Industrial AI will be won on the factory that already exists
Industrial AI is where a model meets a schedule, a tolerance, and a person who will be blamed if the batch is wrong. The productivity story is real, and it is mostly unglamorous: fewer unplanned stops, faster quality decisions, cleaner handoffs through a supply chain, and design cycles that reuse what the plant already learned. The UK’s edge is not a greenfield AI factory. It is high-mix, regulated, export-facing industry that already records more than it uses.
Look at where the work actually sits. Design teams drown in variants and past non-conformances. Production loses hours to changeovers and to quality checks that wait on a person with a gauge. Maintenance is still calendar-based in plants that already stream vibration, temperature, and energy. Supply chains break on exceptions, not on the average week. Each of those is an AI problem only after it is a data and workflow problem. A model trained on a slide of “Industry 4.0” does not know which tag on the historian is the one that matters.
The competitive edge is in sectors that cannot treat data as optional: aerospace and defence supply, automotive, pharmaceuticals, food, speciality chemicals, and the machinery firms that sell into them. Those plants already live with traceability, batch records, and customer audits. Industrial AI that shortens a deviation investigation, flags a drift before scrap, or proposes a maintenance window the planner can accept is worth more there than a generic copilot. The export is the method: how a mid-sized British plant put intelligence on an existing line without replacing the systems of record.
That method is decision support, not a line that runs itself. On a UK operational system we have delivered, outside a factory, the useful pattern was the same shape. The model drafted the next step from records the organisation already held. A named person accepted or rejected it, and the decision was written to a log. The system of record stayed where it was. Nothing was allowed to complete a consequential action on its own. Plants should ask for that shape when a model proposes a quarantine, a setpoint, or a maintenance window: a proposal, a named acceptance, and a record of what the model saw. Buyers should ask vendors for the handover, the log, and the behaviour after a model change, not for a claim that the line will run itself. Autonomy is a later step, and only for tasks whose failure modes are bounded. Most of the value in the next five years sits inside a process that already has a controller, a permit, and a quality signature.
Industry’s job follows from that. Historians, manufacturing execution, quality, and maintenance systems already hold the record. What they often lack is an owner, a dictionary, and a rule for who may train on what. Firms should fund that layer before they fund another proof of concept. The system of record stays. AI proposes. A named role accepts. The best labels for a quality model come from people who already reject parts, so the work has to sit on the floor, not only in a central data team. Competing plants can still align on how a defect is described, how a stop is coded, and how a model update is released. That shared language is what lets suppliers and customers automate across a chain.
Government already has the right instruments, and the Industrial Strategy should point them at this problem rather than at a new showcase site. The Advanced Manufacturing Sector Plan, published with the Modern Industrial Strategy in June 2025, commits up to £4.3 billion for the sector, including up to £2.8 billion in research and development over five years, and up to £99 million from 2026 to continue Made Smarter Adoption. Over the three years to 2024/25 that programme gave more than 1,600 manufacturing businesses a digital roadmap, trained more than 400 SME leaders, and brought in over £16 million of private investment. Those roadmaps are the right unit of help: someone who will sit with the process engineers, name the historian tags that matter, and stop at a workflow a planner will actually accept.
Three uses of that money would match the day’s question. First, score publicly supported projects on whether they run on a live UK line, whether a person still accepts the consequential step, and what did not work. Lessons on data readiness are worth more than another autonomous-cell photograph. Second, spend the skills money on the hybrid role, the person who can read a process and a software interface. Made Smarter’s digital internships are a start. The bottleneck is not a shortage of model demonstrations. It is a shortage of people who are trusted on the floor and literate in the integration. Third, use the data-sharing infrastructure the sector plan funds from April 2026, up to £12 million, to back common descriptions of defects, stops, and model releases. Shared language across a supply chain is an industrial-strategy outcome. A second factory statute is not. Safety and product rules already exist. AI used inside them should be evidenced, not exempted and not duplicated.
The High Value Manufacturing Catapult, including the Advanced Manufacturing Park in South Yorkshire, already sits next to working plants. The next test is whether the adoption money reaches the mid-sized firm whose systems of record are old, audited, and not going to be replaced this decade. The UK will not out-scale every country on volume. It can be the place trusted to apply AI inside difficult industrial processes, with the records to prove it, one workflow at a time.
