Operate
If you cannot score a completed case, you do not have an agent program. You have a demo habit.
Fluent output is easy to like in a meeting. Operators need to know whether the agent finished the right work, at an acceptable cost, with failures they can explain.
How I see it
AI agent evaluation
Agent evaluation should look like operations, not like a chatbot leaderboard. The unit is the case: was the outcome correct, complete, timely, and within policy? Did it take the allowed actions? Did it escalate when it should have?
Build an evaluation set from real historical work. Include the ugly cases. Score the first release before users see it, then keep scoring as prompts, tools, and models change.
Reliability is part of evaluation. An agent that is right 90% of the time and silently wrong 10% of the time is not “pretty good.” It is an unpredictable employee.
A second model that scores the first is not ground truth. If the policy and the evaluator share the same blind spots, the loop will confidently accept a wrong trajectory. Cost belongs on the same scorecard: a perfect case that costs more than the human path is not leverage.
Common mistakes
What teams usually get wrong.
Vibe reviews
A few people saying it looks good is not a quality system.
Accuracy without the business object
A correct paragraph can still produce a wrong CRM update.
Evaluating only the happy path
The cases that create executive pain are the exceptions.
A useful diagnostic
Five questions before you fund the work.
Do you have a labeled set of real completed cases?
If not, start there before changing models.What does a pass look like in the system of record?
The score has to land on a business object.Do you track silent failures separately from escalations?
Escalation can be success. Silent wrong write-back is not.Is cost per completed case on the same dashboard as quality?
Quality that ignores cost will optimize for expensive thoroughness.Can you re-run the set after every change?
If evaluation is a project, drift will win.
Economic model
Release gatequality on the case set + cost per case + escalation rate + critical-miss rate
A critical miss is a wrong action with material business impact. That number should stay near zero.
Three credible paths
How far should you go?
Do not force one solution. Choose the path the economics, the risk, and the organization can support.
Offline case scoring
Score historical cases before the agent touches live work.
You are still choosing whether to pilot.
History may not include the messiest current exceptions.
Supervised live evaluation
Run on a slice of volume with human scoring and a kill switch.
Offline results are strong enough to learn from real traffic.
Needs reviewer time and a clear sample.
Continuous production evaluation
Automatic checks, sampled review, and regression tests on every change.
The agent is in the operating path.
Requires ownership. This is operations, not a one-time test.
When this is the wrong next step
Do not fund an agent here.
- There are no historical cases and nobody will create a score set.
- The team wants to ship because a stakeholder liked the demo.
- Success is defined as “people are using it” rather than completed work.
A useful next step
Bring one workflow. Get guided into production.
We guide the implementation, go deep on the technical path, and stay hands-on through operations — or tell you when a simpler answer is better.
Discuss an AI opportunity