How do I quality-check my AI agents?

Somebody in my masterclass asked how I make sure my agents get it right. An AI agent sounds equally confident when it is right and when it is wrong.

None of what follows needs you to be technical.

1. Agents draft, I approve

None of my agents sends an email, publishes a post or charges anybody without me. They prepare the work. I read it and decide. My automations send only what I have already approved, like booking confirmations. That one rule removes the worst mistakes before they can happen.

2. A second agent checks the first

Every post and page is checked before a stranger reads it. One agent writes. A different agent reads it against my records and flags every claim it cannot find a source for. The writer never marks its own homework.

3. Test every connection with a real write

When I connect a tool, the first test is to write something real, not just read. A connection that can only read looks as if it works, right up to the moment you need it to do something.

4. Read it back

After an agent saves something, I have it read the result back. Once, a draft saved with a title and no text, and the tool reported success. Only reading it back caught it.

5. Fix the brief, not just the output

When an agent gets something wrong, I correct the work and then add the fix to its written brief, so the same mistake does not come back next week. Over time, the brief becomes the quality control.

Before you connect anything to your data, read should I trust AI agents with my client data and my work?

Your next step

If you already run a practice with clients and want somebody to look at your own setup, that is the AI Deep Dive: CHF 750, ninety minutes, one to one, on how you run your practice today. You leave with a written, personalised strategy for where AI can take work off you, and your options for doing it. Not running a practice yet? Start with the free Systems Analysis.

Questions people ask

How do I know if my AI agent is making mistakes?

Have a second check read its work against your own records, and read back anything it saves. Mistakes that sound confident are the ones to look for.

Should AI agents send emails without approval?

Not to clients. Let agents draft, and keep the send with you.