Agentic AI in manufacturing is software that takes actions, such as entering an order, emailing a supplier or drafting a work order, rather than only producing a prediction or a summary for a person to act on. Its use is early. In a 2025 survey by the Manufacturing Leadership Council, the digital manufacturing division of the National Association of Manufacturers, 6 percent of manufacturers were using agentic AI and 24 percent expected to within two years. The Census Bureau's data shows where AI of any kind is used in manufacturing today: between November 2025 and February 2026, 2.2 percent of US manufacturers used AI in producing goods, against 12.8 percent in sales and marketing and 5.8 percent in finance and accounting. That pattern says a good deal about where agents will pay first.
What makes AI agentic
Three kinds of AI are in use in manufacturing, and they do different jobs. Analytical AI makes predictions from data: a model that forecasts when a bearing will fail, or a vision system that flags a defective part on the line. Generative AI produces content for a person: a draft of a standard operating procedure, a summary of a shift log, an answer to a technician's question. Agentic AI adds the ability to act. An agent reads a request, decides on the steps, calls the systems involved, and carries the task through, handing a person the cases it can't decide.
The distinction matters because action carries a different kind of risk. A wrong prediction produces a bad recommendation that a person can ignore. A wrong action produces a wrong order in the ERP, an email to a supplier, or, on the plant floor, a change to how a machine runs. Where an agent is allowed to act, and what checks sit between its decision and the consequence, is the whole design question.
Where agentic AI in manufacturing is being tried
The ambitions are broad. Deloitte's 2026 Manufacturing Industry Outlook describes agents working from the back office to production, with examples such as generating shift handover reports and initiating mitigation steps with suppliers, including contract negotiations, with human approval. The qualifier, with human approval, describes the state of practice accurately.
The most advanced factories show how far agents have come. The World Economic Forum's Global Lighthouse Network, the factories it recognizes as leaders in applying advanced technology, reported in January 2026 that AI agents enabled 5 percent of its top solutions in 2025, against 62 percent for analytical AI and 23 percent for generative AI. The same report states that fully autonomous agentic AI remains theoretical in the production context. In the factories the Forum holds up as leaders, then, the line runs on analytical models, and agents are only beginning to appear.
Why the plant floor keeps deterministic control
Production equipment is controlled by systems built to behave the same way every time: programmable logic controllers, distributed control systems and the safety systems that shut a process down when a limit is crossed. A language model behaves differently. It produces the output it scores highest, which is usually right and occasionally wrong in ways that are hard to predict in advance. That trade is acceptable for drafting an email and unacceptable for a safety interlock.
The security agencies that advise on industrial systems have said so directly. Joint guidance from CISA, the NSA, the FBI and allied agencies, published in December 2025, states that AI such as large language models almost certainly should not be used to make safety decisions in operational technology environments, and advises limiting active control of operational technology by AI without a human in the loop. The same guidance notes that language models are typically used in the business context. On the plant floor, agents read, summarize and recommend, and the controls act.
The paperwork around production is where agents pay first
A mid-market manufacturer's administrative work is large, repetitive and document-heavy. Customer orders arrive by EDI from large customers and as PDFs and emails from everyone else, and someone enters them. Purchase orders go out to suppliers, and someone chases the late shipments and updates the promise dates. Quality documentation, from certificates of conformance to nonconformance reports, has to be assembled for each shipment and customer. Maintenance requests become work orders, parts requests and records. Supplier invoices have to be matched to orders and receipts before they are paid, which is the work we walked through in three-way match, rebuilt.
An agent handles each of these the way an experienced coordinator would, with a record of every step. For an incoming order, it reads the customer's purchase order, maps the customer's part numbers to the manufacturer's own, checks price and lead time against the ERP, and enters the order or flags the lines that disagree. For a late shipment, it reads the supplier's acknowledgments and shipping notices, updates the promise date, and warns the planner when the new date threatens the production schedule. For a shipment's paperwork, it assembles the certificates and inspection records the customer requires. For a maintenance request, it turns a technician's note into a work order with the parts and the machine's history attached. The decisions stay with people: accepting a deviation, committing a date to a customer, renegotiating with a supplier, releasing a nonconforming part.
These tasks share the traits that suit an agent: high volume, documents that vary in format, rules the team can state, and a system of record where the result is written. The Census figures above show manufacturers already moving in this direction, with AI use in sales and marketing nearly six times its use in production, and in finance and accounting more than twice. The same pattern holds in distribution, where order entry and pricing exceptions carry most of the labor, as we described in process redesign inside a wholesale distributor. On the selling side, the path from a customer's order to the cash that pays for it is set out in order to cash.
Why manufacturing AI pilots stall
Manufacturers report the same obstacles as other industries, with more old systems in the way. In the Manufacturing Leadership Council's 2025 survey, 78 percent named legacy systems and integration among their top obstacles, and 66 percent reported difficulty scaling AI pilots into full production. A pilot built on a clean extract of data avoids the old ERP, the machine network and the supplier portals, and the production system has to connect to all of them. Writing results back into systems with no usable interface is its own engineering problem, one we cover in writing back into a system with no usable API. Gartner's broader forecast points the same way: in June 2025 it predicted that over forty percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.
A pilot also needs an owner once it works. A project team that builds an agent and moves on leaves the plant with software nobody is accountable for, and the documents, suppliers and systems it depends on keep changing.
A reasonable counter, answered
A reasonable counter is that the value of AI in manufacturing is on the line, in yield, scrap, throughput and unplanned downtime, and that automating the order desk is small change beside a point of overall equipment effectiveness. At a large plant with sensors, a data historian and engineers to maintain the models, that is often right, and most of the Lighthouse factories' top solutions use analytical AI. Those programs take years and a data infrastructure most mid-market manufacturers don't have yet. The administrative agents run on documents and systems the company already has, their results show up in the monthly numbers, and the hours they return are the hours of the same people a plant-floor program would need. Starting in the office is the order of work that gets a mid-market manufacturer to the line with the budget and the people to do it.
Common questions
- What is agentic AI in manufacturing?
- Agentic AI is software built on a language model that carries out multi-step tasks, calling other systems and taking actions, rather than only producing a prediction or a draft for a person. In manufacturing it is being applied to work such as entering customer orders, following up with suppliers on late shipments, assembling quality documentation and drafting maintenance work orders. On the production floor it mostly advises people and existing control systems rather than acting on equipment. Use is at an early stage: a 2025 Manufacturing Leadership Council survey found 6 percent of manufacturers using agentic AI.
- How is AI used in manufacturing today?
- Mostly in the office rather than on the line. The Census Bureau found that between November 2025 and February 2026, 12.8 percent of US manufacturers used AI in sales and marketing and 5.8 percent in finance and accounting, while only 2.2 percent used it in producing goods. In the most advanced factories, analytical AI such as predictive maintenance and inspection models accounts for most deployments. Agents built on language models are newer, and they are concentrated in administrative and planning work.
- Why do industrial AI pilots fail in manufacturing?
- The most common reasons are integration and scale. In a 2025 Manufacturing Leadership Council survey, 78 percent of manufacturers named legacy systems and integration among their top obstacles, and 66 percent reported difficulty scaling AI pilots into full production. A pilot built on a clean extract of data doesn't have to connect to an old ERP, a machine network or a supplier portal, and the production system does. Pilots also tend to lose their owner when the project team moves on, so they never become part of how the plant runs.