Data is a production input

There is a moment on the floor of every regulated plant that does not exist in most AI pipelines. A truck backs up to the loading bay, a pallet comes off, and before the material can be used for anything it stops. Someone draws a sample, runs it against a written specification, and decides whether the lot is released to production or sent back. The material is guilty until proven conformant. Nothing about its origin, its supplier or its paperwork earns it a pass. Only the test does.

Now look at how most AI systems are fed. The data arrives, and it goes straight in.

That gap, not model choice, is what will separate the companies that make AI work from the ones that keep being surprised by it. It is a question of input control, and manufacturing settled it a century ago. This essay works through what the settlement looked like, why the market has just put a hard price on it, and the three tests any organisation can run at its own receiving dock this quarter, none of which requires a platform purchase.

Volume is a balance sheet idea

The language of AI strategy still treats data as an asset, something to gather, hoard and defend, as if volume were the point. Volume is a balance sheet idea. Qualification is a shop floor idea, and it is the one that matters in production. The companies that will make AI dependable are not the ones with the most data. They are the ones who treat each dataset the way a plant treats a raw material, with a specification it must meet, a record that proves it met it, and a gate it cannot skip.

The intake gate is not caution for its own sake. It is the cheapest place in the entire system to catch a problem. A defect found at goods inwards costs the price of a returned delivery. The same defect found after the product has shipped costs a recall, a regulator and a reputation. Manufacturing learned to move the inspection as far upstream as it would go, and then to build the quality in so the inspection rarely fails.

The encouraging part is that none of the machinery has to be invented. It already exists, under names that have not yet been recognised as the same thing. Every input in a good factory has a specification, the quality criteria it must meet before use. Every batch carries a certificate of analysis, the document that proves it conforms. Every supplier is qualified before a single delivery is accepted. Software is arriving at the same answers, late, under different names. The data contract is a specification. The dataset card, formalised in Gebru and colleagues' Datasheets for Datasets in 2018, is a certificate of analysis. The intake gate is incoming quality control. None of it is new. It is good manufacturing practice, rediscovered by teams who never worked a factory floor.

The market has put a price on the layer

For a long time this argument had to be made on logic alone. In July 2026 the market made it in cash. Schneider Electric agreed to pay $3.1bn, all of it cash, for Cognite, a company whose entire product is taking raw industrial data and making it clean, contextualised and fit for use. Cognite's annual revenue passed $170m last year, so Schneider paid roughly eighteen times sales, not for models and not for compute, but for the layer beneath the machines. The plan is to fold Cognite into Connect, the cloud platform of its industrial software arm Aveva, and Schneider's chief executive Olivier Blum described the prize precisely: a platform that "turns the complexity of operational data into a competitive advantage."

Read that valuation as a statement about where the constraint now sits. The models are increasingly interchangeable. The compute is rentable by the hour. What cannot be rented is the discipline that makes an organisation's own data usable, and a buyer with Schneider's operational pedigree just priced that discipline at $3.1bn. The data layer is not a warehouse to be filled. It is a supply chain to be governed, and supply chains are governed at the point of receipt.

What never got recorded

The same week the market priced the layer, Ford showed what it costs to run one without the discipline. After the AI and automated quality systems it had backed fell short on the factory floor, the company rehired 350 veteran engineers, the grey beards, some out of retirement and some back from suppliers. Ford's chief operating officer Kumar Galhotra told journalists the company had been "relying more and more on automated quality systems" with disappointing results. Charles Poon, its vice-president of vehicle hardware engineering, was blunter: "Mistakenly we thought that by just introducing artificial intelligence and ingesting the design requirements that we had, that that would produce a high-quality product."

The models had been trained on everything Ford had recorded. The problem was everything Ford had never recorded. A dataset is an inventory of what an organisation bothered to write down, and the most expensive knowledge is usually the part that stayed in someone's head. Ford paid for that gap twice, once in the AI that fell short and once in the engineers called back to fill it. The recovery is worth noting too. The rehired specialists now hunt for failure points before a part reaches the plant floor, they are training younger staff and reprogramming the AI tools, and Ford expects the programme to take $1bn out of its costs this year while topping the JD Power initial quality survey among mainstream brands. The lesson is not that the AI failed. It is that the inputs were never qualified, and the qualification turned out to live in people.

Three tests at the receiving dock

Toyota solved this problem for physical inputs forty years ago. Nothing reached the line without a certified supplier, an inspection at goods inwards and a paper trail back to origin. Most AI programmes run the opposite regime, ingest first, ask questions at the output. Three tests belong at the dock, and each can be run this quarter with nothing more than a meeting and a spreadsheet.

The first is provenance. Who supplied this data, and would you certify them the way you certify a component supplier? A plant does not accept a pallet from an unknown foundry because the price was right. Most data pipelines accept feeds on far weaker assurances, a scraped source here, an unexamined vendor extract there, with no equivalent of supplier qualification anywhere in the chain.

The second is freshness. When was this last true? Data has a shelf life, and stale inputs do not announce themselves. A material past its retest date is physically quarantined in a plant. A table that stopped updating eight months ago sits in the warehouse looking identical to one refreshed this morning, and the model consuming it cannot tell the difference either.

The third is coverage. What never got recorded? This is the Ford question, and it is the hardest of the three because the gap is invisible by definition. The test is not run against the data. It is run against the people who did the work before the data existed, which is exactly why Ford's remedy was to rehire the people rather than buy another platform.

The specification written into law

Regulators arrived at the intake gate independently, and they are now writing the specification into statute. The EU AI Act requires that training data for high-risk systems be relevant, representative and, to the best extent possible, free of errors, under documented governance. That is a specification, enforceable from August, and the fine for ignoring it reaches fifteen million euros or three per cent of global turnover, whichever is higher.

The sharper finding is which law is actually biting. New analysis from GovAI shows that data protection rules, not the AI Act, are now the brake on 11 per cent of large language model releases in the EU. The questions the intake gate should have asked, where did this come from, who consented, is it still accurate, arrive anyway, only later and with legal force. In UK health, 82 per cent of patients say governance of their health data is not good enough, a trust deficit that lands on every data-reliant deployment that follows. And the running cost of skipping the gate has a number attached. Organisations estimate that poor data quality costs them an average of $12.9m a year. That is the price of piping the truck straight in.

The cord to pull

Toyota had a word for the alternative. Jidoka. Build quality into the line, and give any worker the authority to stop it the moment a defect appears, because a defect caught at the source is cheap and a defect caught after shipping is a recall. The andon cord was not a technology. It was a delegation of authority, downwards, to the person closest to the defect.

Most AI pipelines have no cord to pull. Data flows in continuously, nobody owns the moment of receipt, and the first person with the standing to stop the line usually sits so far downstream that by the time they act the defect has been trained into the model. The winners will run it the other way. Qualify the source, test against the spec, quarantine until released, keep the record, and give someone at the gate the authority to say no.

Case study: the molecule still has to survive the wet lab

The pharmaceutical industry is about to run the largest live experiment in input qualification anywhere in the economy. The Financial Times reports forecasts that 60 per cent of new drugs will be machine-designed within a decade. The molecules still have to survive the wet lab. A generative model can propose candidates at negligible marginal cost, which means the scarce, expensive, rate-limiting input is no longer the idea. It is the verified experimental result that confirms or kills it.

That inverts the economics of the data layer. Generated candidates are cheap, verified ground truth is not, and the firms that win will be the ones that treat lab results as certified stock rather than another feed, with provenance to the bench that produced them, freshness against the assay that dated them, and honest coverage accounting for the negative results that never made it into a paper. Pharma has run goods inwards on physical materials for half a century under GMP. Its advantage in the machine-design era will be deciding, earlier than other industries, that data from the wet lab deserves the same regime.

Toyota in 1985

None of this is a counsel of caution. It is a description of how production systems mature. Craft becomes industry when inputs stop being trusted and start being qualified, and every industry that made the transition looks back on the trusting era with disbelief. The AI winners of the next decade will look more like Toyota in 1985 than a model lab in 2023, not because they moved slowly, but because nothing reached their line that had not earned its way on.

Data is not a strategic asset you accumulate. It is a production input you qualify.