Insights · October 29, 2026 · 4 min
Three questions to ask about every deployed AI agent
What does it really do, is it worth what it costs, and how far can one mistake go? How to answer with evidence.
A human error is one case. An agent error is every case it handled before someone looked: the same defect, repeated at the agent's volume. That is the strength of AI, which deploys a solution everywhere at once, and it is also its exposure. Three questions let you take stock of any agent in production.
1. What does it actually do?
Compare what the agent was authorized to do with what it does in production, using logs and execution traces, not the design document. Scope drift is the norm: a newly connected tool, an edited prompt, an updated model. The useful question is whether anyone is watching it.
2. Is it worth what it costs?
Add up the real cost to run: compute, human oversight and incident cleanup. Compare it with the cost actually displaced, then review a sample of outputs by hand against what a person would have produced. The business case from six months ago is not a current number.
3. How far can one mistake go?
Estimate how many cases the agent touches, how severe each one is (reversible or not, routine or critical), and how long it would take a person to notice. It is the RTO/RPO question applied to an agent: how much can go wrong, and how fast can it be stopped?
Then match controls to the risk
Approval before action, staged rollout, sampling review, a tested kill switch, a named owner: check that these controls match what the blast radius demands, not just what was convenient to build. Our downloadable audit scorecard helps you record the answers and repeat them every quarter.