Machine logs can be analysed before the integration is built

The pilot that stalled rarely stalled because the model was too weak. It stalled at the integration: live signals from the machines, order data from the enterprise system, quality data from systems that were never built to talk to each other. All the while the machines have gone on writing. Logs, quality records and maintenance history sit in systems that are already paid for. The question is not whether the material exists. It is whether anyone has put a question to it.

The unused data is a known fact, not a known quantity in your plant

Rockwell Automation published the eleventh annual State of Smart Manufacturing in May 2026. It reports that organisations keep collecting growing volumes of data while only 43 per cent of it is used effectively, and concludes that the constraint sits in execution rather than in the supply of data.

That figure is a self-assessment. What it measures is what executives say about their own organisation, not a reading taken from any system, and Rockwell sells industrial automation. The distance between what someone reports and what a register shows is the same distance licence data made visible for software.

Which makes the number useful as an order of magnitude and useless as a measurement of any single plant. An organisation that wants to know how much of its own machine data underpins a decision gets that answer another way, by going to the records.

The maintenance type was wrong on one log in four

A team at the Institute for Energy Systems at the University of Edinburgh, together with a co-author from the operator Nadara, published a preprint in May 2026 in which a language model worked through 16,316 maintenance logs from 280 wind turbines across 32 onshore wind farms. The material runs from 2017 to 2026 and represents more than 22 million operating hours.

The model corrected the maintenance type on 3,997 logs, 24.5 per cent of the whole dataset. It assigned a failure mode profile, with failure mode, dominant mechanism, observed symptoms and possible causes, to 11,662 logs, 71.5 per cent. A review of 300 randomly drawn records gave 80 to 90 per cent agreement between the model's output and the expert's judgement.

What is remarkable is not that a language model can do the job. It is what the correction says about the starting position. The one category that all maintenance statistics rest on, whether the work was corrective or preventive, was wrong on one record in four. Anyone who has calculated the share of corrective maintenance out of that system calculated on a label and not on what happened at the machine.

The Data Act changed what you can get out of the machine vendor

The objection to analysing machine data is rarely that it does not exist. It is that it sits in the vendor's system. Since 12 September 2025 the Data Act, Regulation (EU) 2023/2854, has applied. The European Commission describes the right as giving the user of a connected product access to the data its use generates, and Article 5 gives the user the right to have the data holder make that data available to a third party the user chooses.

Two limits belong to the same text. Information inferred or derived from the data sits outside it, being the outcome of additional investment, which means the vendor's own analyses and indices rather than the readings themselves. And a data holder may refuse a specific request where it can demonstrate that disclosure of trade secrets is highly likely to cause serious economic damage. The requirement that the product itself be designed so that data is accessible reaches only products placed on the market after 12 September 2026, under Article 50. The design requirement is limited to new products in a way the access right is not.

What argues against the logs already holding the answer

The Edinburgh material is a preprint and has not been through peer review. It comes from a single fleet, and it is wind turbines rather than a workshop; nobody has measured whether the share of wrong labels looks the same in a maintenance system at a manufacturing company. One of the four authors is listed on the author line at the energy company Nadara rather than at the university.

That the share is unmeasured in a workshop is a reason to measure it in yours, not a reason to leave it.

The figures carry their own reservation too. Agreement with the expert assessment ran at 80 to 90 per cent, which means that between one in ten and one in five of the model's outputs did not match an expert. And 4,654 logs, 28.5 per cent of the material, got no failure mode profile at all.

Neither measurement has a Nordic population. Rockwell's respondents sit in seventeen countries without a published breakdown, and the Edinburgh material is a single fleet from one operator. A decision-maker in Sweden is therefore weighing evidence gathered elsewhere, and the nearest Swedish measurement covers something else entirely: Statistics Sweden puts AI use at 35.0 per cent of companies with at least ten employees in 2025. That figure says nothing about machine data. What it does say is that a majority of Swedish companies had not started using AI at all in 2025.

The conclusion from the objections is not to leave the logs unasked. It is that the share of uninterpretable records belongs in the delivery alongside the result, and that the share of corrected labels should be worked out on your own material before it underpins an investment decision.

The criterion: your three most recent stoppages

Take the three most recent unplanned stoppages on the line that matters most. Go to the systems, not to the people who were there. Can you read off when each stoppage began, when it ended and which cause was registered, without anyone filling the gaps from memory?

A stoppage with no cause code counts in time but never in cause, and it is the cause that decides what gets fixed.

If the answer is yes on all three, the material carries an analysis today, and the first thing that analysis should do is check the cause codes against the free text instead of trusting them.

If the answer is yes on the times but no on the causes, the work sits in the registration and not in the model. That is a bounded change to how a stoppage is signed off, and it yields usable material only after some months of registration.

If the answer is no on both, the first question is where the stoppage ends up at all. That question nearly always has an answer, and the answer usually sits in a system the organisation already pays for.