The 78 % trap: why GenAI fails in industry, and what the successful 22 % do differently

Gartner confirmed in 2026 what has been visible on the shop floor for years: only 22 % of companies draw significant value from generative AI. The remaining 78 % are burning budget, not because the technology fails, but because the groundwork is missing.

The uncomfortable number

The Gartner report Predicts 2026: Intelligent Applications contains a figure that ought to unsettle any board:

Only 22 % of organisations report that GenAI tools deliver significant value.

Which means more than three quarters of all organisations investing in generative AI see no meaningful return. After two years of intense hype, billions in investment, hundreds of pilots.

The question is not whether the number is right. The question is why.

The reflex, and why it is wrong

The typical response to disappointing AI results follows a predictable pattern.

• We need a better model.
• We need more training data.
• We need a different vendor.

Thirty years of industrial work, from the production line through ERP implementations to AI strategy, point the other way. When technology projects fail, the problem is almost never the technology.

It sits in front of it. In the processes. In the data. In the organisation.

Put as an image: we argue about engine output while the chassis does not exist.

Three causes of the 78 % trap

One: AI is set up as an IT project instead of a leadership task

Gartner supplies the matching finding: 56 % of IT leaders say they cannot drive GenAI adoption on their own. They need the business. They need leadership.

In practice it looks like this. IT is asked to do something with AI. A proof of concept appears. A chatbot. A demo. The board nods. And then nothing happens, because nobody embedded that chatbot in an operational process.

AI is not a tool you hand to IT. It is a strategic decision that belongs to the executive team. Who automates what, to what end? Which processes change? Which roles shift? None of those are IT questions.

Adult supervision is the phrase that fits: the experienced hand that steers where AI is used, and where it deliberately is not.

Two: the data is not ready, and nobody wants to hear it

Gartner forecasts that by 2030, 35 % of large enterprises will have improved the quality of their AI-ready data. The figure for 2025 was 14 %.

That is not an abstract statistic. It is everyday life.

An example. A large automotive supplier near Stuttgart. A transport management system, technically sound, professionally designed. Then the reality check: the system could not calculate optimised routes. Not because the algorithm was poor. Because nobody knew what the parts weighed.

Weights and dimensions, the most basic master data in logistics, were missing. Or wrong. Or in a spreadsheet nobody had maintained for years.

The entire project scope had to be redefined.

AI is only as good as the data it receives. In most industrial companies that data is fragmented, outdated or simply absent. No language model compensates for master data that does not exist.

Three: embedded AI produces noise instead of effect

Another finding from the same report: much of the embedded intelligence will generate more noise for users than actual support in getting their work done.

Every software vendor is currently gluing an AI feature onto its product. A copilot here, an assistant there. It sells well. In operational reality it means ten different AI features in ten different applications that do not talk to each other.

It is the equivalent of fitting a different engine into every vehicle in a fleet, with no shared fuelling network, no common spare parts, no consistent maintenance logic.

The value of AI does not appear in the feature. It appears in the process. Not inside a single application, but across the whole workflow, from source to contract, from requisition to invoice check.

What the successful 22 % do differently

Read the findings together and a clear picture emerges. The companies that genuinely extract value differ in three ways.

AI belongs to the executive team, not to an IT sandbox. Strategy, governance and delivery are centralised and tied to business goals rather than to technical curiosity.

They invest in data quality before they invest in models. The best engine is useless with an empty tank. So they clean up master data, define standards, and build the ground AI can work on.

They integrate AI into existing processes instead of parking it alongside. No isolated chatbot. No demo that impresses in a meeting room and gathers dust in daily operations. AI that reaches into operational workflows, with clear rules, clear governance and clear ownership.

From engine to chassis

The image is worth keeping. The language model is the engine. The chatbot is the bodywork. Together they look impressive, and without a chassis they go nowhere.

The chassis is the existing processes: ERP, MES, bills of material, supplier ratings, contracts, quality data. Anyone who wants industrial value from AI has to fit the engine into that chassis. Not beside it. Not on top of it. Into it.

That is the work most people avoid. It is not glamorous. It demands process knowledge. It means getting your hands dirty, in master data, in interfaces, in organisational resistance.

It is also where the value is.

The real question

The 78 % trap is not a technology problem. It is a leadership problem, a data problem and an integration problem, and no better model solves it. People who understand both worlds do: the technology and the operational day.

The question for a decision-maker today is not whether the company has enough AI projects.

The question is: which group are we in?

---

#AI #GenAI #ITStrategy