
AI is making delivery faster, and that's the headline everywhere. Teams can now build, test and release in a fraction of the time it used to take, and some organizations are making good use of it.
Most organizations only notice that cost once they run into it. The bottleneck has moved. For some teams, work now piles up downstream in code review and testing. For most, the bigger constraint sits upstream.
Twenty-five years ago, the constraint was production. Designing software was manageable, but building it took forever. Agile helped teams build faster, so the problem moved. Fifteen years ago, the bottleneck shifted to deployment. DevOps solved that, and the constraint moved further back again.
Now AI is speeding up both building and deployment. The bottleneck has landed on a step few organizations are equipped to handle.
Output, Outcome and Impact in Delivery Work
Many teams think about Agile in one dimension: output. They track how fast they ship and how many features they deliver. Velocity, story points and deployment frequency all measure output. That's table stakes now.
More mature organizations add a second dimension: outcome. They ask whether they're solving the problems they said they'd solve.
The third dimension is impact, and very few organizations work there. Impact asks whether this was the right problem to solve. It also asks which problems deserve the company's time and money. That's where the most important work sits now.

Impact work belongs in discovery, where a team figures out what to build before engineering starts. Discovery used to happen in strategy meetings and marketing departments. It sat at the management layer, where there was time to think. With AI speeding everything up, discovery has to move closer to the teams doing the work.
Sometimes discovery stays at the management layer, and engineering gets a finished spec. Then managers decide slowly while engineers wait to move fast. The faster engineering gets, the more that lag hurts.
Why Managers Need to Be Player-Coaches
Managers have to get close to the work now. Signing off on plans from a distance can't keep up with the pace AI makes possible.
Baseball offers a useful comparison. As of Opening Day 2026, 11 of MLB’s 30 managers—37%—had primarily been catchers during their professional playing careers. The catcher is the one player facing the whole field. Catchers call pitches, read hitters and track positioning on every play. That's close to the job of a manager.
A player-coach in a delivery organization works the same way. They stay close enough to the work to know what's possible. That lets them push back on a bad spec with confidence. They also coach their people to think about which problems they're solving. And they help them see how the wider system shapes those choices.
That takes real investment in developing people into general managers. Many organizations stopped doing this about thirty years ago.
HP was known for this for decades. It was famed as a talent factory. Top performers could earn full profit-and-loss authority over their own product groups. That kind of ownership teaches people how the business works as a system. They learn how the pieces connect and where the tradeoffs live. Outside Silicon Valley, most companies never really made that investment. They hired functional experts and expected them to stay in their lane.
With AI changing the speed of decision-making, people can't work this out alone. Organizations need people who see the whole system and make good calls quickly.
The Culture That Makes Discovery Work
Getting there takes deliberate cultural choices.
In The Geek Way, Andrew McAfee describes four cultural norms found in high-performing organizations. They are science, ownership, speed and openness. Science means settling disagreements by testing assumptions. Ownership means taking responsibility for results without hiding behind process. Speed means moving quickly. Openness means being willing to be wrong in front of others.
When all four norms are present, the organization learns from what it tries. People decide once they have enough information. They treat failure as information too.
That depends on psychological safety. In Right Kind of Wrong, Amy Edmondson separates two kinds of failure. Organizations should welcome intelligent failures. They should work to reduce basic, preventable ones.
An intelligent failure happens when you test an assumption, design well and execute properly, and it still doesn't work. Maybe the market doesn't want what you built. That result still teaches you something you couldn't learn another way.
A preventable failure comes from skipped testing or a careless mistake. Those are the ones to eliminate.
Many organizations treat every failure the same way and shut it down. That ends discovery. People who fear testing uncertain ideas will only propose ideas they know will work.
Knowing What's Worth Building Fast
AI speed has exposed where the constraint sits. Organizations can now move fast enough to see it. For most of them, it sits outside the engineering layer.
So the question to ask now is: Do you have the people, culture and decision-making to get good at deciding what to build?
If you don't, all that AI speed just means you'll ship bad ideas quicker, and because they aren’t tested, they’ll stay around not working for a long time.
This post draws from Episode AI Changes The Real Bottleneck of Definitely Maybe Agile, featuring organizational consultant Bernie Maloney. You can listen at definitelymaybeagile.com.
Frequently Asked Questions
Why hasn't AI coding sped up what my team actually delivers?
AI made building and deploying faster, so the constraint moved to deciding what to build. If discovery still happens slowly at the management layer, engineers end up waiting on decisions instead of code.
What's the difference between output, outcome and impact?
Output measures how much a team ships. Outcome asks whether the team solved the problem it set out to solve. Impact asks whether that was the right problem to solve at all.
What is an intelligent failure?
An intelligent failure happens when a team tests an assumption, designs and executes well, and the result still doesn't work. Amy Edmondson separates these from basic failures, which come from skipped steps or careless mistakes and should be prevented.
