Most manufacturers I talk to are not short of AI ideas. They are short of AI results. There is a vision pilot that worked beautifully on one line and was never rolled out, a forecasting model nobody in planning trusts, and a chatbot demo that impressed the board and then quietly disappeared. Meanwhile the vendor emails keep coming.
In my 24 years in manufacturing and industry, I've learned that the problem is rarely the technology. It is the choice of the first problem. Pick a use case with no clean data, no owner and no decision attached to it, and even a brilliant model will die in pilot. Pick a dull, well-defined one, and AI quietly starts paying for itself.
The numbers say the same thing. In McKinsey's 2025 global survey, only 39% of respondents reported any enterprise-level EBIT impact from AI, and most of those said it was under 5%. Yet manufacturing was one of the functions where people most often reported cost benefits from individual use cases. Deloitte's 2025 smart manufacturing survey of 600 large manufacturers found 29% using AI/ML at facility or network scale and another 23% still piloting. So the value is there; it just doesn't arrive by itself.
Five use cases that keep showing up in real factories
These are not futuristic. Each one has mature tools today, and each one is tied to a number a plant manager already watches.
- Visual quality inspection. A camera and a vision model check parts at the end of a station and flag scratches, missing components or wrong labels. Pays through fewer escaped defects and less manual inspection.
- Predictive maintenance on one critical asset. Vibration, motor current or temperature data warns you days before a failure. Pays through less unplanned downtime, but needs sensor data and some failure history.
- Demand and sales forecasting. A machine-learning forecast on two or three years of sales history, with promotions and seasonality, gives planning a better starting number. Pays through lower stock and fewer rush orders.
- A generative-AI assistant for maintenance and operations. Technicians ask questions in plain language and get answers drawn from your own manuals, SOPs and past work orders, with the source page shown. Pays through faster repairs and faster onboarding.
- Energy anomaly detection. A model learns the normal energy profile of compressors, ovens or HVAC from meter data and flags waste, like air leaks at the weekend. Pays through lower energy bills.
You will also hear a lot about "agentic AI", software agents that act across several systems on their own. It is coming: in a Manufacturing Leadership Council survey cited by Deloitte, 6% of manufacturers were using agentic AI and 24% expected to within two years. But agents need exactly the same foundations as everything above, clean data, clear owners and a decision to support, so the choice of a first use case matters even more.
Worked example: scoring them in one hour
Here is the simple method I like to use with a management team. For each candidate, score three things from 1 to 5: Value (how much money or risk is at stake), Data readiness (does the data already exist, in usable shape, with an owner) and Effort (integration, change management, hardware). Then calculate Score = Value × Readiness ÷ Effort. It is deliberately crude; the conversation it forces is the point.
The scores below are an illustrative example for a mid-sized discrete manufacturer, not data from a real company.
| Use case | Value | Data readiness | Effort | Score |
|---|---|---|---|---|
| GenAI assistant over manuals, SOPs and work orders | 3 | 4 | 2 | 6.0 |
| Demand / sales forecasting | 3 | 4 | 2 | 6.0 |
| Visual quality inspection on one station | 5 | 3 | 3 | 5.0 |
| Energy anomaly detection | 2 | 3 | 2 | 3.0 |
| Predictive maintenance on a critical press | 4 | 2 | 4 | 2.0 |
Notice what happens. Predictive maintenance, the use case everyone mentions first, lands at the bottom, not because it is a bad idea but because in this example there is almost no failure history and no sensors yet. That doesn't mean "never"; it means "start collecting the data now and come back in a year". The two quick wins are the document assistant and the forecast, and visual inspection is a strong second wave.
A quick payback check on the winner
Before anyone spends money, put a rough number on it. All figures below are illustrative assumptions; replace them with your own.
| Item | Assumption (example) | Per year |
|---|---|---|
| Technician time saved searching manuals and old work orders | 25 technicians × 10 min/day × 250 days ≈ 1,040 h × 15 per hour | ≈ 15,600 |
| Faster repairs on line-stopping breakdowns | 120 breakdowns × 15 min shorter × 2,000 contribution per line-hour | 60,000 |
| Total benefit | ≈ 75,600 | |
| One-off cost (document clean-up, setup, integration, training) | 20,000 | |
| Running cost (licences, hosting, support) | 6,000 |
Payback in this example: 20,000 ÷ (75,600 − 6,000) ≈ 0.29 years, roughly three and a half months. Notice that most of the value comes from downtime, not from saved minutes. That is typical: tie the use case to a number that hurts, and the business case writes itself. If the downtime line were zero, the case would be much weaker, and it is better to know that before the pilot than after.
The one-page use-case card
Every candidate that survives scoring gets one page, no more. If you can't fill it in, you're not ready to start.
| Field | What to write |
|---|---|
| Problem | One sentence a supervisor would agree with |
| KPI and baseline | Today's number, measured the same way you'll measure the pilot |
| Data and owner | Which system, how many months, who signs off on quality |
| Decision it changes | Who acts on the output, and what they do differently |
| Success threshold | The result at day 90 that justifies scaling |
| Kill criterion | The result at day 60 that stops it, with no hard feelings |
A 90-day plan that avoids "pilot forever"
- Weeks 1–2: fill in the card, measure the baseline and check the data with its owner.
- Weeks 3–6: build the smallest working version for one line, one team or one product family.
- Weeks 7–10: run it alongside people, not instead of them, and log every wrong answer.
- Weeks 11–13: compare against the baseline and decide: scale, fix or stop.
The payoff
When the first use case is chosen this way, you get something more valuable than a demo: a result the plant believes, a team that has done it once, and data pipelines you can reuse for the next case. The second project costs less than the first, and the third less again. That is how AI moves from a slide in the strategy deck into the daily rhythm of the factory.
It also connects to the rest of your house. AI needs to know which systems hold which data, which is why I wrote about turning static architecture diagrams into a living model. And the results need to be visible, so a single live dashboard is often the natural partner of a first AI project. If you've read why OEE gains don't always reach profit, the same rule applies here: measure the money, not just the model.
Checklist for this month
- List five to eight AI ideas already floating around your company.
- Score each one for value, data readiness and effort in a single one-hour meeting.
- Do a rough payback for the top two, with clearly labelled assumptions.
- Fill in the one-page card, including the kill criterion.
- Measure the baseline before anything is built.
- Agree who will act on the output and what they will do differently.
- Start the 90-day clock on one line or one team only.
Want help choosing your first AI use case?
If your company has more AI ideas than results, I'd be glad to help. We can run the scoring session together, build a realistic payback with your own numbers, and set up a 90-day pilot that ends in a clear decision, not another slide.
Contact me Read: AI in enterprise architecture
Sources
- McKinsey & Company (2025). The state of AI in 2025: Agents, innovation, and transformation. 39% report any enterprise-level EBIT impact; cost benefits from use cases most often reported in software engineering, manufacturing and IT.
- Deloitte (2025). 2025 Smart Manufacturing Survey. 600 manufacturing executives; 29% using AI/ML at facility or network scale, 23% piloting.
- Deloitte (2025). Agentic AI in the manufacturing industry, citing a Manufacturing Leadership Council survey: 6% of manufacturers using agentic AI, 24% expecting to within two years.
- NIST. AI Risk Management Framework (AI RMF 1.0).