Blog Industries

AI for manufacturing companies: mid-size plant use cases.

Which AI projects pay off first for small and mid-size manufacturers, what each needs to connect to, where a person stays in charge, and how EU and OT security rules apply.

Fig. 0Manuals, drawings and history, one answer

For a mid-size manufacturer, the AI projects that pay off first are rarely the ones in the keynotes. The faster wins are in the office and on the maintenance floor: quoting from customer drawings and specifications, keeping quality and compliance paperwork straight, answering technicians’ questions from manuals and service history, and flagging equipment that’s drifting toward failure. Start with one of them, with an estimator or quality engineer approving the output.

This guide covers those use cases, what each one needs to connect to, where AI is the wrong tool, and which rules apply as of October 2026.

Which AI projects pay off first in a mid-size plant?

Fully autonomous factories and digital twins are mostly big-company projects. These five are not, and each can be built and measured within a quarter.

1. Quoting from RFQs, drawings and specs

Requests for quote arrive as emails with PDFs, drawings and spec sheets. An AI system reads them and extracts materials, dimensions, tolerances, finishes, quantities and the drawing revision. It checks for missing information, then assembles a draft quote from your costing rules for an estimator to approve. Faster quotes win more work, and the estimator spends their time on judgement rather than retyping. Construction tenders and RFIs raise a similar document challenge, covered in AI in construction.

Two design choices matter here. First, the model reads and the code calculates: routings, machine rates and material tables live in your ERP or a costing sheet, and the model never does the arithmetic or fills a gap with a guess. Second, every extracted field carries a pointer to where it came from on the drawing, so the estimator can check a tolerance in seconds. If a callout is unreadable, the system asks the customer rather than assuming.

2. Quality and compliance documents

Certificates of conformity, mill test certificates, inspection reports, first-article packages, non-conformance reports and ISO documentation pile up and have to match each other. AI can extract, file and cross-check them. It flags a material cert that doesn’t match the order, a heat number that isn’t on the purchase order, or an inspection record with a missing signature, before an audit or a customer finds it.

3. A maintenance assistant that knows your machines

Retrieval-augmented generation, an assistant that answers from your own documents, is a natural fit for maintenance. Technicians ask in plain language and get answers from equipment manuals, past work orders and engineering notes, with a reference to the source page. It has to say when it doesn’t know. Our note on retrieval that knows when to say “I don’t know” explains why that matters more than fluency. On a plant floor, a confident wrong torque value is worse than no answer.

4. Predictive maintenance

If your equipment already reports vibration, temperature, current or cycle data, a model can learn each machine’s normal behaviour and flag deviations early. We run this kind of system across more than 200 solar plants with 95% predictive maintenance accuracy. The same approach applies to motors, pumps, compressors and presses. The precondition is data: consistent sensor readings with reliable timestamps, and a maintenance history to learn from.

The usual failure is not the model but the alerts. Joint guidance from CISA and partner agencies on integrating AI into operational technology says false alarms should be rare and carry context, so operators don’t tune them out. We treat alert precision as a number to test before go-live, not after.

5. Production and shift reporting

Turning MES, ERP and spreadsheet data into a readable shift summary, a weekly report or an answer to “why was line 3 down on Tuesday?” is a good job for a language model. Your own data and queries supply the numbers, and the model writes the explanation around them.

Which kind of AI fits which job?

Not every plant problem needs a language model. Match the job to the tool, then check what data it needs and what can go wrong.

Use caseRight kind of AIData neededMain risk
Quoting from RFQs and drawingsDocument AI plus a language model; costing in codePast quotes, routings, rate tables, 50 real RFQsA misread tolerance reaches a customer
Cert and inspection cross-checksDocument AI with rule-based matchingOrders, certs, inspection records, quality proceduresA missed mismatch looks like a pass
Technician questionsRetrieval (RAG) over manuals and work ordersCurrent manuals, work-order history, engineering notesAn outdated manual gives a confident wrong answer
Early warning on equipmentAnomaly detection or machine learning on sensor dataTimestamped sensor history, maintenance recordsAlert fatigue, or missed early signs
Visual inspection of partsComputer vision modelsLabelled images of good and bad parts under steady lightingNew defect types the model never saw
Scheduling and sequencingOptimisation solversOrders, capacities, changeover timesPlans that ignore constraints nobody wrote down
Energy and scrap analysisForecasting and statistical modelsMeter data, batch and process recordsSpurious patterns from short histories
Shift and management reportsLanguage model over your own queriesMES, ERP and spreadsheet extractsNumbers retyped by the model instead of queried

A language model still helps around the specialist tools: explaining a schedule, summarising what a vision system rejected last week, or drafting the corrective-action report. It shouldn’t be the part that decides.

Which systems does it need to connect to?

This is where most of the effort goes. A plant’s information lives in an ERP (SAP, Epicor, Infor or Microsoft Dynamics, for example), a MES, a CMMS such as IBM Maximo, a quality system, shared drives and a lot of email. A useful AI system reads from these and writes back to them, rather than becoming another screen.

  • ERP and quoting: read customer, part and routing data; write the draft quote or sales order in a draft state for a person to release.
  • CMMS and maintenance: read work orders and asset records; create a work-order draft when an alert is confirmed.
  • Sensor and machine data: read from a historian or a copy of the data, not from the controllers themselves.
  • Email and drive: the entry point for RFQs, certs and reports, and the place where the messiest inputs live.

Old systems with limited APIs are common, so integration is usually the biggest piece of work in a project, not the AI. Treat everything arriving from outside as untrusted: a supplier’s PDF or a customer’s email can contain instructions aimed at your system. Our note on prompt injection explains the controls. Outbound shipping and carrier documents are covered in AI for logistics companies.

What about the plant network and cybersecurity?

Office AI, such as quoting and document checking, doesn’t need to touch the plant network at all. Maintenance AI that uses machine data does, and that’s where a connected plant needs care.

The CISA-led guidance recommends push-based architectures, where data is pushed out of the operational technology network for the AI system to use, without persistent access into that network. We design to the same principle: the model works on a copy of the data, outside the control network. The broader reference for securing operational technology is NIST SP 800-82 Rev. 3, the guide to OT security, which covers industrial control systems, SCADA and PLC environments. NIST has a draft of Revision 4 out for comment, so check the current version.

Where AI is the wrong tool, and where a person stays in charge

Some decisions should stay with people, and some should stay in deterministic code. On a plant floor that line is sharp:

  • Safety decisions. The same CISA-led guidance says AI such as language models “almost certainly should not be used to make safety decisions for OT environments.” It also advises limiting active control of OT by AI without a human in the loop, and building failsafe mechanisms that revert to traditional automation or manual operation.
  • Robot and machine motion. Cobot and machine safety functions belong in their certified safety controllers. AI can suggest or explain; it doesn’t switch a guard or change a speed limit.
  • Product release. A quality engineer decides whether a part ships. AI can prepare the file and flag what looks wrong.
  • Quote commitments. An estimator approves quotes with new materials, unusual tolerances or customer-specific terms. Standard repeat quotes can move to review-by-exception once the numbers support it.
  • Writing to PLCs or setpoints. Don’t, in a first project. Recommendations go to people.

Behind each of these sits a fallback: if the model is unavailable or unsure, the work goes to a person with the context attached, and the plant keeps running.

How many manufacturers already use AI?

National statistics measure AI use by all businesses, not manufacturers alone, and each survey asks the question differently. The sources below don’t publish a manufacturing-only figure in their text, so treat these as context, not a league table:

MarketBusinesses using AISource and scope
EU average20.0% (2025), up from 13.5% (2024)Eurostat, enterprises with 10+ employees
Denmark (highest in the EU)42.0% (2025)Eurostat
United KingdomAbout 35% (June 2026), up from about 12% in late 2023ONS Business Insights and Conditions Survey, businesses with 10+ employees
United States17% to 20% (Dec 2025 to May 2026); 37% at firms with 250+ employeesUS Census Bureau BTOS, AI used in business operations

Sources: Eurostat, ONS and the US Census Bureau. The direction is the same everywhere: adoption has risen fast in a year or two, and larger firms lead. A mid-size manufacturer starting now is early, not late. In the UK, the government-backed Made Smarter programme offers regional support to small manufacturers adopting digital technology.

Which rules apply?

This is a summary of what the rules say, not legal advice. A qualified product-compliance or legal professional decides what applies to your machines and your sales.

EU AI Act

Most of the projects above, such as quoting, document handling, maintenance assistants and reporting, are usually not high-risk under the EU AI Act. They carry transparency duties at most. AI built into regulated products is different. The European Commission’s AI Act page says the high-risk rules for AI embedded in products covered by EU product safety law apply from 2 August 2028, after the AI Omnibus entered into force on 27 July 2026. Stand-alone high-risk uses in sensitive areas apply from 2 December 2027. If you build machines with AI-driven safety functions, start designing for those requirements now. Our EU AI Act guide covers the full timeline.

EU Machinery Regulation

The new Machinery Regulation (EU) 2023/1230 applies on a mandatory basis from 20 January 2027, and the Commission’s machinery page says it integrates provisions for machinery with AI-powered safety functions and for cyber-safety. Machinery placed on the market before that date must comply with the current Machinery Directive. This matters to machine builders more than to plants that only use machines.

United States and United Kingdom

In the US, the voluntary NIST AI Risk Management Framework is the common reference for governing, mapping, measuring and managing AI risk, and NIST notes it is being revised. NIST also published a generative AI profile for it in July 2024. We use it as a checklist for what to test and who owns it. For UK-specific rules, see our UK AI rules guide.

How do you start, and what do you measure?

  1. Pick one flow with a number: quote turnaround, hours spent on documents, or unplanned downtime.
  2. Collect fifty real examples: real RFQs, real certificates, real fault records. They become the test set that tells you whether the system works. Our note on evals before features explains how we use them.
  3. Keep a person in the loop at first. Estimators approve quotes and quality staff confirm flags. Automate further only once the numbers support it.
  4. Integrate with what you run, whether that’s your ERP, CMMS, MES or shared drives, rather than adding another place for people to look.
  5. Plan for the months after launch. Drawings, suppliers and machines change, and models drift. Plan monitoring and re-testing from the start.

Track these numbers from day one:

  • Quote turnaround: time from RFQ received to quote sent, and estimator minutes per quote.
  • Field accuracy: share of extracted fields that match what a person would enter, measured on the fifty examples and re-checked monthly.
  • Mismatches caught early: certificate and record errors found before an audit or a customer.
  • Alert precision: share of maintenance alerts that led to a real finding, alongside unplanned downtime.
  • Answer quality: share of technician answers that cite the right source, and how often the assistant correctly said it didn’t know.

The best first AI project in a plant is one your most sceptical engineer can check, and ends up relying on.

If you want help scoping the first one, our services cover document AI, retrieval assistants and predictive systems.

Questions manufacturers ask

Our documents are scanned, old and inconsistent. Is that a problem?

It’s normal. Current models read scans and mixed layouts well. The work is in testing on your worst documents and flagging low-confidence fields for review.

Do we need a data lake first?

No. A focused project needs access to the specific documents or data for one flow. Wider data work can follow once the first system has shown its value.

Can it run on our own servers?

Yes, with open-weight models, with more hosting and engineering work than managed APIs. Many plants keep sensitive drawings on-premises and use managed models for less sensitive work.

Can AI inspect parts on the line?

Yes, with computer vision models rather than language models. They need labelled images of good and bad parts under steady lighting, and they struggle with defect types they haven’t seen, so a person reviews anything the model flags as uncertain.

Can predictive maintenance work if we have few recorded failures?

Often, yes. Failures are rare, so the usual approach is to learn each machine’s normal behaviour from sensor data and flag deviations, then confirm with technicians. Consistent timestamps and a clean maintenance history matter more than a long list of breakdowns.

Start with one flow

Modulus Labs AI builds document AI, maintenance assistants and predictive systems for operations-heavy businesses. Tell us which flow takes your team the most time, and we’ll tell you what we’d build first.