Agentic AI in Manufacturing and What It Can Do

• Agentic AI in manufacturing works best with written limits: agents act alone on low-impact, reversible actions, ask first on rescheduling, and never stop a line.
• A Manufacturing Leadership Council survey in April 2026 found 66% of manufacturers use or plan to use agentic AI, yet half have no AI metrics.
• Guidance from the US Cybersecurity and Infrastructure Security Agency (CISA) in December 2025 says chat-assistant AI almost certainly should not make safety decisions.

What Is Agentic AI In Manufacturing?

Agentic AI in manufacturing is software that plans its own steps toward a goal and carries them out in plant systems. It can order a spare part or raise a work order without waiting for a person to enter the transaction. A generative AI tool drafts a work order and leaves a person to enter it. An AI agent connected to the maintenance system can enter it.

In an April 2026 Manufacturing Leadership Council survey, 66% of manufacturers said they use or plan to use agentic AI tools. Most rated the maturity of their AI use, agentic or not, below 5 on a 10-point scale.

If you run a plant, you answer for safety, quality, delivery, and cost, and that includes what an agent does on your lines. Where earlier AI use cases at your plant never showed a clear return, start the next one with a written permission list.

When An Agent May Act And When It Should Only Recommend

An agent may act alone when an action has low impact and a person can reverse it. High-impact actions, and any action that is hard to undo, need approval first, and safety decisions stay with people.

Cybersecurity agencies from the United States, United Kingdom, Canada, Australia, and New Zealand published joint guidance on agentic AI in May 2026. It covers agents built on chat-assistant technology (large language models) in IT environments. In Accedia's view, the same approval logic applies where an agent touches plant systems. System designers or operators decide when an agent needs human approval, the guidance says, and high-impact actions need it first. In the same survey, half of respondents had no specific metrics for AI results, so a pilot's return is hard to show.

Three Questions That Sort Every Plant Action

The guidance also tells organizations to classify agent actions by impact, likelihood, and reversibility.

• What damage can a wrong action do?
• How likely is the agent to get it wrong?
• Can the action be undone?

Answer all three for every action before an agent gets permission to take it. Guidance from the US Cybersecurity and Infrastructure Security Agency (CISA) and its partners, published in December 2025, covers AI in operational technology (OT). The guidance says AI, such as large language models, almost certainly should not make safety decisions, and that humans stay responsible for functional safety. It also calls for failsafes that revert to traditional automation or manual operation. The guidance is written for critical infrastructure operators, and Accedia reads its reasoning as applying to any plant where software can affect physical equipment.

Which Plant Actions An Agent Can Take Alone

An AI agent can reorder spare parts under a value cap and raise work orders on its own, except on safety-critical assets. Rescheduling a job needs approval, and an agent never stops a line alone. No standard prescribes these limits, so each plant should adjust the table below for its own equipment and suppliers. If a corporate system migration is breaking your manufacturing execution system (MES) interfaces, fix that before an agent acts on any row of the table:

• Work orders. At a Swedish site, the safety lead and maintenance manager agreed the asset list in one meeting, and technicians kept everything on the safety-critical list.
• Rescheduling. Deloitte's road map describes an agent that drafts a revised production schedule after a part shortage or machine failure. The agent then asks a production planner to approve it.
• Stopping a line and the EU. In the EU, most provisions of Regulation (EU) 2023/1230 on machinery apply from January 20, 2027. It adds safety functions built on self-evolving machine learning to its high-risk machinery list. The regulation governs machinery and safety components sold in the EU, so it does not bind a plant's software agent directly. Accedia's reading is that it gives plants in Sweden and the rest of the EU another reason to keep stop functions in dedicated safety systems.

Before An Agent Can Change Anything In Your Systems

Before an agent can change anything in your systems, it needs a read-only trial, its own login and log, a fallback threshold, and tested connections.

Start With Read-Only Access

The operational technology guidance from CISA and its partners describes a predictive maintenance system with read-only access to OT data that produces recommendations. An agent can run in that mode first. Once its recommendations match what the plant sees, it can start making changes with low-impact, reversible actions.

In a client project, the first action Accedia gave an agent was writing the shift handover report. It could read the maintenance and MES data and write the report, but it could not change anything on the line. Supervisors edited the first reports, and we widened its role once they stopped editing. An early agent raised work orders alone for conveyor and packaging equipment, with a cap of ten open orders per shift. Every order went to the maintenance planner, who could cancel it. After eight weeks with no canceled orders, the agent's scope widened to more assets.

Give Each Agent Its Own Login And Activity Log

Each agent needs its own login, separate from any person or machine account, so every action traces to it. Record each action in an activity log with that login and the name of the person who approved it. An early reordering agent bought bearings and filters from two approved suppliers, with a $500 cap on each order. Anything above that waited for the purchasing lead. The agent logged every order under its own login, so finance reviewed them each Friday.

Set A Threshold For Handing Control Back

Accedia recommends deciding in advance when the agent stops, and a person or the previous system takes over. If its outputs fall below a performance or safety threshold, the plant reverts to the earlier process or to manual operation. Name the person who can trigger that handover on every shift.

Check Integration And Data Readiness

An agent needs a working connection to every system it changes, and current data to act on. A corporate system migration that changes those interfaces means you need to retest the agent's permissions. If your plant runs older equipment, pulling live data and sending changes back are the hard parts. Many manufacturers lack in-house engineers for this work and turn to software development companies that already build for plants.

Conclusion: Who Should Sign Off On Your Agent Permissions

Agentic AI in manufacturing works when the people who answer for each action sign off on its limits. An agent’s authority should match the cost of a wrong decision, and those people bear that cost. Before any pilot, take the table to the people who would answer for each row. Ask each of them what the agent would do after a wrong decision and who would see it first. Write their answers down, and leadership can check the pilot's results against a list of what the agent was allowed to do.

Reach out to Accedia’s manufacturing team to review your plant’s permission list with the plant manager and IT director, from system interfaces to approval steps. The team works with manufacturers in the US and the Nordics on legacy systems and AI.