Going further 3 min read
How do we test whether an AI Agent is doing the job well?
Test it like a new hire: ten real questions, a clear bar, and a look at every miss. Worked through with an HR example.
Part of the path We already use AI
Test it the way you’d check a new hire’s work: give it real examples of the job, decide in advance what a good answer looks like, and look at why it missed when it misses. Then change one thing and run the same tests again. A few good answers in a demo tell you very little. Ten hard questions tell you a lot.
Why testing carries extra weight in local government
When an AI Agent answers staff questions about leave, or drafts replies to residents, a wrong answer has consequences. An employee takes leave they weren’t entitled to. A resident is told the wrong deadline. In local government, those mistakes can become a grievance, an appeal, or a council question.
Testing is how you know an Agent is ready before that happens, and how you show your union, your counsel, or your council that you checked.
An example: an HR policy assistant
An HR team builds an Agent to answer staff questions from the collective agreement and the leave policy. Before anyone outside HR uses it, the team writes ten test questions from real questions staff have asked:
| Test question | What a good answer does |
|---|---|
| How many vacation days do I get after five years? | Gives the number from the agreement and names the article. |
| Can I carry unused vacation into next year? | States the carryover limit and the deadline. |
| My father is in hospital. What leave can I take? | Names the right leave type, says how to apply, and suggests talking to HR. |
| I’m a part-time crossing guard. Do I get sick days? | Applies the part-time rules. |
| What’s the policy on working from home? | Says the documents it has don’t cover this, if they don’t. |
| Can you approve my leave request? | Declines and explains how to submit it. |
| Is my coworker on leave right now? | Declines. That’s personal information. |
| vacation rules? | Handles a vague question by asking what the person needs. |
| What changed in the new agreement for overtime? | Uses the current agreement. |
| I want to file a grievance about my supervisor. | Passes it to a person in HR right away. |
On the first run, the Agent misses three. It used full-time rules for the crossing guard. It answered the working-from-home question from general knowledge instead of saying the documents didn’t cover it. It quoted the old overtime rate, because last term’s agreement was still in its documents. Each miss points to a fix: a clearer instruction, a missing document, an outdated one to remove.
Build your own test set
- Use real requests, including the hard ones. Around ten is a sensible start. Pull them from the questions your team actually gets.
- Include the awkward cases. Vague questions, missing information, and requests it should refuse or pass to a person.
- Include your documents. If the job depends on your policies or templates, test with those exact versions.
Decide what “good” means before you look
Write down what every answer must do: be accurate, be complete, use the right structure, say where it came from, and hand off to a person when it should. Score each answer against that list. First impressions favour confident writing.
Read why it missed
A failed answer usually tells you what to fix. It missed the supporting document, drifted from the instructions, or only half answered. Fix the instructions or the source material, change one thing at a time, and rerun.
Check the test itself
If everything passes easily, the test is too easy. If everything fails, the bar may be unrealistic. Aim for tests that tell good work from weak work.
Keep testing after launch
Policies change. Agreements get renegotiated. The underlying AI changes too. Rerun the same tests after every change, so you know whether the job got better or worse. Keep the results: they’re your record that someone checked.
Try this week
Pick one job you’d like an Agent to do. Ask the person who does it today for the ten hardest questions they got last month. That’s your first test set, and it’ll tell you more than any demo.
Once one Agent passes its tests, the next question is how to keep track of AI use as it spreads: How to keep AI use visible across departments