Nexly AgentsBlog

Nexly Agents › Blog › Manufacturing

Manufacturing

Factories start AI in the office, not on the floor

Published 2026-09-16·3 min read·Author: Jakub Milewski, Nexly Lab

Photo: Unsplash

Manufacturers start AI on the floor, and that is why they so often end at the pilot. The data says those who succeed start in the office.

I say this as someone who deploys AI agents at manufacturers and hears almost the same thing at every first meeting: "we want to predict failures".

Eurostat shows what industry really uses AI for. In manufacturing, most often for marketing and sales (30 in 100 companies using AI) and for organizing administrative processes (27 in 100). Production processes come further down. The biggest difference between large and small plants concerns precisely production processes: 35 against 22 in 100.

Manufacturing has the lowest AI adoption among industries, around 12 percent in US data, but the fastest growth, more than two and a half times year on year. Deloitte reports that 84 in 100 manufacturers already see measurable value from AI, but only 20 in 100 use cases have been scaled beyond the first plant. The OECD, after interviews with European manufacturers, calls adoption modest and fragmented, with three areas of greatest potential: maintenance, quality control and the supply chain.

Now why "predicting failures" is the wrong first step. About 60 percent of predictive maintenance projects fail in the first year. 70 to 80 percent of the effort is not the model but data engineering. And without a failure history and sensors on the machines, the model has nothing to learn from. Most plants that want to start with prediction don't have the three to five years of consistent data it requires.

The order that follows from the data, from the software layer to the hardware layer:

  1. Search in machine documentation and manuals. A service technician asks about a torque value or an error code and gets an answer from the technical documentation, not from the memory of a colleague who is on holiday. Purely text-based, no sensors, no integrations.
  2. Quoting and requests for quotation. A first draft of the RFQ response from a template and quoting history. A human checks before it goes out.
  3. Shift reports. Summaries and classification of the causes of downtime and defects from what people enter anyway.
  4. Visual quality inspection, but only when there is steady volume and images of defects. Vendor data speaks of a 40 to 50 percent reduction in defects and payback in 7 to 8 months, but those are their best deployments. Start with one defect on one line, where a camera already stands.
  5. Demand forecasting, when there is history in the ERP. Data preparation is 60 to 80 percent of the work.
  6. Predictive maintenance at the end, after building the data and sensors.

The first point is exactly what we do for manufacturers as a technical assistant: machine documentation becomes the knowledge base of an agent that answers technicians and customers, also at night and in another language. Nobody on the floor changes how they work, and the engineer stops taking the same calls.

Limitation: an agent on documentation is only as good as the documentation. If the manual is a scan from 2009 without text, the first week is spent tidying it up.

A question for you: how many times this week did someone from service call an engineer with a question whose answer is in the manual?

Want to see how the agent handles your customers' questions? Enter your website address and e-mail in the demo form. In 20 minutes you get an agent trained on your website's offer, ready to be quizzed.

Sources

See the AI agent on your own website

Give us your website address and e-mail, and in 20 minutes you get an agent trained on your offer, ready to be quizzed.

Get a free demo

Read also

AI in e-commerce: start behind the scenes Accounting firms and law firms: three processes to start with and one trap Don't start AI with a revolution