AI Agent Performance Metrics: 15 KPIs for Customer Outcomes

An AI agent performs well when customers complete the intended task correctly and the business can support the work at an acceptable cost. A fast response is useful, but a fast response that promises an unavailable appointment still creates work for your team.

This guide gives business owners a menu of 15 key performance indicators, or KPIs. Start with four. Use the rest to explain a problem or assess a specific improvement.

What counts as a successful customer task?

Define success using evidence outside the conversation where possible. The agent saying “done” is a claim to verify.

Agent job	Completion evidence	Incomplete outcome
Book a service	Valid appointment recorded and confirmed to the customer	A preferred time collected without a booking
Answer a support question	Correct answer or completed remedy, checked against policy	Conversation closed while the customer still needs help
Prepare an itinerary	Plan meets the agreed dates, budget assumptions, and preferences	Attractive plan missing a required day
Qualify an inquiry	Required details are accurate and reach the assigned person	A name and email with no usable request

Keep two views: completion across all eligible requests, and completion without human assistance. The first measures service quality. The second helps explain staffing needs. Treat an appropriate handoff followed by resolution as a successful team outcome.

Which four metrics belong on the first scorecard?

Use verified completion, material error rate, handoff completion, and cost per completed task. Record the number of requests behind every percentage.

Illustrative weekly scorecard: a repair business receives 80 eligible appointment requests. Fifty reach a valid booking through the agent, 18 go to the office, and 12 end without a result. The office books 10 of its 18 requests. Overall booking completion is 60 / 80, or 75%; direct agent completion is 50 / 80, or 62.5%. These are examples of arithmetic, not performance targets.

If four customers needed a promised slot corrected, record those errors even when the corrected booking eventually succeeded. A completion rate alone would conceal that problem.

What do the 15 metrics mean?

Metric	Calculation or evidence	Decision it supports
1. Verified task completion	Completed eligible tasks / eligible task starts	Does the service work?
2. Direct agent completion	Tasks completed without staff / eligible starts	How much work can run independently?
3. Material error rate	Reviewed tasks with consequential errors / reviewed tasks	Which mistakes need correction?
4. Repeat contact rate	Resolved issues reopened within your defined window / resolved issues	Did the answer hold up?
5. Handoff completion	Handoffs resolved by staff / handoffs due for resolution	Is the customer reaching a person?
6. Unnecessary handoff rate	Reviewed handoffs within the agent’s approved scope / reviewed handoffs	Is useful automation being missed?
7. Abandonment	Eligible starts with no outcome or handoff / eligible starts	Where do customers leave?
8. Time to useful response	Time until relevant information or a usable next step appears	Is the experience responsive?
9. Time to resolution	Elapsed time from request to verified outcome	Is the whole service faster?
10. Customer satisfaction	Positive responses / responses to a consistent survey	How do responding customers assess it?
11. Customer effort	Repeated questions, steps, and reported difficulty	Is the interaction burdensome?
12. Human minutes per task	Staff handling, review, and correction minutes / tasks	Is workload actually falling?
13. Cost per completed task	Allocated service costs / verified completed tasks	Are the economics sustainable?
14. Completed sales or bookings	Recorded transactions, later adjusted for cancellations	Is commercial activity reaching an outcome?
15. Connection success	Verified successful actions / attempted actions	Are connected systems completing the work?

Define each denominator once. Separate test traffic, spam, and genuinely out-of-scope requests using a recorded rule. Do not exclude difficult eligible requests after seeing the results.

There is no universal good completion rate for every agent. A public policy explainer and an account-specific refund service have different access requirements and consequences. Set thresholds for the particular job, with stricter release conditions for consequential mistakes.

How do you collect useful evidence without a large dashboard?

A shared worksheet can be enough for a small pilot. Record a request identifier, task type, outcome, handoff reason, correction needed, human minutes, and confirmation reference. Use only the personal data needed for that review.

Review a sample of ordinary completions as well as every complaint and failed action. Complaints expose some errors; silent customers can leave without reporting one. Satisfaction results should include the response count because satisfied and dissatisfied customers may respond at different rates.

NIST’s AI Risk Management Framework provides broader guidance for evaluating AI in its intended context. This article’s scorecard is a practical editorial framework, not a NIST certification or an industry benchmark.

What should you do when metrics disagree?

Investigate the customer journey before changing the agent. Use these combinations as diagnostic prompts:

• Faster responses, slower resolution: check whether the agent is adding questions before handing off.
• More bookings, more cancellations: examine fit, availability, and the promises made during qualification.
• Fewer handoffs, more corrections: inspect whether the agent is guessing when it should stop.
• More completed tasks, higher total staff time: count review and reconciliation work that moved downstream.

Change one material part of the workflow at a time and record when it changed. Keep a small set of past failure examples to repeat after updates.

How does this apply to an Agentplace agent?

An Agentplace agent serves customers through its own interface and URL. It can use business knowledge, ask follow-up questions, change the page to show a plan or service options, and use connected tools to complete approved actions.

Measure that entire experience. For a travel planning agent, assess whether the displayed itinerary fits the client’s requirements and whether the final request reaches the advisor. Message count alone tells you little about the quality of the plan.

Start from a customer task template, define its completion evidence, and use the ROI framework when you are ready to connect outcomes to costs. Compare actual usage with the current pricing and limits.

For the next weekly review, bring three things: the four-metric scorecard, examples of unsuccessful requests, and one change to test. That is enough to turn measurement into a better customer service.