Manufacturing AI, lagging vs leading: the anatomy of a widening gap

Stage 1 is not a technology failure — most disconnected plants are competently run and some are very profitable. What defines the stage is that the plant produces no machine-readable record of its own behaviour. The PLCs cycle, the line runs, the product ships, and none of it leaves a trace an application could learn from. Every improvement conversation therefore starts with anecdote: the filler 'always' struggles on Mondays, the oven 'tends to' drift in summer, line 3 'never' hits rate on the thin-gauge material.

The economics of the stage are deceptive because the cost is invisible. Nothing fails loudly; the plant simply pays a permanent tax in the form of questions it cannot answer. Which machine loses the most availability? Nobody can say without a two-week clipboard study. Did last quarter's changeover programme actually reduce downtime? The data to answer was never captured. In a disconnected plant, every argument is settled by seniority rather than by measurement, and the improvement agenda belongs to whoever argues best.

The exit is cheaper than most stage-1 leaders assume, because it is deliberately narrow. The move is not a plant-wide digitisation programme — it is instrumenting one bottleneck line well enough to answer one question. Retrofit sensing on the three worst machines, an OPC UA gateway, a historian, one agreed set of downtime codes. Plants that frame the exit this way cross in a quarter; plants that frame it as 'Industry 4.0 transformation' write a strategy deck instead and are still at stage 1 two years later.

Stage 2 is the largest single population in manufacturing and the easiest to misread from the outside, because the plant looks digital. SCADA screens glow in the control room, the historian ingests thousands of tags, quality records are electronic. The tell is what the data is for: it is written far more than it is read, and read almost exclusively backwards — for an audit, a customer complaint, a post-incident investigation. The plant has built the sensory apparatus of a leading operation and connected it to nothing that decides anything.

The structural problem is context, not collection. Historian tags carry values without meaning: pressure_04 spiked at 14:02, but which product was running, which recipe, which crew, whether the spike was during changeover — reconstructing that context takes an engineer half a day per question, so questions go unasked. Meanwhile every department that touches the data has evolved its own definitions. Production's OEE excludes planned maintenance, engineering's does not, finance computes a third number for the board pack. None of them is wrong; there is no shared definition to be right about, and every cross-department improvement discussion begins with an argument about whose number is real.

This is also the stage where AI pilots begin, and where the fork ahead is set up. A pilot built on hand-cleaned historian exports can absolutely produce a good model — that is precisely what makes stage 2 sticky. The model works, the deck lands, and the organisation concludes it is 'doing AI' while the loop that would make the second model cheaper than the first — shared definitions, contextualised data, a route into a workflow — remains unbuilt. Stage 2 plants do not lack proof that AI works. They lack anywhere for it to live.

Stage 3 is a genuine achievement and a genuinely dangerous place to rest. The plant that reaches it has done the unglamorous work: definitions agreed, context modelled, dashboards trusted enough that the morning meeting argues about causes rather than numbers. Waste is visible, and visibility alone usually buys a real one-off improvement as the obvious losses get actioned. That improvement is also the trap, because it feels like the payoff when it is actually the entry fee.

This is the fork of the whole curve, and what forks is a wiring decision that looks minor at the time. The pilot's output can terminate on a dashboard, or it can be written into the system that runs the decision — a predicted-failure work order in the CMMS queue, a recommended setpoint in the MES, a revised sequence in the APS. The dashboard path is faster, needs no integration approval and no argument with the MES vendor, so it is the path most plants take. It is also, on the evidence of the whole industry, the path that leads back to the start: dashboard adoption depends on a person voluntarily adding a step to their day, that discipline decays under production pressure, the pilot's champion moves on, and eighteen months later a new pilot starts from zero. That is the lagging loop, and stage-3 plants run it with excellent data.

The plants that cross take the slower path: one model, one workflow, one write-back, with a human approving every action and a one-switch fallback to the old process. It is politically harder — writing into the MES means change control, validation in regulated environments, and a conversation with operations about trust. But it changes the physics of the programme. Output that lands in a work queue gets consumed as part of the job rather than checked as a favour; the approval log starts accumulating the evidence that later justifies autonomy; and the pipeline built for the first write-back is the template the second use case reuses. The gap between manufacturers does not open at stage 5. It opens here, at the dashboard.

Stage 4 is where the programme's character changes from analytical to operational. The questions that matter are no longer about model accuracy — they are about freshness alerting on the gateway feeds, drift behaviour across product changeovers, who is paged when the scrap-prediction service degrades, and how the quarterly saving is attributed when three initiatives touched the same line. Plants at this stage talk about their AI the way they talk about their utilities: mostly invisible, occasionally on fire, owned by someone specific.

The economics are now visibly different from the lagging loop. A deployed use case pays a measured, holdout-attributed return; that return funds the next deployment without a fresh capital case; and the operators' relationship with the system has shifted from being shown insights to processing recommendations — accepting most, overriding some, and generating with every override the labelled data that makes the next model version better. This is the compounding mechanism in miniature, and once a site has run it two or three times the internal argument about whether AI 'works here' simply ends.

What stage 4 has not yet solved is replication. The loop runs on one site because it was hand-built for that site: this historian's quirks, this MES vendor's API, this planner's queue. Copying the downtime use case to the sister plant means rediscovering all of it, and the copy costs eighty per cent of the original. That is the boundary between stage 4 and stage 5, and it is where most leading-loop programmes stall — not for want of ambition, but because templating your own work is nobody's urgent job. The plants that pull away are the ones that treat the second copy as a product decision: extract the template, standardise the interfaces, write the replication sheet, and make the next site a configuration exercise.

Stage 5 is what the lighthouse plants documented by the World Economic Forum's Global Lighthouse Network have in common, and it is less futuristic than its reputation. The defining property is not autonomy or lights-out operation — plenty of compounding plants keep a human on every consequential decision. It is that deployment has become standard work: a template library holds the proven use cases, a replication sheet tells a new site what to install and what to measure, and the marginal cost of the next deployment falls with each one shipped. The plant has industrialised its own improvement, which is the most manufacturing-native idea imaginable — it is what the industry did to production itself a century ago, applied to its decision-making.

Autonomy, where it exists, is narrow and evidence-bound. A compounding plant lets specific enumerated decisions — a setpoint nudge inside SPC limits, a routine reorder, a schedule swap below a value threshold — execute without approval, because months of approval logs showed the human accepting effectively all of them. Everything else escalates. The decision policy is versioned and reviewed like code, the audit trail can reconstruct any automated action months later, and the kill switch is drilled. In regulated non-automotive sectors — food, pharma adjacent, aerospace suppliers — this evidentiary discipline is not overhead on the way to autonomy; it is the thing that makes any autonomy permissible at all.

The uncomfortable property of stage 5 is what it does to everyone else, and it is why this page is about a gap rather than a ladder. A compounding plant's advantage grows without further strategic decisions: each deployment cuts cost or lifts yield, funds the next, and enlarges the data and template estate that makes the next cheaper. A lagging competitor is therefore losing ground during quarters in which it makes no mistake whatsoever. The gap is not a ranking that updates when you act; it is a differential equation that runs continuously. That is the strategic case for crossing early, and for crossing on one line rather than waiting for the perfect plant-wide moment.