AI in Manufacturing Execution: What It Actually Does in 2026

Almost every defensible AI use on a shop floor falls into one of three verbs. Drafting: the model writes text a human would otherwise write from a blank box — a deviation narrative, an investigation summary, a recipe proposal from an existing SOP. Ranking: the model orders a queue by importance — which exceptions in a closed batch actually matter, which overdue task to chase, which of forty alarms is the real one. Explaining: the model correlates messy operational data and proposes the two or three most likely causes with the evidence attached — yield drift against shift, scale, ambient conditions and raw-material lot. Notice what is missing: deciding. Nothing in that list releases a batch, signs a record, or overrides a tolerance. That boundary is not timidity; it is what makes the rest of it installable in a GMP plant.

The single highest-value AI use in execution is also the least glamorous. When a step fails — a weight out of band, a temperature excursion, a missed in-process check — someone has to write a deviation. In most plants that is 30 to 60 minutes of an experienced person's time, spent reconstructing what the system already knows: which equipment, which lot, which operator, what time, what the tolerance was, what the reading was, and whether this has happened before on the same line. A model with access to the live step data writes that narrative in seconds. The investigator then does the part only a human can do: judge the root cause, decide the impact, and sign. The gain is not just time — it is that deviations get written at all, and written with the evidence attached, rather than compressed into 'operator error' at the end of a shift.

Review by exception is a twenty-year-old idea that most implementations get wrong: the system flags anything unusual, the reviewer receives 200 flags, and the review becomes scrolling again. Ranking is where a model helps. Given a closed batch, it can order exceptions by regulatory weight and product impact, group repeats that are obviously one event, and put the two that matter at the top. The reviewer still reads everything they are accountable for — but they read it in the right order, and the ones that are genuinely trivial are presented as a group rather than as forty separate decisions. Batch release cycle time falls because reviewer attention stops being spent uniformly on everything.

The slowest part of an MES implementation is rarely the software. It is converting hundreds of Word and PDF procedures into structured, executable steps with tolerance bands, signature points and witness rules. A model reads the procedure and proposes that structure: here are the steps, here is the sequence, here is where a second signature is implied by the wording, here is a tolerance that the text states ambiguously and a human should resolve. Engineering then edits and approves through normal change control — nothing goes live because a model suggested it. In practice this compresses first-recipe build time from days to an afternoon, which is the difference between a 90-day rollout and a two-year project.

Operators, supervisors and quality staff routinely need an answer that exists somewhere in the system: what does this step mean, which version of the SOP is effective, has this material failed before, what happened on the last three batches of this product. A grounded assistant answers from the effective record — not from general knowledge, and not from a document that expired last quarter. The engineering that matters here is grounding and version-awareness, not model size. An assistant that cheerfully answers from a superseded procedure is worse than no assistant at all, because it is confidently wrong in a way that shows up in an audit.

Three claims deserve scepticism. First, autonomous disposition — 'AI releases your batches'. No regulator anywhere accepts a model as the signing authority, and no vendor will indemnify you for it. Second, predictive quality presented as certainty — correlation on a few hundred batches is a hypothesis, not a control strategy, and treating it as one creates a validation problem rather than solving a quality one. Third, generic large-model chat bolted onto a database. If the assistant is not version-aware, not permission-aware and not grounded in the effective record, it will eventually cite a retired procedure to an operator. Ask any vendor to demonstrate their assistant answering from a superseded document — the good ones refuse and point you to the current one.

Four artefacts make AI in execution defensible. A documented intended use for each assist, narrow enough to test. A risk assessment treating the AI as a supporting, non-decision-making function under GAMP 5 Second Edition, with critical thinking applied to what happens if an output is wrong. Evidence of human review: the audit trail must record whether each suggestion was accepted, edited or rejected, and by whom. And a data statement — where inference runs, what leaves the tenant, and whether your records are used as training data. Under the EU AI Act, assistive human-in-the-loop use with documented purpose, oversight and traceability is the posture you want on file well before enforcement bites; retrofitting it after a deployment is considerably more painful.