AI Agents on the Shop Floor: From Shift Logs to Scheduling
Every factory runs on two things at once: machines with their controllers and planning systems, and a steady stream of text around them. Shift logs, downtime notes, questions about job sequence, work instructions and emails to purchasing are written by hand every day. That second layer is where AI agents are useful. This guide explains where they help, where classic manufacturing systems stay in charge, and which security rules apply.
What are AI agents in manufacturing, and what are they used for?
An AI agent on the shop floor is software built on a large language model that reads information from production systems and turns it into summaries, explanations or suggestions. It works in the communication and coordination layer of a plant, not in machine control. It is good with messy, unstructured information and unsuited to real-time control.
Common tasks include:
- Shift handover: condensing the shift log, alarm history and production confirmations from the last shift into a short, structured handover.
- Explaining downtime: finding earlier incident reports, manual sections and fixes that match a current fault, and presenting them to the crew.
- Commenting on the schedule: explaining the plan from the scheduling system in plain language, flagging conflicts and drafting questions for sales or purchasing.
- Maintaining work instructions: drafting updates to operating and inspection instructions when a material, tool or test characteristic changes.
- Answering questions: "When did part 4711 last run on line 3, and what went wrong?" – answered with references, not memory.
The difference from traditional manufacturing IT is the kind of work. An agent reads a lot, combines sources and writes. It does not make binding decisions, and it stays out of processes that already have certified, deterministic systems behind them.
Where is the line between an agent and MES, APS and PLCs?
The line follows accountability. PLCs control machines deterministically, the manufacturing execution system (MES) records orders, confirmations and batches, and an advanced planning and scheduling (APS) system calculates capacity and sequence. An AI agent uses their data, explains it and prepares decisions. It replaces none of them.
The ANSI/ISA-95 model, published internationally as IEC 62264, is a useful map. Siemens summarises the levels as follows: level 0 is the physical process, level 1 senses and manipulates it through intelligent devices, level 2 monitors and controls it with systems such as PLCs and distributed control systems, level 3 coordinates production with systems such as an MES, and level 4 covers business activities with an ERP system.
| Level | Typical system | Role of the AI agent |
|---|---|---|
| 4 – Business | ERP | reads orders, dates, master data; drafts queries |
| 3 – Manufacturing operations | MES, APS | reads confirmations and plans; explains them |
| 2 – Control | PLC, distributed control system | at most filtered alarms via an intermediate layer |
| 0–1 – Process, devices | sensors, actuators | no access |
The rule of thumb: the closer you get to the physical process, the less a language model belongs there. On levels 3 and 4, a read-only agent can be very useful. Level 2 should only ever be reached indirectly, for example through a historian or alarm archive that already sits outside the control network.
Why draw the line so firmly? Language models are probabilistic. They give good suggestions but not guaranteed identical answers to identical inputs. That rules them out for control loops, interlocks and safety functions. It makes them a good fit for "What happened on line 2 during the night shift?", where clarity matters more than bit-for-bit repeatability.
How does a shop-floor agent work?
A manufacturing agent has four building blocks: a language model, read-only connections to data sources, a search index over documents, and a small set of tightly scoped output tools. All of it runs outside the control network, usually in the office or data-centre network, and reaches production data through defined interfaces.
1. The language model. The model understands questions, reads data and writes text. It can be a hosted model from a cloud provider or a model you run yourself. Plants with strict rules on data leaving the site often find a self-hosted model the shorter route to approval; the page on sovereign AI covers the background.
2. Read-only connections. The agent queries the MES, ERP and scheduling system through their interfaces, ideally through a read-only database view or an API with its own service account. That account should only read, and only see the tables the use case needs.
3. Document search. Manuals, work instructions, incident reports and shift logs are indexed so the agent can find and cite the right passage. Without this, a model answers from general knowledge – which on a shop floor is usually wrong, because every line has its own quirks.
4. Output tools. The agent writes only to clearly defined places: a draft entry in the shift log, a message in the line's Teams channel, a draft work request in the maintenance system. Everything it writes is marked as an AI draft and confirmed by a person. The guide to human-in-the-loop design for AI agents explains how to build those approval steps.
A typical handover run looks like this. At the end of the shift the agent pulls the alarms from the last eight hours, fetches production confirmations from the MES, reads the free-text shift log and produces a structured summary with open items. The shift lead reviews it, adds what is missing and releases it. The agent has decided nothing, but it has done most of the reading.
Which use cases exist on the shop floor?
Use cases can be ordered by how close they sit to the machines. The easiest starting points are read-only, text-heavy tasks such as shift handovers and downtime analysis. Further along are tasks where the agent proposes changes to schedules or processes, which then go through existing approval routes.
Shift log and handover. The most obvious entry point. Shift logs are often patchy, written in different styles and hard to search. An agent condenses them, enriches them with MES and alarm data and makes past shifts searchable. A dedicated guide on AI-assisted shift handovers is in preparation.
Downtime analysis and knowledge search. For a recurring fault, the agent finds similar past incidents, the actions taken and the relevant manual pages. This is closely related to general knowledge management with AI agents, with one difference: machine ID, time window and shift context filter the search.
Explaining and coordinating the schedule. The APS calculates the plan; the agent explains it. It summarises why a job slipped, which materials are missing and which questions need to go to purchasing or sales – and drafts those questions. It does not move jobs itself.
Work and inspection instructions. When a material, tool or inspection characteristic changes, the agent drafts a revised version and highlights what changed. Release still follows your document control procedure.
Where maintenance begins. Downtime analysis touches maintenance but is not the same thing. The guide to AI agents in maintenance covers maintenance plans, spare parts and work orders. A production agent looks at the running shift instead: what happened, what it means for the next jobs, and who needs to know.
Coordinating with neighbouring teams. Many production problems are really interface problems: missing material, unclear drawing revisions, rush requests. The agent collects these cases and drafts messages to the right team with all the data they need.
For industry-specific detail, such as engineer-to-order production and variant design, see AI agents for machinery manufacturers. That guide focuses on specifications, spare parts and field service; this one is about what happens inside the production hall.
What does an AI agent need before it can work in production?
A production agent needs three things: readable data with unambiguous keys such as machine and order numbers, a secured route to that data outside the control network, and an organisation that builds the agent into an existing routine. If one of these is missing, projects tend to fail for reasons that have nothing to do with AI.
Data. The agent must be able to match records. If the shift log says "Line 3", the MES says "L03" and the maintenance system says "Asset 3100", you need a mapping table. Free text can be messy – language models handle that well. Keys cannot.
Systems. You need read access to the MES, ERP, scheduling system and document store. For machine and process data, the right entry point is a historian or alarm archive outside the control network, never the controller itself.
People and process. Someone has to review the output, and that review has to fit the shift routine. A handover the agent writes but nobody reads is wasted. If the agent processes shift logs or confirmations that could reveal individual performance, involve your works council or employee representatives early, and check the local rules on employee consultation with your legal or HR team.
Security. Operational technology (OT) security is a precondition, not an add-on. It gets its own section below.
How do you roll out an agent step by step?
A five-step approach works well: scope one use case tightly, set up read-only data access, agree the security design with whoever owns OT security, run a pilot on one line or shift, then review and extend. Each step should have a clear result before the next one starts.
- Scope the use case. One line, one routine, one group of users – for example, handovers on line 2 for shift leads. Pick a metric you measure before and after, such as time spent on handover or follow-up questions in the next shift.
- Connect data read-only. Create a service account with read rights, document the sources and fields, and settle how asset IDs map across systems.
- Agree the security design. With the person responsible for OT security: where the agent runs, which connections it uses, who may do what. Write it down.
- Run the pilot. A few weeks of real use with feedback from shift leads. Collect failures: wrong matches, missing sources, unclear wording.
- Review and extend. Compare the metric, fold in the feedback, then move to the next line or use case.
Step three is often underestimated. Responsibilities between IT, production and maintenance frequently need to be clarified before any new system is allowed near the shop floor, so start that conversation early.
What does OT security mean for AI agents?
For AI agents, OT security mainly means: no direct path into the control network, read-only access at level 3 and above, traceable connections through defined crossing points, and a clear answer to what happens if the agent fails. Production must keep running without it.
Germany's Federal Office for Information Security (BSI) lists the main threats to industrial control systems in Industrial Control System Security – Top 10 Threats and Countermeasures 2022. They include compromise of extranet and cloud components, malware infection via internet and intranet, and intrusion through remote maintenance access. On cloud components, the paper notes that plant operators have only very limited control over their security, even though such components can be directly connected to local production. An AI agent with access to production data is exactly that kind of component.
For remote access, the BSI recommends, among other things, sufficiently fine-grained network segmentation, access points in a demilitarised zone (DMZ), firewalls that open only the required addresses and ports, and logging of every access. The same principles apply to agents: the agent pulls its data from a DMZ or from level-3 systems, never from the control network itself.
The IEC 62443 series provides the framework. According to the IEC overview, it is developed jointly by the IEC and ISA99, and part 3-2 covers dividing a system into zones – groups of assets based on risk – and conduits, the communication paths between zones. For an agent this means it gets its own zone, and every link to the MES or historian is a deliberately designed conduit with defined rules.
A checklist covering network, permissions and fail-safe behaviour is in preparation. Until then, three principles cover most of it:
- Read, don't write. The agent has no write access to the MES, scheduling system or controllers.
- Failure without consequences. If the agent goes down, nothing stops; the shift works as before.
- Traceability. Every query and every output is logged, so you can later see what a suggestion was based on.
What are the limits and risks?
The main limits are wrong or invented answers, poor key data, unclear ownership, and the temptation to grant write access once the first results look good. An agent is the wrong tool for anything that needs real-time behaviour, determinism or safety functions.
Wrong answers. Language models can produce plausible but false statements. The countermeasure: the agent cites the source of every statement – the shift log entry or alarm record – and says plainly when it found nothing.
Data quality. Inconsistent asset IDs and missing timestamps lead to wrong matches. These problems usually only surface during the pilot.
Scope creep. When the agent works well, someone will soon ask it to reschedule jobs or set parameters directly. Every expansion of that kind needs a fresh security and ownership review.
Not a fit for control loops, interlocks, safety functions, binding quality decisions or the scheduling calculation itself. An APS with explicit rules and constraints calculates more reliably than a language model; the agent is useful there as an explainer and communicator, not as the planner.
Acceptance. If workers feel watched, the project fails. Being open about which data is used, and agreeing this early with employee representatives, is part of the rollout, not an afterthought.
How do I get started with AI agents in production?
Start with a read-only use case that happens every day and that users can judge themselves – usually the shift handover or a summary of downtime reports. In parallel, agree with the person responsible for OT security where the agent may run and which data it may see.
A practical first step is a short inventory built around three questions:
- Which texts are written by hand on the shop floor every day, and who reads them?
- Which level-3 and level-4 systems could feed those texts, and do they offer read-only interfaces?
- Who decides whether a new system may be connected near production?
With those answers you can scope a pilot for a single line. The overview of AI agents for business functions explains how such agents are built in general and which departments make good starting points. On the shop floor one rule always holds: the agent works on the information around the machine, never on the machine itself.
Frequently asked questions
Will an AI agent replace our MES?
No. The MES remains the system of record for orders, production confirmations and batches. The agent uses that data to write summaries, explanations and suggestions.
Can an AI agent change machine parameters?
Not as a starting point. Writing to controllers is outside what a language model should do; parameter changes keep going through your existing engineering and approval process.
Does the agent have to run on site, or can it use a cloud model?
Either works, as long as the agent never sits directly on the control network. If process data is sensitive or plant policy is strict, a self-hosted model in your own data centre is usually easier to get approved.
What is a good first use case?
Usually the shift handover or a summary of downtime reports. Both are read-only, have a clear group of users, and shift leads can judge the quality straight away.
Sources
- BSI-CS 005: Industrial Control System Security – Top 10 Threats and Countermeasures 2022 (version 1.50, 3 May 2022) allianz-fuer-cybersicherheit.de
- IEC Systems Committee Smart Energy: IEC 62443 – overview of the standards series syc-se.iec.ch
- Siemens: ISA-95 framework and layers (ANSI/ISA-95 / IEC 62264 level model) siemens.com