Everyone wants AI. Far fewer have the data foundation it quietly depends on. Across industry, predictive maintenance, quality prediction and energy optimisation have moved from workshop slides into real roadmaps - yet most initiatives stall not on the algorithm, but on the data beneath it. AI readiness is, in the end, a question about that foundation. The gap rarely shows up in a model’s accuracy; it shows up when a pilot that shone on one line cannot be repeated on the next, because the second line’s data lives in different systems under different names. What closes it is not a cleverer algorithm, but data that is connected, consistent and trusted before the model ever runs.
IN BRIEF
- Most industrial AI initiatives stall on the data beneath them, not on the algorithm.
- Predictive maintenance, quality analytics and digital twins all need data that is available, integrated, contextualised and trustworthy.
- Gartner estimates around 60% of AI projects are abandoned through 2026 for lack of AI-ready data.
- The precondition is a harmonised, governed data foundation - built with human judgement first, then extended with AI.
Part of the series , The Industrial Data Foundation
The shift
AI has moved from experiment to operational plan
Artificial intelligence is now on nearly every industrial agenda. Manufacturers are exploring predictive maintenance, production planning, quality management, energy optimisation and operational decision support, and in many organisations these have graduated from innovation workshops into strategic roadmaps.
It makes sense. A modern plant generates enormous volumes of data across machines, sensors, lines, ERP, MES, historians and quality systems. In principle, all of it can feed better forecasts, faster decisions and more resilient operations. The catch is that value does not flow from data automatically - it has to be built up, layer by layer.
Why pilots stall
The pilot works; the rollout does not
A predictive-maintenance model that shines on one line’s clean historian data often falls over at the second plant - where the same sensor is tagged differently and “downtime” is defined another way. The reasons are rarely exotic. Operational data sits in disconnected systems, so a consistent view of production or assets is hard to assemble. Data points arrive without context, so a model can spot a pattern without knowing which asset, batch or order it belongs to.
Underneath those sit two quieter problems. Definitions and units differ between departments and sites, so the same figure rarely means the same thing twice. And ownership is unclear: If no one can say who validates and maintains a number, the resulting outputs are hard to trust and even harder to audit. Add an initiative that started from the technology rather than a value case, and you have a pilot that impresses in the room and never reaches the floor.
AI does not clean your data. It amplifies whatever it is fed.
The real question
Is the foundation ready - or just the model?
AI does not remove the need to bring data together, keep it clean, add context and govern it - it raises the stakes on all of them. For a business-critical decision it is not enough that a system produces an answer; you have to know where the data came from, how it relates to the process, and whether the result moves a measurable objective.
The industry numbers are blunt about it. Gartner predicts that through 2026, organisations will abandon 60% of AI projects that are unsupported by AI-ready data, and at least 30% of generative-AI pilots to be dropped after proof of concept, with poor data quality among the leading causes. Put simply: on a fragmented foundation, the share of AI initiatives that reach reliable production is close to none; on harmonised, governed data, it is a different story.
So readiness starts not with a tool but with an outcome - less downtime, more throughput, lower energy use, faster reporting - and then asks whether the operational data those outcomes depend on is available, connected, contextualised and reliable. Where the answer is no, the work is not a better model. It is a better foundation.
| Response to a stalled pilot | What it actually changes |
|---|---|
| A bigger model or a new tool | treats the symptom - the data stays fragmented |
| Another data lake | more copies of the same conflicting numbers |
| A layer above the systems | connect once, bring the data together, keep it traceable |
The way out
Build the foundation first, the intelligence on top
The foundation that works is not another model or a bigger data lake - it is one layer above the systems that connects each source once, brings the data together into a shared business view, and lets analytics and AI run on top. In practice it does a handful of concrete jobs:
- Connect once, not point-to-point. Each system joins a shared layer instead of wiring to every other, so a new tool does not multiply the integration bill.
- Model the meaning. Data is linked to assets, orders, batches, materials and time, and harmonised so the same figure means the same thing across sites - without erasing each system’s own valid view.
- Link, don’t copy. Values stay live at their source, so nothing silently goes stale between exports.
- Keep it traceable and governed. Reported figures stay tied to an authoritative source - which is exactly what makes an AI output explainable.
- Start where the value is. One outcome, one process first; grow by need.
Build the data foundation with human intelligence first - then add artificial intelligence on top.
The lesson travels
Four signs your AI plan is really a data plan
The signs are easy to spot once you look for them:
- Pilots that look great in the lab but never scale to a second line or plant.
- Two systems that report different values for the same KPI.
- A number no one can trace back to an owning source.
- A use case that stalls the moment definitions or units change between sites.
Any one of them says the blocker is not the model. Building the foundation beneath it - reliable, harmonised, traceable data the whole business can stand on, the sequence explored in why you need a data foundation before advanced analytics - is exactly what the boTec DataBridge is built to be.