ALGS

All articles

Product3 min read

Define "done" before anyone prompts

The code was the easy part. What AI agents couldn’t decide was what good output looks like.

I rebuilt my company’s website with AI coding agents. The code was the easy part.

The agents wrote pages, scripts and a catalogue of 63 releases. What they couldn’t do was decide what "done" meant.

<h2>Acceptance checks, not opinions</h2>
<p>So I wrote acceptance checks, the same way I did as a Product Owner at Avanade: text contrast, banned hype words, one primary button per page. Every change runs against them. When a check fails, the agent fixes it before I see it.</p>

<h2>The step most AI rollouts skip</h2>
<p>Teams get the tool, nobody defines "done", and output quality drifts. Then the tool gets blamed for a missing process.</p>

<h2>If you’re rolling out AI in a team this quarter</h2>
<ol>
  <li>Write down what good output looks like before anyone prompts.</li>
  <li>Make the check automatic, not a manager’s opinion.</li>
  <li>Measure usage and quality, not licences bought.</li>
</ol>

<p>How does your team define "done" for AI-generated work? <a href="https://www.linkedin.com/in/gillesalexandre/" rel="noopener">Tell me on LinkedIn</a>.</p>

The same problem on your side?

A 20-minute call to talk it through, no pitch deck.

Reply within 24 working hours · first call free · no pitch deck

Next article: Measure usage, not licences