There’s a specific moment I started recognising in agile transformations. Usually around month four or five. The teams were running sprints. The retrospectives were happening. Someone had bought a Jira licence. And the organisation, structurally, was exactly the same as it had been before anyone said the word agile.
Priorities still arrived from above, unclear and subject to change. Decisions that needed sign-off from three functions still took a month, because nobody had sorted out who was actually allowed to say yes. The problems teams raised in retrospectives got written on sticky notes, grouped into themes, and left there, because fixing them needed authority that sat above the team.
I watched this happen at enough places to stop being surprised by it. The vocabulary changed. Velocity, ceremonies, iterative delivery. The system didn’t. Leadership had new words for old habits, and the problems went nowhere.
That was agile. The same thing is happening now with AI, except faster and with bigger numbers attached.
The move is identical
Nobody’s running AI standups yet. But the structural move is the same one: adopt the practice, skip the change underneath it.
AI lands inside firms where decision rights are unresolved, data is scattered across functions that have no reason to share it, and governance exists as a document nobody reads. A model gets bought. Pilots run. A few produce interesting demos. Then the initiative stalls, or fades, or produces results nobody can measure because nobody defined what success looked like before they started.
If there’s a post-mortem, it usually blames the technology. Wrong model, not enough data, integration trouble. Sometimes that’s true. It’s rarely the whole story. The harder thing to write down is that the firm running the AI initiative was the same firm that couldn’t agree on a sprint goal.
Why it costs more this time
Agile transformations failed expensively. AI fails faster and more visibly. A sprint that doesn’t deliver costs you a fortnight. An AI programme that fails after eighteen months of procurement and change management costs you a great deal more, including, sometimes, the appetite to try again.
There’s also a clarity gap that wasn’t there with agile. Most firms doing agile knew roughly what they were trying to fix. Slow delivery, poor quality, teams working in the dark. The problem was legible even when the fix was botched.
With AI, the problem is often undefined. Firms adopt it because it seems necessary, because competitors are doing it, because a client asked. The specific thing AI is meant to fix is sometimes genuinely unclear. That isn’t a technology problem. No model resolves it.
What actually comes first
The firms I’ve seen get real value from AI had a few things sorted before anyone wrote code. They knew which decisions AI was meant to improve. They had data trustworthy enough to feed a model. They knew who owned the output when it was wrong.
None of that is glamorous. None of it makes a good case study. It’s the boring foundation that makes the interesting part possible.
The lesson from agile is simple and most firms didn’t learn it. The technology is not the constraint. The organisation underneath it is. That was true in 2015. It’s still true, and the bill is bigger now.