Companies don't have an AI policy. Their people are already using AI anyway. This episode is how to think about it: the mindset, the principles, and what's green, yellow, and red.
Conversational, back and forth, concrete examples. Frame at the top, then talk. Security is its own episode later.
Mindset: not much changes
Less changes than you think. You never wanted company data in software you don't control, or people sending work they haven't reviewed. AI just makes the old bad things easier to do.
People don't know how to think about AI. The media narrative is doom. You can't rally a team around doom.
Blocking AI doesn't work either. Your people are already on free accounts on their phones. Now you have a real security problem.
Both are true: not much changes, and this is as big as the internet.
Address replacement head on. "Are we getting rid of people?" is where most minds are stuck.
Don't forget your product. The client work still has to happen while you build.
Principles
Don't create what you don't understand. You can't own what you don't understand. This covers "don't build software" without saying it.
Go slow. 2x with the same quality bar, then 4x. Not 10x with 36 agents.
Just because you can doesn't mean you should.
Manual before automated.
Be iterative. No one-shotting. Work with the AI, don't fire and forward.
Yellow. Fine with something in place: a read-through, a second opinion, a checklist. Sending anything. Publishing anything that touches sensitive info.
Red. Sending something you haven't read. Bespoke systems that replace real ones. Skunkworks tools nobody owns. Tasks that need to be deterministic.
One more: require people to work in central systems. Everyone scattering into their own workflows and skills breaks the company later.
Examples on deck
The Matzner email: "I was the hypocrite in my own AI policy." Use AI, don't build with it.
The vibe-coded Jira replacement that got rolled back.
The doc example: writing went from 4 hours to 5 minutes. The review moved to the back, and skipping it is where slop comes from.
The 36-agents confession. Quality drops off and is hard to win back.
Fred's spreadsheet: the skunkworks tool half the company uses and nobody maintains.
Lower thirds
Shadow AI. AI tools employees use outside the company's approved systems and visibility.
One-shotting. Firing a single prompt and pushing the output forward instead of iterating.
Skunkworks. A project built in the background, outside leadership's view.
Deterministic. A task that must produce the same correct result every time. Not a place for AI judgment.
The close
Go try this: sit down with your leadership team and talk about how you want AI used, what people are worried about, and what your policy should cover.
Offer: the AI handbook template, link in show notes.