Why 86 % of all AI projects fail on data, before they even start

The number nobody wants to hear

Gartner confirmed it in 2025: only 14 % of companies hold data that is ready for productive use of generative AI.

Fourteen per cent.

Which means the other 86 %, all of them investing in AI right now, are building on a foundation that will not carry the weight. They are buying the most powerful engine available and bolting it into a chassis with no axles.

The number is no surprise. It is statistical confirmation of something visible in three decades of industrial work, from ERP rollouts through digitalisation to the current wave.

The algorithm is never the problem. The problem sits in front of it.

A case from the field: the missing weights

One example that stands in for dozens.

A large automotive supplier near Stuttgart invested in a transport management system. Technically sound, cleanly designed, professionally implemented. The goal: optimised routing, lower freight cost, better utilisation.

Then came the reality check.

The system could not calculate a single optimised route. Not because the algorithm was faulty. Not because the software had defects. Because the most basic master data was missing: the weights and dimensions of the parts.

That information was either absent, obviously wrong, or asleep in spreadsheets nobody had maintained for years. No system on earth, classical optimisation or state-of-the-art AI, can work with data that does not exist.

The project scope had to be redefined. What began as a software implementation became a master data cleanup first. The schedule slipped by months. The budget did not survive.

Three levels of data failure

That case is not an outlier. Across three decades, from furniture manufacturing through ERP consulting to AI strategy, the same pattern keeps appearing in new clothes. Data failure comes in three levels.

Missing data. The information simply is not in the system. Weights, dimensions, lead times, quality metrics. Fields that exist in the ERP and were never filled. Everyone knows. Nobody owns it.

Wrong data. The information exists but is stale or incorrect. A supplier moved and the old address is still on file. A material weight was estimated at creation and never validated. Prices from a previous decade sit next to current terms.

Scattered data. The information exists and is correct, but it lives in the wrong system. Or in a spreadsheet on a planner's desktop. Or in an email from 2019. The truth is fragmented, and no algorithm can assemble fragments it does not know about.

Why AI makes this worse, not better

Here is the decisive error in reasoning. Many decision-makers believe AI is the answer to their data problem. The model will find the right patterns. Machine learning can learn from bad data.

It cannot. Not in a way that can carry an operational decision.

Classical software fails on bad data visibly. A wrong address produces a wrong delivery. Somebody notices. Somebody corrects it.

AI fails invisibly. A language model trained or grounded on faulty master data produces answers that sound plausible and are wrong. It does not only hallucinate where knowledge is missing. It hallucinates where knowledge is wrong. In an industrial context, a plausible and incorrect recommendation can cost millions before anyone thinks to check.

Bad data in a classical system is a flat tyre. You feel it immediately and pull over. Bad data in an AI system is a bearing failing slowly. Everything runs smoothly until the wheel comes off, at speed.

What the 14 % do differently

Three traits stand out among the companies that are genuinely AI-ready.

Data quality has an owner in the executive team. Master data quality is not an IT topic managed in the second row. There are clear responsibilities, defined maintenance processes and regular audits. The CFO or COO asks about the state of the data before a project fails, not after.

Cleanup before innovation. These companies have the nerve to prioritise boring work. Master data migration, data governance, process harmonisation. None of it makes a good conference talk. All of it is the foundation everything else stands on.

Process thinking instead of tool thinking. The 14 % do not ask which AI tool to buy. They ask which process carries the greatest leverage, and what data is needed for AI to create value there. The process defines the data requirement. The data requirement defines the cleanup. Technology comes last.

Foundation first

This is where the work actually starts. Not as an evangelist promising the next revolution, but as someone who knows the gap between what the technology can do and what the organisation can deliver.

From time-study engineering on the shop floor through ERP consulting to AI strategy, one lesson holds: every technology is only as good as the foundation under it.

Put differently: you can build the most powerful engine in the world. If the chassis will not carry it, nothing moves.

The 86 % do not fail at AI. They fail at what comes before AI. That preparatory work, the translation between what the technology needs and what the organisation can supply, is the whole job.

The honest question

Before releasing the next AI budget, one question is worth sitting with.

If an algorithm analysed your master data today, what would it find?

Clean, current, complete records? Or a mosaic of gaps, legacy entries and spreadsheet workarounds?

The answer decides which group the next project joins.

---

#AI #MasterData #ITStrategy