6 min read

Why AI Is Breaking Your Definition of Done

AI doesn't change what Done means. It raises the cost of proving it. Three things to add to your Definition of Done for AI-written code. Part 3 of 3.

Part 3 of 3. Part 1 is Why AI Won’t Speed Up Your Scrum Team (The 1999 Trap). Part 2 is Code Got Cheap. Accountability Didn’t.

Short answers

AI doesn’t change what Done means. It raises the cost of proving it. Parts 1 and 2 covered where the extra code piles up and what an AI-native delivery loop looks like. This video asks who decides when AI-written work is Done, and what a team should add to its Definition of Done so the answer is still “the Scrum Team.”

It starts with Stan TheeMan and his two charts: Team A did 400 points, Team B did 50, so why can’t Team B be more like Team A? My answer is to add a zero to every estimate. Story points were never the work, and now that AI makes output cheap, outcomes matter more than output. There is no outcome until the work is Done. Then a Chicken skips the Product Owner and says the AI can build it in an afternoon. Tab, tab, merged, lunch. That isn’t Scrum.

The video uses the Friday, November 13, 2026 FAA ground stop as a reminder of what a change can cost. The FAA said there was “zero evidence of anything AI-related.” Officials tied it to routine maintenance at a remote site in Arkansas, and the root cause was still under investigation when I recorded. It fits the pattern I see most in postmortems: somebody made a change. Google’s DORA research found that more AI meant more throughput and more instability.

Don’t write a second, lower bar for AI. Make the one bar you have explicit. Add three things to your Definition of Done: a named human owns every change and can explain it, every change says what was asked and why, and a person writes or reviews tests for your system’s edge cases. Stan can count points. Your Scrum Team defines Done.

Video transcript

Lightly edited from the video’s captions: timestamps removed, punctuation added, Scrum terms linked to the Guide at first mention.

Who decides Done?

If your team uses AI to write code, your Definition of Done is about to break. AI doesn’t change what Done means. It raises the cost of proving it. Stick around: by the end, you’ll have three things to add to yours.

The points trap

Stan TheeMan walked into my office with two charts. “Team A did 400 points this Sprint. Team B did 50. Why can’t Team B be more like Team A?”

I’ve heard that question for twenty years, and I stopped teaching story points in my CSM classes over a decade ago. My answer is a little sarcastic. Team B, add a zero to every estimate. Congratulations. You just did 500.

It’s kind of a joke. But it starts the right conversation. Story points were never the work.

Now AI is in the room, and Stan thinks output just got cheap. He’s right. But output isn’t what counts. Outcomes over output, as Josh Seiden put it. And there’s no outcome until the work is Done.

Who gets to ask?

On a lot of teams, Done looks like this. A Chicken, someone involved but without skin in the game, skips the Product Owner and goes straight to the team. Quote. The AI can build this in an afternoon. Just do it. Unquote. Tab. Tab. OK. Merged. Lunch.

That’s not Scrum. The Scrum Guide says the Product Owner is accountable for maximizing the value of the product. And those wanting to change the Product Backlog can do so by trying to convince the Product Owner. Every Chicken. Stan included.

And Pat, our developer, gets the call when it breaks.

The FAA ground stop

On Friday, November 13th, 2026, the FAA stopped every departure in the country for more than four hours. Over 14,000 flights were canceled in two days, and it took all weekend to get planes and crews back in place, less than two weeks before Thanksgiving. Now imagine that happening the Sunday after Thanksgiving. Nobody has a final price tag yet.

What happened? Routine maintenance. A contractor was swapping old copper wire for fiber at a remote site in Arkansas, and replacing a router. The new router started flooding the network: about 1,000 attempts to log on every four to five seconds. The FAA says there’s “zero evidence of anything AI-related.” As I record this, investigators are still asking: installer error, or a faulty router? When the full root cause comes out, go read it.

I’ve read, and helped write, many postmortems. This one fits the pattern I see most: somebody made a change. And untangling a bad rollout can take longer than doing it right the first time.

Google’s DORA research, that’s DevOps Research and Assessment, found something similar with AI. More AI meant more throughput. And more instability.

I’ve seen this movie before

I’ve seen this movie before. Mid-nineties, IBM in Boca Raton. I worked on the team porting OS/2 Warp from Intel to the PowerPC platform. There were no online updates. We actually cut a Golden Copy in Boca and someone physically flew it to IBM’s processing center in Denver for production.

The code worked. It shipped. It was a genuine flop, and production support was buried in defects, no matter how done we thought it was.

That helped shape how I think. It’s in The Enterprise Unified Process, the book I wrote with Scott Ambler, and in Implementing Scrum since 2004, with nearly five thousand people trained. And today I’m Chief Product Officer at Edvizion, still shipping software.

Scott’s new book, Not Just Data, makes the case that continuous enterprise data is what feeds an organization’s AI. My endorsement says it shows leaders how to clear the data debt holding their AI initiatives hostage.

Quick question. How is your team handling AI-written pull requests right now? Tell me in the comments.

My own backyard

My own website runs on code that Claude helps me write. Private repo. Astro. GitHub. Cloudflare.

On my How I Use AI page, I admit I don’t always read every line the same way. That’s okay for me. If my site breaks, nobody’s flight gets grounded.

Now apply that to your world. Payments. Patient records. Air traffic.

Three things to add

The Scrum Guide doesn’t mention AI. It does say this: quote. “The Definition of Done creates transparency by providing everyone a shared understanding of what work was completed as part of the Increment.”

So don’t write a second, lower bar for AI. Make the one bar you have explicit. Here’s what I’d add.

One. A named human owns every change and can explain it in their own words. If you can’t explain it, it’s not done.

Two. Every change says what was asked, and why. Review checks the code against that intent, not just the syntax.

Three. Tests for your system’s edge cases, written or reviewed by a person. Not just tests the AI wrote to grade its own homework.

Still standing

If you’ve been doing this a while, you may wonder why they still need you.

Here’s my answer. You’ve seen production. You know what Done means when it’s real. That’s the hard part to automate.

Still standing after all these layoffs? Some food for thought. Every review you do and every fix you make teaches the model something. What are you teaching it? And who will decide what Done means this time next year?

Stan can count points. Your Scrum Team defines Done. That’s how output becomes an outcome.

Get the checklist

Want this checklist as a one-page PDF? Join my Saturday morning Implementing Scrum email at michaelvizdos.com/focus, and it lands in your inbox right away. Twenty-five hundred plus leaders already receive it. Thank you.

Focus. #deliver

Previous: Part 2, Code Got Cheap. Accountability Didn’t. shows what an AI-native delivery loop looks like. Part 1 is Why AI Won’t Speed Up Your Scrum Team. For the short answer on the checklist, see How should the Definition of Done change when developers use AI?

Get the free Study Kit

Full audio book + printable PDF. Includes the Saturday morning cartoons.

Get the Free Study Kit
Michael Vizdos

Written by Michael Vizdos

Certified Scrum Trainer® since 2006. About me · LinkedIn

35+ years working with teams around the world on Implementing Scrum. Found this useful? Feel free to share it with your team, or send feedback. I read all of it.

I sometimes use AI tools to brainstorm, draft and create, including images and video visuals. How I use AI.