Lessons learned: five AI errors in work instructions that reviewers miss on a quick read

So we did something tedious. We took drafts generated from real field-service recordings — cleaning routines, a calibration, a motor test, a belt replacement, recorded by technicians in French and English — and compared them line by line with what was said and shown in the video. Every step, every value. The recordings come from a pilot with the field-service team of an international recycling group in Canada, in July and August 2026. The recordings and the review are described in more detail in where a video-to-SOP draft loses detail.

The errors fall into five patterns. None of them are spectacular. All of them matter, because they are exactly the kind a busy reviewer will not catch on a quick read.

A technician sets a tester to a specific value and says the number while doing it. The draft says “adjust the sensitivity and start the test”. The sentence is correct. The step is useless, because the number was the point.

This is the error that hurts trust most. Once a reviewer has found one dropped value, they stop trusting every step and start re-watching every video. At that point, the time saving is gone.

Not every number in a recording belongs in the SOP: some are guesses, some are corrected a minute later, some are small talk. But a value that is stated as the setting, limit or target for a step does belong there. If the draft leaves it out, the reviewer should be told, not left to notice.

Rule: every number and unit in the recording is either carried into the draft or listed for the reviewer as “stated, not used”, with its timestamp. That comparison is mechanical and should not depend on anyone's attention.

Reviewer check: before reading the steps, scan the transcript for numbers. Decide for each one whether it is relevant. Then find the relevant ones in the draft.

The opposite error. A technician adjusting a belt says: “I'm not sure, maybe two millimeters, or one, or one and a half. Just one turn.” The draft says: “Give the adjustment screw exactly one full turn to set the final tension.”

Two things went wrong. The hesitation disappeared, and “exactly” appeared. The technician was aligning the belt, not tensioning it, but the draft turned a rough guess into a specification.

A related version: the draft states “insert the plug into a 110/220 V outlet”, where nobody mentioned a voltage. The model filled in something that sounded plausible. In an electrical procedure, plausible is not good enough.

Rule: a value in the draft needs a source — the recording, an approved specification, or the reviewer. Uncertain speech stays uncertain: “approximately one turn; confirm against the specification”.

Reviewer check: search for “exactly”, “always”, “never” and every number. Ask where each one came from.

People correct themselves while they work. In one recording the technician first says a motor is turning in the right direction, and a few minutes later, with a colleague, concludes that it is not, and explains why. The draft kept the first statement and built the quality check around it — the pass/fail criterion of the whole test, reversed.

A model asked to produce a clean document will tend to resolve contradictions, because clean documents do not contain them. But the reviewer needs to know that the recording was ambiguous, especially when the ambiguity concerns the acceptance criterion.

Rule: conflicting statements about the same thing are flagged in the draft, both of them, with timestamps. The reviewer decides.

Reviewer check: read the quality-control criteria first and compare them with the last few minutes of the video. Corrections tend to come late.

When two people are on camera, one usually asks questions. “Is this only for the disc, or also for the belt?” “How often do you clean this?” In a draft, a question can turn into a step, and the answer to a hypothetical (“if it were turning forward, you would swap the wires”) can read like an instruction for something that did not happen.

The same applies to recordings that are discussions rather than demonstrations. We saw a recording in which the team talked through error messages for a minute without touching the equipment. The draft produced action steps anyway, with confident reasons attached.

Rule: a step generated from a recording needs an action in that recording. A conversation without action produces notes, not steps. Steps added from other sources — a specification, the reviewer's own knowledge — are fine, but they should be marked as such, not presented as if they were filmed.

Reviewer check: for each step, find the moment in the video where the action happens. If there is none, either find its source or remove it.

This one has nothing to do with speech. Risk assessment fields in a work instruction often start with a default value. If a reviewer fills in the hazard (“electric shock”) and the measure but never touches the risk rating, the document can end up printing the default — the lowest rating, in green — next to an electrical hazard.

Nobody assessed that risk. The document says somebody did.

Rule: an unassessed field renders as unassessed, never as a value. Absence must look like absence.

Reviewer check: in the safety section, check every rating you did not set yourself.

Each of these errors makes the document look more finished than the recording was. A missing value leaves a clean sentence. Invented precision reads more authoritative than the hesitation it replaced. A resolved contradiction reads cleaner than an open one. A question turned into a step fills a gap. A green default looks like a decision.

That is the real risk of AI-generated work instructions: not that they look wrong, but that they look right. A reviewer reads a fluent, well-structured draft, nothing jumps out, and approves it.

The fix is partly in the tool and partly in the review. The tool should do the mechanical checks — values in versus values out, unassessed fields, conflicting statements — and show the results instead of hiding them. The reviewer can then spend their attention where no tool helps: whether these are the right steps, and whether the emphasis is on the parts of the job that actually matter.

This is not a full technical approval. It is the first pass we recommend before the subject-matter review, because it finds the five errors above quickly:

1. Scan the transcript for numbers and units. Mark the relevant ones and find each in the draft.

2. Search the draft for “exactly”, “always”, “never” and every number. Find its source.

3. Compare the quality-control criteria with the end of the video.

4. For each step, confirm there is an action in the video or a named source.

5. In the safety section, check every rating you did not set yourself.

If this pass finds several problems, look at the recording before correcting the draft. Silent stretches and conversations on camera are the usual causes, and it is often faster to re-record two minutes than to rewrite ten steps. Every AI-generated work instruction still needs approval by a qualified person before it is used.