AI Agents in Manufacturing: Where They Actually Fit

AI agents in manufacturing fit best in the information work around the plant, not in machine control. That means supplier and purchase-order exceptions, nonconformance reports and CAPA drafts, maintenance work-order triage, scheduling questions, RMAs, and compliance records. The agent reads and drafts in ERP, MES, CMMS, and QMS. A person approves record changes. Controllers keep running the line.

Below: survey evidence, a use-case table to score your plant against, a worked example, and why closed-loop control is a different problem.

What are manufacturers actually using AI agents for?

Adoption is broad and shallow. In the RSM Middle Market AI Survey 2026, 56% of the 129 manufacturing respondents said they use agentic AI and 79% said they use generative AI (RSM, July 2026). The Manufacturing Leadership Council’s 2026 Industrial AI survey found 70.9% of respondents use generative AI tools such as ChatGPT or Microsoft Copilot in manufacturing operations, and 66.6% use or plan to use agentic AI tools (MLC, April 2026). MLC does not state its sample size on that page.

Among MLC respondents, knowledge management led the functions where generative AI was implemented (56.6%), followed by process improvements and communications (38.3% each) and root cause analysis (33.3%). CNC machining and adaptive control came in at 8.3%. And 50% said they have no specific metrics for AI’s impact in operations.

The leading plants show the same pattern from the other side. McKinsey’s January 2025 update on the World Economic Forum’s Global Lighthouse Network covered the 18 sites added that month (13 Fourth Industrial Revolution and five Sustainability Lighthouses). For that cohort, it reported that 77% of the top five use cases were enabled by analytical AI and 9% used generative AI (McKinsey). Analytical models drive the plant-side results. Language models land first where the inputs are documents.

My read on the mechanism: an LLM agent is good at reading messy text, finding the right record, and drafting the next step. Plants generate a lot of that work in quality, maintenance, purchasing, and customer service. Those records also have a reviewer and an undo button. A PLC output does not.

Where does an agent sit in the plant? Use the ISA-95 levels

ISA-95, the enterprise-control integration standard, splits a manufacturing company into five levels. The OPC Foundation’s reference text summarizes them. Level 0 is the physical process. Level 1 is sensors and actuators. Level 2 is PLCs and distributed control systems, with “real-time responses measured in sub-seconds.” Level 3 is manufacturing operations management: maintenance, quality, inventory movement, scheduling, historians, and MES, working in “days, shifts, hours, minutes, and seconds.” Level 4 is ERP, working in “months, weeks, and days.”

An LLM agent is a Level 3 and Level 4 tool. It responds in seconds to minutes, its output can vary for the same input, and its work is easiest to check when the output is a record a person can read. Those three traits are fine for a work order and wrong for a control loop.

The decision rule I’d use: an agent may read from any level through an approved data path, and may write to Level 3 and 4 systems only through the approval steps a human clerk uses. It gets no write path to Levels 0 through 2. If a use case needs one, it is a control-system change, owned by controls engineering under the plant’s management-of-change process.

So most of the integration work is ordinary: APIs and permissions in ERP, MES, CMMS, and QMS, plus read-only historian summaries. For the general scoring method, see which workflows to automate first.

Use-case table: eight manufacturing processes scored

Use this table to rank candidates. “Write risk” is what happens if a wrong draft goes in and nobody catches it. “OT exposure” is whether the agent needs any data path into the control network. “First-pilot fit” is my judgment, not a benchmark.

Two rules for reading it. First, if write risk is medium, run draft mode for the first 30 to 60 days: a named person commits every record, and you log how often they edit the draft. Where to put those gates is covered in human in the loop for AI agents. Second, if a row needs anything beyond a read-only, brokered feed from OT, move it to the back of the queue. The best first pilots have no OT exposure at all.

Supplier, scheduling, and RMA exceptions on the ERP side

Deloitte’s 2026 Manufacturing Industry Outlook lists what supply chain agents can do: monitor disruption sources, alert the right people, quantify impact, recommend alternative suppliers, and “initiate mitigation steps, including contract negotiations, with human approval.” Even that example keeps the approval step.

The ERP failure mode is specific. A PO date change is not a local edit. It feeds MRP, which can reschedule planned orders, move work orders, and change what the floor builds next week. An agent that updates confirmed dates from every supplier email will make the plan jumpy, even when each update is correct. Planners call this nervousness.

So the agent drafts and the buyer commits. Set a threshold. If a supplier moves a date by less than the item’s safety lead time, the agent queues the change for batch approval. If the move is larger, or the part feeds a constrained work center, the agent escalates with the affected orders listed.

Scheduling follows the same rule. MLC respondents rated production planning among the least mature AI areas, with 0% calling it advanced. An agent that answers “why is job 4471 late, and what are my options” is useful today. One that resequences the line by itself is not.

RMAs are document matching too: serial number, ship date, and warranty terms checked against the claim before anyone issues a credit. The freight exceptions post covers that pattern.

Quality records: NCRs, CAPA drafts, PPAP, and ISO 9001

Quality is the most document-heavy part of a plant, and its rulebooks just changed. ISO 9001:2026 was published on September 16, 2026 and replaces ISO 9001:2015. Per ISO/TC 176 transition guidance, organizations certified to ISO 9001:2015 have until September 30, 2029 to complete the transition. In US medical devices, the FDA’s Quality Management System Regulation took effect February 2, 2026 and incorporates ISO 13485:2016 by reference.

Here is where an agent helps. An NCR usually starts as a short operator note, a photo, and a lot number. The agent can structure it, pull the part’s prior nonconformances, the supplier’s recent lots, inspection data, and open deviations, then draft containment and a 5-why outline for the quality engineer. For PPAP, it can check that every document carries the same drawing revision.

Here is where it hurts. A language model writes a plausible root cause as easily as a true one. If an engineer accepts “operator error” from a draft because it reads well, the CAPA closes and the defect comes back.

Detect it with two numbers: CAPA recurrence (same failure mode, same part, within 12 months) before and after the agent, and how often engineers keep the drafted root cause unchanged. If acceptance is high and recurrence rises, treat it as a warning signal and review a sample of cases. Rubber-stamping is one explanation. A shifting defect mix, a process change, or corrective actions that never worked are others. The hallucination budget post covers how to set a tolerable error rate per process.

The rule: the agent drafts evidence and structure. A named engineer owns root cause, disposition, and sign-off.

Maintenance work orders and technician lookup

Maintenance has the most published evidence, though all of it is self-reported. Siemens said in March 2025 that “first pilot use cases have shown” its Industrial Copilot for maintenance “helps save on average 25% reactive maintenance time” (Siemens press release). Siemens did not publish the pilots’ size. McKinsey’s 2024 write-up of the Lighthouse network describes ACG Capsules, a contract manufacturer in Pithampur, India, which built an assistant over its SOPs and policies in about two weeks. Within five weeks, nearly three-quarters of operators and technicians were using it, and ACG reported a 40% reduction in mean time to repair and unplanned downtime (McKinsey, April 2024).

My read on the mechanism: a lot of repair time goes to finding things, such as the asset’s history, the last fix that worked, the right manual page, and part stock. Work-order text is short, abbreviated, and inconsistent across shifts. A language model reads that text better than keyword search does.

The failure mode is the wrong document. Plants run several revisions of the same machine, often with scanned manuals. An answer from the Rev B manual when the line runs Rev C sounds right and is wrong. Require every answer to cite document, revision, page, and the matched asset ID. If it cannot match the asset, it should say so and stop.

Emergency and safety calls bypass the agent. A line-down call or a lockout question goes straight to a person.

Worked example: work-order triage at one plant (hypothetical)

Say a plant receives 1,200 corrective work requests a month. A planner reads each one, finds the asset, checks history, sets priority, and attaches manuals and parts. Assume 9 minutes per request: 180 hours a month.

With an agent, each request arrives as a draft. Assume the agent marks 25% as low confidence, which still take 9 minutes. The other 75% take 3 minutes to review.

The saving is 90 planner hours a month. At an assumed $55 loaded hourly cost, that is $4,950. Assume a running cost of $0.40 per request for model calls and hosting, or $480. Net: $4,470 a month, before build and upkeep.

The formula: hours saved = requests × (1 − low-confidence share) × (manual minutes − review minutes) ÷ 60. If review takes 5 minutes instead of 3, the saving drops to 1,200 × 0.75 × 4 ÷ 60 = 60 hours. If half the requests are low confidence, it is 1,200 × 0.5 × 6 ÷ 60 = 60 hours. In this example, those two inputs move the result by the same amount, so measure both in a shadow period before you promise a number.

Planner hours are the smaller prize. Faster triage on requests that become downtime is bigger but harder to price, so I’d make time from request to correct priority the hero metric.

Why is closed-loop control of equipment a different problem?

Closed-loop control means a controller measures a process variable, compares it with a setpoint, and adjusts an actuator to close the gap, over and over. In a plant those loops run in PLCs and DCS at Level 2, and each correction takes effect without anyone reviewing it first. Three things make putting an agent there a separate problem.

First, safety. In an IEC overview of IEC 61508, functional safety is the part of overall safety that depends on a system or equipment working correctly in response to its inputs. IEC 61508 sets the requirements for electrical, electronic, and programmable safety-related systems. The December 2025 joint guidance from CISA, Australia’s ACSC, and partner agencies in six other countries is blunt: “AI such as LLMs almost certainly should not be used to make safety decisions for OT environments” (CISA joint guidance, PDF). It also says “humans are responsible for functional safety.”

Second, security. The ISA/IEC 62443 series sets requirements for securing industrial automation and control systems, including security risk assessment for system design. Plants that apply it segment the control network, with controlled paths between segments. The CISA guidance says to “prefer push-based or brokered architectures that move required features or summaries out of OT without granting persistent inbound access.” Rockwell’s 2026 State of Smart Manufacturing report, with 1,560 respondents in 17 countries, found 46% had at least one cyber incident in the past year (Rockwell, May 2026).

Third, speed. Level 2 loops respond in sub-seconds. An agent calling tools takes seconds or longer.

Decision rule: the agent reads summaries that OT pushes out to a DMZ or broker, and it never holds credentials to a controller. If your plant cannot provide that push path yet, pick a pilot that needs no OT data. For network patterns, see private AI agents in your own cloud and the air-gapped oil and gas post.

Bottom Line

AI agents in manufacturing earn their keep in Level 3 and 4 work: ERP exceptions, quality records, maintenance triage, and RMAs.

Pick one row from the table with no OT exposure and a named reviewer. Run it in draft mode, measure how often people change the draft, and only then let it commit records.

Leave control loops to the control systems.

FAQ

Can an AI agent write to MES or ERP directly?

Yes, with the same transactions and approvals a human clerk has, and no more. Give it its own non-human identity. The CISA guidance asks for AI decisions to be logged under an identity distinct from any user or machine ID.

Do AI agents replace predictive maintenance models?

No. Predictive maintenance runs on analytical models over sensor data. An agent sits after the model and turns an alert into a work order with history and parts attached. Siemens’ Scale Package pairs the two: Senseye Predictive Maintenance with the Maintenance Copilot.

Does ISO 9001:2026 say anything about AI?

Not as a headline requirement. ISO’s summary lists leadership and quality culture (clause 5.1), strategy alignment (5.2), separate treatment of risks and opportunities (6.1), and new awareness requirements (7.3). The new Annex A clarifies key terms and the intent of requirements. Agent-drafted records are held to the same requirements as any other record.

How long does a first manufacturing agent pilot take?

ACG Capsules went from build to use by most of its operators and technicians in about five weeks, per McKinsey. That was a Lighthouse site with mature data. My read: without clean asset IDs in the CMMS or a document store that tracks revisions, the data cleanup will take longer than the agent build.

Which manufacturing AI surveys should I trust?

Check the population first. RSM’s 56% agentic figure covers 129 middle-market manufacturers. Rockwell’s 1,560 respondents span companies from $100 million to over $30 billion in revenue. Figures from different populations do not combine into one industry rate.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect