An alarm tells me what happened. I want to understand why.

Monitoring Notes No. 01 — AI should help explain events, assess what is happening and anticipate what may come next. The final decision remains a human responsibility.

The next useful answer

A temperature reading gives me a value. An alarm tells me that a rule has been triggered. I still need to understand what happened, whether it is continuing and what I should examine next.

That is the help I want from AI: an explanation I can examine and use to improve, anticipate or correct something.

Start with a question the operator actually has

Consider a hypothetical cold room with an overnight temperature excursion. The operator has the temperature history, door events, defrost schedules, equipment status and the actions recorded by the team. Reading those sources separately takes time.

AI could assemble the timeline and identify possible explanations. A long door opening may coincide with the rise. Recovery may have slowed over several days. The records may also be insufficient to explain the event.

Those possibilities must remain distinguishable. A coincidence in time can suggest a cause without establishing it. The system should show the records behind its interpretation and make missing information visible.

Explain the past, assess the present, anticipate the next problem

I judge the proposed use against three questions: does it help explain the past, understand the present or anticipate a problem? Each requires appropriate methods and evidence.

A summary of yesterday’s event should preserve the facts and their sequence. An assessment of a current problem needs timely, relevant inputs. A prediction needs evidence that its signal performs usefully in the operating conditions where it will be used.

A convincing paragraph does not establish any of those capabilities. If the system says that equipment is deteriorating, I want to see how that conclusion was tested and what uncertainty remains. If it cannot distinguish an equipment problem from changed loading practices, that limitation matters.

Keep the original evidence within reach

The NIST AI Risk Management Framework organises risk-management activities around Govern, Map, Measure and Manage. Its trustworthiness discussion includes reliability, explainability and accountability. I find that useful for asking what an AI application is meant to do, how it will be evaluated and who remains responsible for its use.

I expect every explanation to remain connected to its evidence. The reviewer needs direct access to the measurements, events and configuration used in the analysis.

AI should never fill a data gap with a plausible story presented as a measurement. If the evidence is incomplete, the analysis should say so. That may be the most useful conclusion it can offer.

The product decision remains with people

I want AI to help the responsible team understand the exposure and gather the evidence needed for a disposition decision. The final decision to retain, release, reject or destroy a product should remain with the authorised human decision-maker under the relevant quality procedures.

That decision may depend on product-specific stability evidence, the full exposure history and information outside the monitoring platform. An explanation of the equipment event does not by itself establish whether the product remains suitable for use.

For corrective and preventive action, AI can help organise the investigation and prepare a working text. The responsible people must challenge the proposed cause, determine the actions, finalise the CAPA and approve and sign it through the applicable process.

That allocation of responsibility is the position I defend for these uses. It needs to be reflected in the organisation’s procedures.

Test the explanation before relying on it

Start the evaluation with historical events the team understands, protecting any sensitive information. Check the reconstruction against the records. Look closely for invented causes, ignored contradictions and unjustified certainty.

Then test incomplete records and unfamiliar situations. Does the application recognise when it lacks the evidence to answer? In GxP use, the intended purpose, validation and changes also belong within the applicable quality system.

Before I sign a decision informed by AI, I need to distinguish the recorded facts, the system’s interpretation and my own judgement. The explanation should make that review easier.