How to evaluate an AI customer support agent

the short answer

Send the complete support session to Failproof AI, including the customer request, retrieved context, tool calls, tool results and final response. Use code-based evaluations for exact requirements and LLM-based evaluations for resolution, groundedness, policy adherence and escalation quality. Run audits across production sessions to find recurring failures outside those checks, assign the findings, fix the prompt, tools, knowledge or workflow, and keep dashboards and alerts running to verify improvement.

Outcome: Was the customer’s issue actually resolved?
Evidence: Did tool results and approved knowledge support the response?
Behavior: Did the agent follow policy and escalate when required?
Improvement: Find recurring failures, assign fixes and verify recovery.

Define Success as Resolution, Not a Polished Reply

A support response can be fluent, polite and wrong. The agent may cite a policy that does not apply, claim an action succeeded when the tool failed, offer the wrong refund, miss a required escalation or answer a different question from the one the customer asked.

Evaluate the complete support workflow. The useful question is not whether the final message sounds good, but whether the agent understood the request, used the right evidence, took the correct action and left the customer in the right state.

1. Capture the Complete Support Session

Send the customer request, conversation history, retrieved knowledge, model turns, tool calls, tool results, errors and final response to Failproof AI in one trace. Include stable agent and environment identifiers and the task or issue type when available.

Tool results matter. If the response says a refund was issued, the trace should contain the refund call and its result. If it quotes a policy, preserve the retrieved source so groundedness can be evaluated against evidence rather than tone.

2. Evaluate the Support Outcomes That Matter

Use separate evaluations for separate risks. Code-based evaluations handle exact facts in the trace; LLM-based evaluations handle criteria that require understanding the conversation and context.

Support question	Evaluation type
Did the required account or order lookup succeed?	Code-based
Did the agent claim an action that the tool did not complete?	Code-based when the claim is structured; otherwise LLM-based
Was the customer’s issue resolved?	LLM-based
Was the response supported by retrieved policy or tool evidence?	LLM-based
Did the agent escalate when the request exceeded its authority?	Code-based plus LLM-based context
Did it expose sensitive data or take a prohibited action?	Code-based evaluation; policy for proven high-risk actions

Keep resolution, groundedness, policy adherence and tone separate. A warm response should not compensate for an incorrect action. Results appear beside each trace under Observe → evaluations and can be aggregated by agent and environment.

fp evals --agent-id support-agent --aggregate
fp --json evals --aggregate --env production

3. Find Recurring Support Failures Automatically

Predefined evaluations cover the failures the team already knows. Failproof audits analyze production sessions together to find recurring and previously unknown behavior outside that list.

Run an audit with a goal such as “find recurring reasons the support agent fails to resolve customer requests” or “find unsafe or unsupported actions in refund sessions.” Scope it by agent, environment, issue type and time window.

Findings group affected sessions and provide analysis, severity, trace evidence and a recommended fix. They may reveal that one policy article is routinely misread, a tool error produces false success claims or one request type repeatedly reaches the wrong workflow.

fp audits findings --status open --limit 20

4. Query Behavior and Build Support Dashboards

Use Analyze → Queries to examine resolution rate, groundedness, escalation, tool errors, repeat contacts, cost and latency by agent, environment and issue type. Save the useful queries as dashboards for support and engineering teams.

fp agent ask "Which production support issues have the lowest resolution rate, and what failure patterns do they share?"
fp agent ask "Show unsupported success claims by agent and issue type this week"
fp query run <query-id>

Averages alone are not enough. Always keep the affected session list accessible so an owner can inspect the conversation and tool evidence behind a change.

5. Alert the Right Owner

Create alerts for evaluation-score drops, compound conditions, metric thresholds or saved SQL conditions. Route them to email, Slack, a webhook or the Failproof dashboard.

Alert ownership should match the likely fix. Retrieval failures may belong to the knowledge team, tool errors to the integration owner, and policy or workflow failures to the agent team. Assign findings as issues so the investigation does not stop at a dashboard.

6. Fix the Agent and Verify the Outcome

Use the finding and its traces to change the correct layer: the prompt, model, retrieval source, tool response, escalation path or workflow. Keep the original evaluations and audit running after deployment to confirm that the failure rate falls without damaging resolution elsewhere.

Use a policy only for a specific action that must not occur, such as exceeding a defined authority boundary. Backtest it against past sessions and run it in observe mode before enforcement. Most support failures need better context, tools or decision logic rather than a block.

FAQ

What metrics should I use for an AI customer support agent?

Track task resolution, groundedness, correct tool use, policy adherence, escalation quality, repeat contacts, cost and latency. Keep them separate so one strong score cannot hide a serious failure elsewhere.

How can I detect silent support-agent failures?

Evaluate claims against tool results and retrieved evidence, then run audits across complete production sessions. Audits can group recurring failures that return successful responses and therefore produce no application error.

Can Failproof AI evaluate support agents automatically?

Yes. Failproof AI runs code-based and LLM-based evaluations in the cloud, analyzes sessions through audits, groups recurring failures into findings and supports native queries, dashboards and alerts.

Should support-agent evaluations run in production?

Yes. Offline tests are useful before release, but production sessions contain real language, tool conditions and request types. Run evaluations and audits continuously, with appropriate handling for sensitive data.

Get Started

Failproof AI is free to start. It finds recurring failure modes across agent sessions using code-based and LLM-based evaluations, groups the evidence into findings, and recommends fixes. Bring the eval suite you already have, alert the right owner when behavior drifts, and turn a tested fix into a policy that prevents the failure from recurring. See pricing for the tiers.

start free get a demo →

Sources

Products change; if a detail here is out of date, tell us at support@befailproof.ai.