Artificial intelligence can help teams investigate events, search technical guidance and review operating history. But the fact that a model can consume PLC data does not mean the first step should be a direct connection to the control system.

Before selecting a model, protocol or infrastructure, define the decision you want to improve and the role AI is allowed to play. Many early evaluations can run on historian exports, alarm records and documents without touching the live process. The following questions help establish that boundary.

1. What operating question should the solution answer?

“Apply AI to the plant” is not a testable objective. Start with a concrete question: What happened before a stop? Which procedure applies to this alarm? Which events should the next shift review?

Identify who will use the answer and what they will do with it. Maintenance, production and engineering may need different evidence from the same event. A precise question also makes it easier to decide whether AI adds value beyond a query, rule or clearer visualization.

2. Does AI need to observe, recommend or act?

These levels carry different risks. An observational tool can organize events or retrieve information. A recommendation tool may suggest checks while leaving the decision to authorized personnel. A system that acts could change setpoints, states or parameters.

State the permitted level explicitly. Read-only access is a strong default for an early evaluation because it isolates the assistance layer and makes results easier to review. Any write capability requires permissions, interlocks, rejection conditions and a safe recovery path designed with automation and operations owners.

3. Which data exists, and which context is missing?

An isolated tag rarely explains a machine state. Interpretation may depend on operating mode, recipe, cycle phase, adjacent equipment or a manual intervention.

Inventory process values, alarms, events, historian data, maintenance records, manuals, procedures and shift notes. For every source, identify its owner, format, location and access conditions.

The inventory will also expose gaps. A model cannot reliably reconstruct information that was never captured simply by inferring from indirect signals.

4. Is the time resolution sufficient to reconstruct what happened?

Sequence analysis needs more than the right values. Those values must also be ordered consistently.

Review sampling frequency, event-based logging, clock synchronization and transformations made by historians or intermediate platforms. Distinguish the time an event occurred from the time it was stored.

The required resolution depends on the case. A daily summary and the investigation of a fast transition operate on very different timescales.

5. What happens when data is missing?

Plants experience communication gaps, frozen values, failed sensors, naming changes and outdated documents. The solution should recognize these conditions rather than present an incomplete answer as conclusive.

Agree on rules for marking missing data, detecting stale inputs and communicating uncertainty. A responsible response should state the evidence it used, what is absent and which check still requires a person.

Also decide what happens when the AI service itself is unavailable. The control system should not accidentally depend on assistance that was never designed as a critical function.

6. Who owns the decision?

The interface should distinguish recorded facts, correlations, hypotheses and recommendations. Mixing these categories can give an inference more authority than it deserves.

Assign owners for reviewing results, authorizing access and responding to unexpected behavior. Operations, automation, maintenance and the solution provider may each have a different role. Documenting those roles avoids assumptions during commissioning.

Where required, the audit trail should include the inputs used, the applied configuration, the generated answer and the human validation.

7. How will the team decide whether the pilot worked?

Write acceptance criteria before reviewing results. Depending on the case, they may test whether the system retrieves relevant evidence, orders a sequence correctly or abstains when information is insufficient.

Use representative cases, including incomplete and ambiguous episodes. Compare the solution with the current process and record failures, not only favorable examples.

The pilot should allow three legitimate outcomes: proceed, adjust or stop. A convincing demonstration alone is not enough reason to integrate with production.

What to validate in your plant

Before designing a PLC connection, bring operations, automation, maintenance and security together to confirm:

  • The operating question and responsible user.
  • The boundary between observing, recommending and acting.
  • Available sources, their context and their owners.
  • Data quality and time resolution.
  • Behavior when inputs or services are unavailable.
  • Required permissions, records and human review.
  • Test cases and criteria to proceed, adjust or stop.

Architecture should follow these limits. Technology can be selected once the team knows what information it needs, which risks it must contain and what a useful answer looks like.