Skip to content
Hakan KaracaENGINEERING JOURNAL
AI & Intelligent AutomationEki 8, 20264 min read

LLMs in OT: Put Language Models Beside the Control System, Not in Its Fast Path

Large language models are excellent at language and weak at being treated as a control authority. In OT, that distinction is non-negotiable.

A language model can be genuinely useful in an industrial environment. It can help locate relevant documentation, explain unfamiliar alarm text, draft a handover note or turn an engineer’s question into a structured investigation plan. None of those capabilities requires putting the model inside a PLC scan cycle.

This boundary is not hostility to AI. It is a distinction between a probabilistic language interface and a deterministic control function. The control system has known timing, defined transitions, safety constraints and testable failure modes. A general-purpose language model may produce a plausible answer without being able to prove that answer is correct for this machine, recipe, software version or operating state.

Use the model for interpretation and navigation

Good candidate tasks are those where the output supports a qualified person rather than directly actuating equipment:

  • Search manuals, work instructions, electrical drawings and approved troubleshooting procedures.
  • Summarise an alarm history while preserving event timestamps and source references.
  • Generate a first draft of a maintenance report from technician-entered facts.
  • Compare a fault description with known failure modes and suggest diagnostic checks.
  • Explain a code block or help prepare a unit-test outline for engineer review.

The output should be treated as decision support. It should not silently become an interlock bypass, a setpoint change, a recipe approval or a quality disposition.

Draw the boundary in the architecture

A safer arrangement is to place the model in a supervisory application layer with explicitly limited interfaces. The application retrieves approved records, removes data the user is not authorised to see, and passes only the relevant context to the model. The response returns to the application, which attaches citations and displays uncertainty or missing evidence. Any proposed action passes through an existing, approved workflow with authentication, permission checks and a human decision.

PLC / SCADA / MES
      ↓ read-only, bounded data interface
Context and access-control service
      ↓ approved documents and selected events
LLM / retrieval service
      ↓ answer + supporting source IDs
Engineer review → authorised workflow (if appropriate)

The diagram is deliberately boring. In OT, boring is often a quality attribute.

Do not confuse fluent language with evidence

Language models predict plausible continuations. They can invent a menu entry, cite a document section that does not exist, or combine two nearly identical equipment variants. Industrial questions frequently depend on context: the installed firmware version, product contact status, the validated configuration, a particular alarm class or a site-specific standard operating procedure.

Require every consequential answer to show its evidence. For a document question, include document ID, revision, section and link. For a process-history summary, keep source timestamps and tag identifiers available. When the source material does not answer the question, the correct response is “the evidence is insufficient”, not a confident paragraph with excellent punctuation.

Security and availability considerations

Read-only access reduces risk but is not a complete security strategy. The application still needs least-privilege identities, network segmentation, audit logging, controlled updates, data-loss prevention and a plan for compromised credentials. Treat retrieved documents and operator prompts as untrusted input. A malicious or simply misleading instruction embedded in a maintenance PDF must not override the application’s permissions or system policy.

Decide what happens when the model service is offline, slow or unavailable. The underlying control system must continue to operate according to its designed behaviour without waiting for a generative service. Do not make an operator’s ability to stop equipment, handle an alarm or access an approved procedure depend on an internet connection or a model API.

Evaluate a use case like an engineered system

Start with a bounded dataset of realistic questions, including poorly phrased prompts, conflicting revisions, ambiguous alarms and cases where no answer exists. Measure factual correctness, source support, refusal behaviour, response time and the frequency of misleading omissions. Have domain experts review failures. Average accuracy is insufficient if the remaining errors cluster around high-consequence decisions.

Also measure operational value: time to find a controlled procedure, first-time resolution, repeat fault rate or the quality of maintenance handovers. A demonstration that impresses a visitor is not yet a use case that deserves an operational dependency.

Where the control boundary should remain firm

Keep safety functions, deterministic interlocks, time-critical control and validated control sequences in systems specifically engineered and verified for those tasks. An LLM may explain why a permissive is false by consulting approved documentation and read-only state. It should not invent a new permissive condition because the operator sounds impatient.

The practical objective is not “AI everywhere”. It is shorter distance between a real question and reliable evidence, while the control system retains authority over the physical process.

References and further reading

  1. NIST AI Risk Management Framework
  2. NIST SP 800-82 Rev. 3: Guide to Operational Technology Security