Almost every AI conversation I have starts with a use case and ends with the same problem: the data underneath it was never built. The model is rarely where these projects fail. They fail much earlier, in the meeting where someone decides the data can be sorted out later.

When an AI initiative stalls, the post-mortem usually points at the model, the vendor, or the use case. People retune prompts, swap platforms, and re-scope the pilot. They are debugging the wrong layer.

The real failure point sits underneath all of that. The project was scoped to run on data that is scattered across systems that do not talk to each other, that nobody owns, and that the business quietly does not trust. None of that shows up in a demo. All of it is fatal in production.

That is what "fail before they start" really means. By the time the first sprint kicks off, the outcome is already decided, because you cannot retrofit a foundation under a build that is half finished. You cannot get to agentic AI, an agent that takes action on your behalf, without getting the data right first. And you cannot get the data right after the agent is already acting on it.

The pattern is remarkably consistent. My team gets called in once an organisation has already decided it wants to do something with AI: an agent in front of the service desk, automated forecasting, a copilot for the sales team. The ambition is rarely the issue. What sits underneath almost always is.

And it is not a niche failure mode. Gartner surveyed 248 data management leaders and found that 63 per cent of organisations either do not have, or are unsure whether they have, the data management practices AI requires. On the back of that, Gartner predicts organisations will abandon 60 per cent of AI projects that are not supported by AI-ready data through 2026.

Not long ago we were brought into an enterprise that wanted an AI agent answering customer questions. A few weeks in, we found three systems that each held a different version of who the customer even was. Marketing, sales, and finance had quietly disagreed for years, and nobody had been forced to resolve it, because no human decision had ever depended on all three being right at the same moment. I set this pattern out at length in The Data Conversation. The definitions are not wrong individually, which is exactly why nobody ever resolves them:

An agent does not have the option of picking one. It reads all three and treats each as fact.

An AI agent depends on exactly that. The technology did not create the mess. It removed the human in the loop who used to quietly absorb it. That is the recurring autopsy: not a model that underperformed, but a foundation that was never agreed on, exposed the moment something automated started relying on it.

The data foundation gets skipped precisely because it is the least exciting thing in the room. The agent demos beautifully. The governance model does not. There is no board-meeting moment in defining who owns customer data, and no quick win in reconciling two systems that have disagreed for a decade.

When the board asks "are we doing AI", the honest answer, that you are about six months of unglamorous data work away, rarely survives contact with the quarter. I have watched this play out in organisations of every size and sector, and the temptation is identical in all of them. It is not a budget problem or a maturity problem. It is a human one. Skipping the foundation is rational in the moment and fatal over the year, and almost everyone makes the rational choice.

When I say data foundation, I do not mean a tooling purchase. I mean four specific things, and every one of them is a commercial outcome wearing a technical costume.

None of these are technical luxuries. They are the difference between an AI project that compounds and one that quietly gets shelved while everyone agrees not to mention it.

You do not need a data maturity assessment to know where you stand. You need one honest conversation in the approval meeting, before anyone writes a sprint plan.

Ask this: what is the data foundation this will run on, who owns it, and would our own people trust the numbers without checking a spreadsheet on the side? If the room cannot answer that, you have not approved an AI project. You have approved an expensive way to discover, six months from now, that the data was never ready.

Get that answer first. It is the cheapest insurance you will ever buy, and it is the one step almost everyone skips.