---
title: Building a knowledge agent your technicians actually trust
url: https://opero.pro/resources/trust
collection: resources
---
**Kind:** guide · **Read time:** 9 min · **Author:** Anna Lindholm · **Persona:** Service mgr · **Tag:** Operations · **Industry:** Cross-industry · **Published:** 2026-01-22
Technicians don't distrust AI because it's AI. They distrust it because the first three answers were almost right, and "almost right" on a 40-ton press is how people get hurt. Trust is operational — you build it in the corpus, the citations and the audit log, not in the marketing copy.

## Document-level citations, on purpose

Every answer the agent gives points back to the source PDF and page — not a highlighted line, not a paraphrase. That's a deliberate choice. When a technician opens the citation they see the answer in *context*: the warning above it, the torque spec below it, the revision date in the header. A line-level citation hides all of that and quietly trains the reviewer to skip the source.

Document-level surfacing also keeps the agent honest. If the model paraphrased something the page doesn't say, the technician sees the page and notices. That gap is the most valuable feedback signal in the whole system.

## Revision awareness is half the work

Service bulletins get superseded. Manuals get reissued. The most common trust failure we see in the wild isn't hallucination — it's a confident answer cited to a bulletin that was withdrawn eighteen months ago. Generic RAG pipelines don't catch this because they don't model "supersedes" as a relationship; they just index every PDF in the share.

Getting this right is unglamorous. Someone has to tag the corpus with effective dates, supersession chains, applicable model and serial ranges. There is no shortcut. But once it's tagged, the agent will refuse to cite a withdrawn document, and that single behaviour earns more trust than any UI polish.

## Role and document permissions

Sales does not need to see engineering drawings. Field technicians do not need to see commercial pricing. The same agent serves both, but the retrieval is filtered by the user's role before the model ever sees the candidate documents — not after. Filtering post-generation is how leaks happen.

Document-level ACLs sit alongside the revision tags in the same metadata layer, which is also where customer-confidential material gets walled off per tenant. Get the permissions schema right at week one; retrofitting it later is painful.

## Audit logs for every retrieval and every action

Every question, every retrieved chunk, every cited source, every outbound action — PO draft, ticket update, work-order note — is logged with the user, timestamp and model version. This matters for two reasons. One: when a technician disputes an answer six weeks later, you can replay the exact retrieval and see whether the corpus has drifted. Two: regulated customers will ask for this, and "we have logs" is a much shorter conversation than "let me check."

## The feedback widget is the eval set

Every answer carries a thumbs-up / thumbs-down with a one-line "why." Every thumbs-down becomes a row in the evaluation set, reviewed weekly, and the failure mode gets traced to either the corpus (missing or mis-tagged doc), the retrieval (wrong chunk surfaced) or the generation (model paraphrased badly). Each layer has a different fix.

The teams that get the most out of their agent are the ones who treat the eval set as a living document, not a launch artefact. Three months in, your eval set is more valuable than the model.

## Trust is operational, not a feature flag

There is no "trustworthy AI" toggle. There's a corpus that's tagged correctly, a citation pattern that puts the source in front of the reviewer, a permissions model that doesn't leak, an audit log that lets you replay any answer, and a feedback loop that closes. Do those five things and the technicians start opening the agent before they pick up the phone. Skip any one and they won't, no matter how good the model gets.