My best PO had a problem: his team had become too fast
He told me one morning, without irony. 3 minutes to produce an increment. 3 days to get it validated on the back-end side.
The engine is running at full throttle, the transmission stayed on the old tempo.
AI did not speed his team up. It made visible what was already slowing it down. The bottleneck did not disappear, it moved: the team did not manufacture fluidity, it shifted the traffic jam one notch along.
I see it every week in the field. An augmented team waiting on a validation stuck somewhere, and a handover that exists only in the governance slide shown at kick-off.
Transformation does not start at the prompts
This is the most misunderstood consequence of AI arriving in an organization. As long as producing was the limiting factor, accelerating production made the whole flow win. That is no longer true. When one step goes from 3 days to 3 minutes, the constraint immediately shifts to the next one — most often a validation, an arbitration, a queue between two teams.
Lean and the theory of constraints had established this well before AI. Optimizing one link does not make the chain win. What AI changes is the brutality with which the imbalance shows up.
If you want augmentation to hold, the move is simple to describe and uncomfortable to carry out: trace the real path of an increment, from idea to deployment. Note every waiting point. Identify who holds the next link, and go ask them face to face what actually slows them down.
A fast team inside a slow ecosystem does not build momentum. It burns out.
AI transformation therefore does not start at your people's prompts. It starts at the boundaries between teams — exactly where nobody holds a clear mandate, and exactly where the IAgile™ principles actually play out.
Three framing mistakes on a train already running
At the scale of an Agile Release Train, the same mechanism takes a recognizable form, and I keep seeing three framing mistakes the training courses never mention.
The first: framing the train on the org chart rather than on the value stream. The dependencies discovered at PI Planning are then the mechanical consequence of the chosen split, not a coordination accident. No dependency-mapping tool will change that.
The second: treating system and platform teams as internal suppliers. They become the bottleneck nobody anticipated, and accelerating the product teams merely lengthens their queue.
The third: believing that a train already running cannot be reframed. That one costs the most, because it turns a framing error into a trajectory.
They had launched the train, and still said "we pivot"
I saw an organization do it. Not out of fear of failure: out of respect for the field.
The vision was wrong. The people on the ground knew it. It took a year for that to travel up, be heard, and actually change something. That is not a failure in itself; it is what happens when you advance in silos and each product optimizes in its corner without seeing the whole game.
What is rare is what came next. It was the operational people who carried the message, not the managers, not the sponsors. And someone at the top had the courage to genuinely hear it. They were not afraid to lift the carpet, nor to pivot mid-launch, nor to say "we start again differently" rather than defend a dead direction.
Changing strategy along the way is not a weakness. It is the mark of an organization that learns faster than it clings. A year to sow — and at the following PI Planning, the collective energy in the room was visible to the naked eye.
Apparent chaos is not the problem
One objection comes up as soon as these subjects are opened: freeing the teams installs disorder. Everyone with their own agent, their own prompt, their own method.
Apparent chaos is often freedom expressing itself. That is not a bad thing: it is the sign of a team daring to create, explore and experiment in every direction. The real question is not how to suppress the ferment, but how to channel it into a coherent force at company scale.
This is where continuous improvement resumes its role, and it is more demanding than an end-of-sprint ritual: turning creative disorder into a trajectory without extinguishing what produces it. A governance that answers ferment with prohibition gets calm, and then nothing at all — the mechanism I describe in the missing link in AI governance.
What I look at first
In an organization that is accelerating, I no longer look at team velocity. I look at the waiting times between teams, and at the delay separating the moment information arrives from the moment someone decides.
That is where the lost weeks sit, and it is the only place where an AI transformation produces a durable effect. The rest — licences, tools, training — is only the entry ticket.