How Do You Measure an AI Employee's Performance?
Measure AI employee performance with six numbers: completion, failure, rework and escalation rates, time saved and cost per task. How to count them weekly.
Reviewed by WaveBrain Operations TeamReviewed 6 min read
The short answer
You measure an AI employee's performance by counting outcomes against its job, not how busy it looks. Track six numbers each week: completion rate, failure rate, rework rate, escalation rate, time saved and cost per task. Pull the counts from one activity log of everything it did, and check quality by hand on a small sample, such as 10 tasks a week.
An AI SDR that sent 600 emails last week looks productive on paper. Then you read 10 of them and find three went to the wrong contact and one promised a discount you don't offer. The send count went up and the work got worse.
That's the trap with AI employees (the software agents you've given a job, such as an AI SDR, a support agent or an n8n workflow). They produce a lot of activity, and activity is easy to count. This guide covers which numbers say whether an AI worker is doing its job well, and how a non-technical owner can collect them in about 20 minutes a week.
How do you measure an AI employee's performance?#
You measure an AI employee's performance the way you'd measure a new hire on a narrow job: write down what a finished, correct task looks like, then count how often it delivers that. Volume only matters once quality holds up.
At Asymbl, a company profiled by Salesforce in April 2026, each AI agent gets a job description, KPIs and documented feedback from the people who work with it. The same article cites Deloitte's State of AI in the Enterprise report, which found that 85% of companies expect to customize AI agents but only 21% have a mature governance model for them. Measuring is the part most teams skip.
Which numbers tell you if an AI employee is good?#
Six numbers cover most AI workers, whatever their job. Count them per AI employee, per week.
| Metric | How to count it | What it tells you |
|---|---|---|
| Completion rate | Tasks finished ÷ tasks started | Whether it gets through the work |
| Failure rate | Tasks logged as failed ÷ all tasks | Broken logins, bad data, tool errors |
| Rework rate | Sampled tasks a person had to fix ÷ tasks sampled | Whether the finished work is right |
| Escalation rate | Tasks handed to a human ÷ all tasks | Whether its scope fits what it can handle |
| Time saved | Tasks done × minutes a person would have spent | The value you're getting |
| Cost per task | Monthly tool and model costs ÷ tasks done | Whether it's cheaper than the alternative |
Set your own targets from the first two or three weeks of data instead of copying someone else's. A support agent that escalates 30% of tickets might be fine for billing disputes and too timid for password resets.
Why is activity a misleading way to judge an AI employee?#
Activity is misleading because an AI worker can raise every activity number while doing worse work. More emails sent, more tickets closed and more posts published look like progress until you check whether the emails reached the right people or the tickets stayed closed.
That's why rework rate matters most. It's the only number on the list that needs a human to read the output, and it catches the mistakes the AI worker can't see in its own work. Pick 10 finished tasks at random each week, mark each one "fine as is" or "needed fixing", and write down what was wrong.
How do you collect the numbers without a data team?#
You collect them from one activity log where every AI employee records each finished task with a status. Completion, failure and escalation rates come from counting entries by status for one agent, and the weekly sample for rework comes from the same list.
The log needs a few fields for this to work: the agent's name, a one-line title, a status of success, failure or pending, and a link to the result so you can open the email or ticket. Our explainer on what an AI agent audit trail is lists the full set. Time saved and cost per task come from outside the log, from your own estimate of minutes per task and your monthly bills.
How do you run a performance review for an AI employee?#
You run it like a short one-to-one with yourself and the person who owns that AI worker. Once a month, spend 15 minutes per AI employee on these steps:
- Reread its job description and the limits you gave it.
- Compare this month's six numbers with last month's.
- Read the three worst tasks from your rework samples.
- Decide on one change: clearer instructions, a new approval step, more access or less.
- Write the change and the date down, so next month you can see whether it helped.
If you also need a record for clients or an accountant, our guide to auditing AI agents in a small business covers that side.
What should you do when an AI employee underperforms?#
When an AI employee underperforms, narrow its job before you replace it. Most bad results come from a task that's too broad, instructions that leave room for guessing, or access to data it shouldn't use.
Work through the fixes in this order: rewrite the instructions with two or three examples of good output, add a human approval step before anything goes to customers or moves money, shrink the job to the part it does well, and only then swap the tool. If the rework rate stays high after two monthly reviews, retire it from that job.
Where does WaveBrain fit in measuring AI employees?#
WaveBrain is the activity log the counts come from. AI workers that can call an MCP server or send an HTTP request record each task with a title, details, the agent's name, the project, a status and a link, and you filter the timeline by agent, project or action type to see one AI employee's week.
WaveBrain calculates only the basics: its stats page shows actions, success rate and failures per AI worker. It has no performance scores or HR features, so for anything else you count entries yourself or ask an agent to read them through the list_actions tool and tally them. It only knows what your AI workers report, and closed products with no MCP, webhook or API step can't report to it yet. The free plan keeps 7 days of history, so do the weekly count before entries age out. For the bigger picture, see how to keep track of AI agent work.
Frequently asked questions#
What is the single most useful metric for an AI employee?#
Rework rate, the share of sampled tasks a person had to fix. It's the one number that measures whether the work was right.
How many tasks should I check by hand each week?#
Ten random finished tasks per AI employee is enough to spot patterns for most small teams. Check more for anything that talks to customers or touches money.
Should I compare AI employees with human employees?#
Compare them on the same task, using time saved and rework rate. Comparing overall output misleads, because AI workers usually handle narrower jobs.
Does WaveBrain give each AI employee a performance score?#
No. WaveBrain records what each AI worker did, with a status and a link, and you count from that. It has no scoring or HR features; its stats page only counts actions and failures.
How often should I review an AI employee's performance?#
Look at the numbers weekly and do a fuller review monthly. Review sooner after any change to its instructions, tools or access.