is a cheaper model a good way to cut costs?
Only if it passes the same tests. A cheaper model that fails more cases pushes work back to people and costs more overall.
Should agent costs sit with IT or the business?
With the business that owns the process. The savings and most of the costs, especially review time, land there, so the decision about depth should too.
Do agents cost more in a bank than elsewhere?
The model and platform costs are similar. Validation, record keeping and the depth of oversight are higher, because the rules and the consequences are.