Two things are true in most organizations we walk into. There is appetite for AI, usually coming from the top and usually with a deadline attached. And there is data — years of it, in more systems than anyone can list from memory.
What is missing is the part in between: a clear path from the data that exists to the AI that is wanted, and the confidence that whatever gets built will survive contact with production.
That gap is where stalled pilots live. Not in the modeling. Not in the tooling. In the unglamorous distance between having data and having data that something can be built on.
What the gap actually looks like
It rarely announces itself as a data problem. It shows up as a series of small, reasonable delays that each sound like someone else’s department:
- The pilot works on an extract somebody pulled by hand, and nobody is sure how to refresh it on a schedule.
- Two systems disagree about what a customer is, and the disagreement has been survivable for years because humans quietly reconciled it.
- The model is accurate against last year and nobody can explain why it drifted, because nothing recorded what the inputs looked like at the time.
- The output is good, and no one will act on it, because the number underneath it has never reconciled to the number finance reports.
- The person who understood the source system left, and the logic that made it work was never written down.
None of these is an AI failure. Every one of them is a foundation problem that AI surfaced faster and more publicly than reporting ever did.
Why the gap is wider than it looks
Reporting is forgiving in a way that AI is not. A dashboard that is slightly wrong gets corrected by the person reading it, who knows the business well enough to spot a number that looks off. That human correction layer has been quietly holding a great many data estates together for years.
Automation removes that layer. A model does not know that the March figure is always restated, or that one region codes returns differently, or that the feed fails silently on holidays. It takes what it is given. Work that was good enough to inform a judgement is often not good enough to drive a decision without one.
This is the uncomfortable part: the gap is usually not news. Somebody in the organization has known about it for years and has not had a reason compelling enough to get it funded. AI supplies the reason. That is genuinely useful — but only if the response is to close the gap rather than to build over it.
Closing it
The sequence matters more than the technology, and it is shorter than most roadmaps suggest.
Start from the decision, not the data. Establish what decision the work is meant to improve and who has to defend it. This sounds like a formality and it is the single largest determinant of whether anything useful gets built. Without it, teams optimize for what is measurable rather than for what matters, and the result is a technically correct answer to a question nobody asked.
Assess the estate as it actually is. Not as the architecture diagram says it is. The real constraint is usually not the obvious one, and finding it takes reading the data and talking to the people who work around its quirks every day.
Fix only what the decision depends on. The instinct to clean everything first is how foundation work becomes a two-year programme that gets cancelled in month fourteen. The discipline is to repair the specific path from source to decision, prove it, and move.
Build where the work will live. In your environment, alongside the people who will own it afterwards, in checkpoints they can review. A system nobody internal understands is a liability that arrives dressed as an asset.
Then automate. AI on a foundation that holds is unremarkable to deploy and hard to argue with. AI on a foundation that does not is a demonstration that never becomes a system.
The honest version
Some of this work is not interesting. Reconciling two systems that have disagreed since 2019 is not a story anyone tells at a conference. But it is usually the difference between a pilot that becomes infrastructure and one that becomes a slide.
And sometimes the answer is to wait. Not every organization should be spending on this yet, and the ones that should are not always the ones with the most enthusiasm. A firm that will not tell you that is not being helpful.
The gap between data and AI is not a technology problem. It is a sequencing problem, and sequencing is a choice.