An important distinction before anything else: "AI in manufacturing" covers three different technologies — computer vision for inspection, statistical models for predictive maintenance, and language models for information work. They solve different problems with different engineering. This article covers the third; conflating them during scoping causes expensive confusion.
Language-model AI has a narrower but genuine role on the production side. It is in the information layer, not the control layer.
Where it belongs and where it does not
| Layer | Example | Language models? |
|---|---|---|
| Machine control | Servo loops, PLC logic | No — deterministic real-time only |
| Safety systems | Interlocks, E-stop | No — certified hardware |
| Inspection | Defect detection | Computer vision, not language |
| Condition monitoring | Predictive maintenance | Statistical models, not language |
| Documentation | Work instructions, reports | Yes |
| Communication | Supplier and customer correspondence | Yes |
| Knowledge access | Procedures, specifications, history | Yes |
| Planning support | Drafting schedules for review | Yes, with human decision |
The rule is simple: language models handle information about production, never production itself. Anything with a physical consequence or a safety implication belongs in deterministic systems.
Where the value actually is
Technical knowledge access
Machine manuals, maintenance procedures, quality standards, past fault records. Scattered across binders, shared drives and individual memory. An agent grounded in this, accessible from the shop floor, answers questions that currently require finding the one person who knows.
Supplier and customer communication
Chasing deliveries, requesting certificates of conformity, responding to specification queries, following up on non-conformances. High volume, repetitive, and currently consuming production staff time.
Quality documentation
Writing up inspection findings, drafting non-conformance reports, summarising trends across batches. The inspection itself is human or vision-based; the documentation around it is the burden.
Maintenance history and triage
"This machine is doing X — has it happened before?" Searching unstructured maintenance logs semantically rather than by keyword surfaces past occurrences that keyword search misses.
Work instruction drafting
Producing first drafts of work instructions and setup sheets from engineering data, for review by someone who knows the process.
Maintenance log search is the most underrated application. Decades of fault records typically sit in free text nobody can search usefully. Semantic retrieval over that history turns "has this happened before?" from an unanswerable question into a thirty-second lookup — and the answer frequently contains the fix.
Shop-floor realities that shape the design
- Connectivity is unreliable. Anything that stops working offline gets abandoned.
- Hardware is modest — the terminal will not be new.
- Hands may be dirty or gloved — voice input and large touch targets matter.
- Noise affects voice interaction.
- Time pressure — if it is slower than asking a colleague, it will not be used.
- Language — regional language support is frequently necessary.
What the documentation problem really is
Manufacturing knowledge is disproportionately undocumented. The setup trick, the machine's quirk, the supplier who always ships late — this lives with experienced staff and leaves when they do. An AI grounded in your documentation can only surface what was written down, which means the documentation exercise is the real project. Frequently it is worth doing regardless of whether AI follows.
Integration with production systems
Value increases substantially when the agent can see live data:
- ERP — order status, material availability, schedules.
- MES — production status, batch records.
- Maintenance system — work orders, history, spares.
- Quality system — inspection records, non-conformances.
Older on-premise systems frequently lack clean APIs, which makes the access layer the substantial part of the project rather than the AI.
Where to start
- Index maintenance history and machine documentation. Internal, low risk, immediately useful.
- Deploy where staff already are — Teams, WhatsApp, or an existing terminal.
- Measure what gets asked — the questions reveal your documentation gaps.
- Add supplier communication drafting with human send.
- Connect to production systems once the access layer justifies itself.
Running production and wondering where AI genuinely fits? Tell us what your team looks up or writes repeatedly — we will be straight about what belongs in deterministic systems instead. See our AI agent service, CNC controller service, and internal knowledge bases.