A European industrial-equipment manufacturer

The situation

The manufacturer’s installed base operates in environments where a wrong repair step is a safety matter, not an inconvenience: presses, compressors, and drive systems running on customers’ factory floors. Field-service engineers had begun using a maintenance copilot that drew on the service knowledge base to suggest diagnostic sequences and repair procedures against reported fault codes.

Three concerns stalled wider rollout. The copilot would occasionally assemble plausible advice that appeared in no approved procedure. There was no reliable record of what it had actually told an engineer on a given work order, which left safety reviews reconstructing events from recollection. And the knowledge base and underlying model both changed regularly, with no way to establish whether an update had altered the advice given for a known fault.

What changed

The team deployed the copilot inside a Hyperpriors harness scoped to documented procedure. Every recommendation must resolve to a versioned procedure in the approved manual set, and each one carries its citation. Where a reported fault falls outside documented coverage, or confidence is low, the harness does not permit improvisation: the case is escalated to the engineering support desk with full fault context, and the field engineer is told plainly that no approved procedure applies. Tool access and retries are bounded, so a failed lookup ends in escalation rather than a confident guess.

Every interaction is traced against its work order: the fault codes reported, the procedure versions consulted, the exact advice returned, and the identity of the reviewing engineer where escalation occurred. Traces are replayable and feed directly into the company’s safety-audit process.

Evaluation became a gate. Senior service engineers curated a golden set of known faults with agreed diagnoses and procedures. Every knowledge-base revision and every model update must pass it before release, with judging criteria calibrated against engineering sign-off; a regression on the golden set blocks the update.

Where it landed

The decisive shift was in accountability. Advice now either stays inside documented procedure or arrives at the engineering desk as an explicit escalation — there is no third category. Safety audits work from complete, replayable work-order traces rather than reconstruction, and questions about what an engineer was told on a specific date are answered from the record. Updates no longer change fault advice silently: when the golden set catches a regression, it is corrected before any field engineer sees it. Escalation volume itself became useful management information, indicating where documentation was thin and which fault families needed new approved procedures.

This is a hypothetical scenario, illustrating how strict auditing, harness control, and valid evals apply in industrial field service.