Walk into almost any manufacturer right now and you will find the same conversation happening at two different volumes. In the boardroom, it is loud: the directors want an AI strategy, the competitors are said to be "doing something," and nobody wants to be the business that moved too late. On the factory floor and in the data team, it is quieter, and more anxious: another pilot has stalled, the proof-of-concept that dazzled in the demo has gone strangely silent, and no one is quite sure why.

That gap - between AI enthusiasm at the top and AI outcomes everywhere else - is the defining manufacturing technology story of the moment. And the reason it exists is almost never the thing everyone assumes.

The model is rarely the problem

The numbers are sobering, and they are remarkably consistent across the analysts who track them. Depending on whose research you read, somewhere between 88% and 95% of AI pilots never reach meaningful production. IDC found that for every thirty-three proof-of-concepts an enterprise starts, only about four make it into real use. MIT's Project NANDA put it more bluntly still: roughly nineteen in twenty generative-AI pilots deliver no measurable return at all. Gartner expects organisations to abandon a majority of AI projects that lack AI-ready data and attributes the bulk of outright failures to poor data quality rather than poor algorithms.

Here is the part that should change how leadership teams think about it. When organisations are asked why a pilot failed, they tend to blame the model, the immaturity of the technology, or the difficulty of the use case. The research says they are looking in the wrong place. RAND's analysis of failed projects found the dominant causes were organisational and data-related - not technical limitations of the AI itself. The failure, in other words, is upstream of the algorithm. The cleverest model in the world cannot rescue a foundation that was never built.

The pilot trap

There is a specific, almost universal pattern behind these collapses, and it is worth naming because once you see it you cannot unsee it.

A pilot succeeds because a data science team, by hand, assembles exactly the data it needs - cleaned, reconciled, complete - into a tidy set that proves the concept beautifully. Everyone is delighted. Then the project moves towards production, where that curated dataset does not exist and never did. The same model now has to run on the full, messy reality of the business: data scattered across systems, inconsistent between them, duplicated, incomplete, and arriving far faster than anyone can hand-clean it. The manual process that made the pilot shine cannot run at production volume. The model that looked brilliant on prepared data produces confident nonsense on real data. And a project that took six to twelve weeks to demonstrate quietly enters the six-to-twelve-month limbo that most never escape.

The pilot did not lie. It simply tested the wrong thing. It proved the model could work; it never tested whether the data foundation could feed it.

Why manufacturing feels this more acutely

Every industry struggle with this, but manufacturers carry a heavier version of the problem, and it is not their fault. Manufacturing data is spread across more silos than almost any other sector: ERP systems holding the commercial picture, MES and historians on the plant floor, separate quality and maintenance systems, supplier records, and - inevitably - a constellation of department-level spreadsheets that have become load-bearing without anyone deciding they should be.

On top of that sits the operational-technology divide we have written about before: the plant-floor systems that were never designed to share data cleanly with enterprise IT, and the multi-site reality where the "same" data means subtly different things in different factories. The result is that most manufacturers do not have a single, trustworthy version of the truth for an AI system to learn from. They have a dozen partial versions that disagree. You cannot automate, predict or optimise on a foundation you cannot trust - just as, on the security side, you cannot protect what you cannot see. It is the same discipline, viewed from a different angle.

AI demand is really data demand

The most useful thing a leadership team can do with the current wave of AI enthusiasm is to translate it. When the business says "we need to do something with AI," what it actually needs, almost always, is a connected, governed data layer that makes AI possible - and a great deal of the value it is chasing turns out to be achievable from that layer long before a single model is trained.

This is the reframe that separates the manufacturers who succeed from those who cycle through pilot after pilot. The successful ones treat data readiness as a precondition for the pilot, not as something to be improvised during it. They build a layer where data is integrated across systems, matched and de-duplicated so the same customer, part or process is recognised consistently, quality-checked, and governed so that everyone knows what a figure means and can rely on it. One version of the truth, continuously maintained, that an AI system - or a dashboard, or a regulator, or a human being making a decision - can actually depend on.

It is less glamorous than the AI conversation the board wants to have. It is also the only version of that conversation that ends in production.

What good looks like

The manufacturers getting this right tend to do a handful of unglamorous things well.

They define the outcome before they pick the technology. Not "we need AI," but "we want to cut unplanned downtime on this line by a measurable amount" or "we want to predict quality failures before they reach the customer." A quantified business objective keeps everyone honest and gives the eventual pilot something real to be judged against.

They run an honest data-readiness assessment before they build - mapping where the relevant data lives, how good it is, and what it would take to make it trustworthy. This is the step almost everyone skips, and skipping it is what the failure statistics are made of.

They build the foundation incrementally, starting where the data is good enough and the value is real, rather than attempting to perfect everything at once. A single well-chosen use case on solid data beats five ambitious ones on sand.

And they treat governance as an enabler rather than a brake. As frameworks such as ISO/IEC 42001 and the NIST AI Risk Management Framework mature - and, for those exporting into the EU, as the EU AI Act bites - structured governance is increasingly a source of advantage, not a compliance tax, because it catches the problems that kill projects before they reach production. Knowing where your data came from, who can use it, and whether it can be trusted is not bureaucracy. It is the thing that lets you scale with confidence.

The foundation that pays for itself

There is a final reason to start with the data layer rather than the model, and it is the most commercially persuasive one. A model is single-purpose; a good data foundation is not. The connected, governed layer you build to make one AI use case work becomes the layer that powers the next one, and the one after that - as well as your analytics, your reporting, your compliance evidence, and the everyday decisions people make without any AI involved at all. The investment compounds. The pilot-by-pilot approach, by contrast, spends the same effort again and again and keeps arriving at the same disappointment.

The manufacturers who win with AI over the next few years will not, for the most part, be the ones who piloted first. They will be the ones who built the foundation first - quietly, deliberately, while their competitors were still wondering why the last demo never made it to the floor.

So, when the board asks for an AI strategy, the most valuable answer a technology leader can give is a slightly different one: a data strategy that makes AI inevitable. Get that right, and the pilots stop stalling. They start scaling.

At VE3, we help manufacturers build exactly that foundation - integrating data across fragmented systems, raising and maintaining its quality, and governing it so that AI, analytics and decision-making can all draw on a single, trustworthy source. If the AI conversation in your business keeps ending in a stalled pilot, the place to look is underneath it.

We help manufacturers and industrial operators turn fragmented data into a connected, trusted foundation for analytics and AI. Talk to us about a data-readiness assessment.