Predictive AI adoption in manufacturing rose 12 percentage points to 48% in the most recent survey round, with process optimisation at 36% and supply chain planning at 35%. Those numbers describe intent and pilots. The systems genuinely running on a plant floor are a narrower and duller set, and the gap between the two is infrastructure rather than ambition.
Here is what the split actually looks like.
What are manufacturers deploying right now?
Three clusters, in descending order of how often they reach production: predictive maintenance, process and schedule optimisation, and supply chain planning. The reported adoption movement is real and the direction is consistent — 2026 is the year the sector moved from pilots toward execution.
| Use case | Reported adoption | Where it usually runs |
|---|---|---|
| Predictive maintenance | 48% | Sensor data, off the control loop |
| Process optimisation | 36% | Historian and MES data |
| Supply chain planning | 35% | ERP and supplier data |
Notice what those three have in common. None of them requires a model to touch a control system. They read data that industrial plants already emit and produce a recommendation a person acts on — which is precisely why they clear approval.
The systems that get talked about more and deployed less — autonomous agents acting on the line, industrial foundation models, full digital twins — are real, but they carry integration and safety costs that push them well past a first project.
Why do plant-floor projects stall more often?
Because the constraints are physical and organisational rather than technical. Roughly 88% of AI agents never reach production across all sectors, with infrastructure gaps the single largest attributed cause at 41% — and industrial environments have more infrastructure constraints than most.
The network is segmented, often deliberately. Plant networks are air-gapped or tightly firewalled for sound safety reasons. A design that assumes a cloud API call is dead on arrival, which is why on-premise or self-hosted inference is a design input in industrial work rather than a preference.
The data is in formats models find hard. Drawings, scanned specifications, twenty-year-old ERP tables, maintenance logs in free text. Extraction from these is solved, but only if you accept confidence scoring and human review on ambiguous cases.
OT and safety systems are rightly off limits. Anything touching a control loop carries a different approval burden entirely. The defensible scope is the data those systems emit, not the loops they close.
The knowledge is in people, not documents. The person who knows why line three runs at 82% has been there thirty years. That is a real asset and an unindexed one.
What ships before anything touches the plant floor?
The commercial workflows, almost always. They have the same shape that makes any automation survive — high volume, rule-heavy, low-judgement, with a human checkpoint that already exists — and none of the network or safety constraints.
RFQ and quoting. An RFQ that takes four days to price loses to a competitor who took four hours, and most of that time is retrieval: finding the comparable job, the current material cost, the right tolerances. Extraction plus matching against historical jobs turns an estimator's build into a review.
Specification and catalogue data. Turning PDF spec sheets into structured, searchable product data fixes site search, feeds the ERP, and makes capability data legible to the engines buyers now query — three returns on one piece of work.
Knowledge capture. Structured interviews with your longest-serving staff, transcribed and folded into a retrieval corpus alongside the drawings and maintenance logs. Unglamorous, and the only version of this that works is the one done before someone retires.
Each of these is a business systems and data foundations problem before it is a model problem. That ordering is not a preference — it is why the commercial side ships while the plant-floor project is still in review.
Are digital twins and foundation models worth it?
Eventually, and rarely first. Industry-specific agents running on industrial foundation models are a genuine development, and digital twins let you test configurations before changing anything physical. Both are substantial undertakings with long payback.
The honest sequencing test is whether you already have the inputs. A digital twin needs reliable, reconciled, well-instrumented data about the process it mirrors. If the historian disagrees with the MES and both disagree with the ERP, the twin inherits that disagreement and produces confident nonsense.
So the question is not whether the technology works. It is whether your data layer could support it today — and for most operators the answer is not yet, which makes the foundations work the actual first project.
How do you choose what to start with?
Score candidates on four things before writing code: volume, whether a written procedure exists, whether two experienced people would produce the same output, and whether someone already checks the result. Anything missing one of those will stall.
Then add the industrial filter. Does it require network access the plant will not grant? Does it touch a control loop? Is the source data trapped in a format nobody has extracted? Each yes moves it later in the sequence, not off the list.
The full version of that test, with the sector-by-sector comparison, is in which AI automation applications actually reach production — and the constraints specific to industrial environments sit on the manufacturing and industrial page.
The short version
What manufacturers deploy successfully is unglamorous: quoting, specification data, knowledge capture, maintenance prediction off sensor data. What stalls is anything that needs plant network access, touches a control loop, or assumes clean reconciled data that does not exist yet.
Start on the commercial side where the constraints are fewest and the payback is measurable, fix the data layer while you are there, and let the plant-floor work follow the infrastructure rather than lead it. If knowledge capture is the urgent one — because someone is retiring — that is a staffing deadline, not a technology one, and it moves first regardless of what AI enablement would otherwise sequence.