Start with the people around the agent, then the model bill. Rework, reviewer time and cases the agent was never going to finish usually cost more than tokens. After that, trim what the agent reads, stop runaway loops and route simple capabilities to smaller models, recertifying each change. Never save money by cutting the record or the review.
Where does the money actually go?
Most teams look first at the model provider's invoice because it arrives as a single line. It is rarely the largest cost. The full picture includes reviewer minutes on every draft and held action, rework when a draft is rewritten, integration upkeep, certification runs, and the time spent on incidents. Our argument that the cost of an agent is not the model bill sets this out in full.
You cannot reduce what you have not measured per agent and per capability. If you only know the monthly total, start with tracking the cost of each AI agent and a shared cost tracking template before you change anything.
Which levers reduce cost, and what does each risk?
| Lever | What it saves | What it can break | Check before keeping it |
|---|---|---|---|
| Improve drafts so reviewers edit less | Reviewer and rework time | Nothing, if done through the brief | Edit rate falls on live work |
| Triage out cases the agent will escalate anyway | Model calls and wasted reviewer reads | Cases wrongly triaged out | Sample of triaged cases |
| Trim the context to the sources each task needs | Tokens per task | Grounding, if a needed source is removed | Rerun certification |
| Set step and retry limits | Runaway loops | Hard cases cut short | Escalation rate stays stable |
| Batch work that is not urgent | Peak usage charges | Service levels | Time to complete per batch |
| Route simple capabilities to a smaller model | Model cost per call | Refusal and escalation behaviour | Full recertification |
| Retire agents nobody uses | Licence and upkeep | Little | Owner sign-off |
The table is ordered on purpose. The top rows are cheap to try and hard to get wrong. The bottom rows change what the agent is, and must go back through certification.
How do you run a cost reduction safely?
- Analyst: breaks cost down per agent and capability, including reviewer minutes. Evidence: a monthly cost line per capability.
- Process owner: picks one lever and states the expected saving and the risk. Evidence: a short change note.
- Agent, in a test environment: runs the change against the certification case set. Evidence: scores against the floor, gate by gate.
- Process owner: approves or rejects on the scores, not on the saving. Evidence: the decision, filed with the scores.
- Analyst: checks live results after a month against the expected saving. Evidence: before and after cost lines.
One lever at a time. Two changes in the same week make it impossible to tell which one moved quality.
What does a sensible cost cut look like?
Say an insurer's complaints team runs an agent that reads every new complaint, the policy, the claim history and a year of contact notes, then drafts an acknowledgement and a summary at depth 2.
The analyst finds that a third of model spend goes on reading contact notes that the summary rarely uses, and that reviewers rewrite most acknowledgements to change the tone. The team does two things in sequence. First, they rewrite the brief's tone guidance with examples from approved letters; the edit rate falls and reviewer time with it. A month later, they limit contact notes to the last 90 days and rerun certification. Grounding still clears, so the change stays.
They also test moving classification to a smaller model. It clears every gate except refusal, where it labels unclear complaints with confidence instead of flagging them. That change is dropped.
What should you never cut?
- The filed record of verdicts. It is your evidence for every decision.
- The daily review of the held queue.
- Certification runs, even when they feel repetitive. A lapsed certificate costs more than the run.
- Drift monitoring. Cheaper configurations often drift faster.
For planning the overall spend, see budgeting for running AI agents and the guide to deploying an AI agent.
Questions people ask
Is a smaller model always cheaper overall?
Not always. If it produces more escalations or drafts that need heavier editing, people absorb the difference. Compare total cost per completed case, not cost per call.
Should agents share one model to get volume discounts?
Pricing is a commercial matter, but a shared model should not mean shared certification. Each agent's capabilities are certified on their own cases with the model they use.
How often should we review running costs?
Monthly for each live agent, and after any change to model, sources or depth. Cost often moves when hold and escalation rates change.