AI-assisted work orders should preserve the evidence

Research perspective | Logic Foundry | October 9, 2026. This is an original design proposal and suggested evaluation, not a report of a completed experiment, validated diagnostic method or available AI work-order feature.

1. The operational question

Imagine a fictional office building where an occupant reports a warm meeting room. A sensor shows a higher temperature than nearby rooms, and an earlier maintenance record mentions a damper issue. An assistant could compress those details into a confident instruction to repair the damper. Yet those observations do not establish that the damper caused the current condition.

The question is how to make a work order easier to understand without making its evidence seem stronger than it is. A useful summary should preserve what the team knows, what it suspects and what still needs checking.

2. A proposed three-part record

Observation: retain the reported condition, source, relevant time and associated equipment or space. Distinguish a person reporting discomfort from a measured temperature. If the sensor location, unit or timestamp is uncertain, keep that uncertainty visible. Missing readings should remain missing.

Interpretation: label possible explanations as hypotheses. In the fictional example, a damper issue might be one possibility, alongside occupancy, schedule, sensor placement or another cause. An old maintenance note can inform an investigation without proving that the same fault has returned.

Proposed next step: state the information or review needed to decide what happens next. A qualified person should select appropriate checks under existing site procedures. The assistant should not turn a suspected cause into an authorised repair or override established control and safety responsibilities.

Keeping these parts distinct would let a reviewer challenge one element without discarding the whole record. An observation may be credible even when the first explanation is wrong.

3. Preserve context through revision

A proposed interface could let reviewers trace generated statements to source records and see when those sources were captured. If a technician corrects an equipment identity or rules out a suspected cause, preserve the earlier version and reason for correction. Otherwise, a later summary may recycle an obsolete interpretation as current evidence.

Across several buildings, property and asset associations matter as much as wording. A plausible note attached to the wrong air handler is still a poor basis for a decision. Design firms, controls providers and integration partners could help define which evidence needs to survive a handover into the maintenance record.

4. Where the proposal can fail

Source links do not guarantee source quality. Stale sensors, incorrect asset mappings, incomplete complaints and copied historical notes can remain misleading even when neatly cited. Reviewers also need to recognise contradictory evidence rather than assuming the longest or newest record is correct.

The format could create excessive administration if routine requests become lengthy reports. A short summary with expandable evidence is one interface hypothesis to evaluate, not a demonstrated improvement. Maintenance records may also contain occupant names, access details or unrelated personal information. Traceability should respect existing access permissions and should not expose every underlying note to every reader.

5. A small evaluation readers could design

Choose a small set of de-identified, completed work orders whose outcomes are already known to a reviewer. Using only an approved environment and permitted records, compare ordinary summaries with summaries divided into observations, interpretations and proposed next steps. Keep the exercise separate from live dispatch and building controls.

Ask reviewers to identify unsupported diagnoses, lost qualifications, incorrect asset references and missing context. Record corrections and review time. Then check whether a different reviewer can recognise what remains uncertain. Do not infer operational savings or diagnostic accuracy from a small convenience sample.

This would evaluate information quality. It is not a field trial performed by Logic Foundry. Even promising results would leave questions about different building types, incomplete histories, unusual faults and the effort required to sustain the process.

6. Two discussion prompts

Which statements in your work orders are direct observations, and which are interpretations that have gradually become accepted as facts?

What evidence would a technician need to disagree confidently with an AI-generated suggestion, and where should that disagreement be recorded?

Logic Foundry is researching AI-assisted work orders alongside predictive maintenance, sensor anomaly detection and future decision tools. CORE-FM currently supports properties, assets, work orders, preventive maintenance and sensor monitoring. This proposal describes a research direction, without claiming a working AI diagnostic integration or promising a release.