From the PMI talk · Partnering with AI

AI Experiment Playbook

1. Find one real task

   Something already on your plate this week — not a hypothetical. Prefer work you would do anyway.

2. Try AI on it

   Approved tools only. Give context: audience, format, and what “good” looks like. One clear ask beats a vague “help me.”

3. Review the output

   Read it like a junior teammate’s draft. Cut what’s wrong, fluffy, or overconfident. Fix what’s weak before anyone else sees it.

4. Measure value

   Did it save time? Improve clarity? Or create cleanup work you wouldn’t accept? Be honest — speed without trust is not value.

5. Keep or discard

   If it helps and you can stand behind it — keep. If not — stop. No sunk-cost loyalty to a tool.

After one experiment — score it

• Time: faster, about the same, or slower after edits?
• Quality: would I put my name on the final version?
• Trust: could I explain every claim in the room?
• Repeat: would I use this pattern again next week?

If you can’t say yes to quality and trust, discard — even if it felt fast.

What “keep” looks like

• You have a prompt or pattern you actually reuse
• You still review before send — every time
• You know when not to use it

When an experiment fails

• Wrong task — too vague, too high-stakes, or no source material
• Weak prompt — no audience, format, or “don’t invent” rule
• Skipped review — the output went out unedited
• Wrong tool — unapproved, or the wrong fit for the job

Failing once is data. Sending bad output is a craft problem.

This week

• One Copilot (or approved) task on real work — you check the whole thing
• One meeting → action table you actually send

This month

• Pull a story from a few different sources
• If you build something, start with a written spec

Need prompts? Use the PM AI Prompt Library. Need the boundary line? See What Stays With the PM.

Fix it or drop it. Don’t send a draft you haven’t read.

Accelerate the work. Don’t dilute the craft.

← All take-home resources