NewIntroducing the Vexum advisory boardRead the announcement
All posts · 4 min read

Can AI update work instructions from what happens on the floor?

Yes, if it starts from the gap between what's written and what actually happens, and a person approves every change. Here is what that looks like and where it goes wrong.

VexumVexum
A laptop showing a procedure beside an approval stamp
In short
  • AI can help keep work instructions current, but only if it can see what actually happens on the floor, not just the document.
  • Drafting a new instruction from a prompt is easy. Knowing which step is wrong, and why, requires the work to be recorded against each step.
  • When the real work keeps differing from a step, AI can propose the change with the evidence attached.
  • It should never publish on its own: a person approves every new version, and everyone trained on the old one is flagged.

Most of the talk about AI and work instructions is about writing them: turn a PDF into steps, turn a video into a procedure, draft an SOP from a prompt. That part is useful, and it is no longer hard. The harder question is what happens next, when the instruction has been written and the floor starts doing something else.

Work instructions go stale because the work moves and the document doesn't. The gap between what's written and what actually happens grows a little every week. An AI that only reads the document can't see that gap.

What does AI need to update a work instruction well?

The real work, recorded step by step. To know that step 4 is wrong, you need to know how step 4 is actually being done: the values entered, the photos taken, what operators said they did and the questions they asked. Without that, an AI can rewrite the instruction, but it is guessing at what to change.

Enough runs to see a pattern. One operator doing a step differently is a data point. The same difference across shifts and weeks, with good outcomes, is evidence that the instruction is behind the work.

The source documents. A good proposal cites where the current step came from, so the reviewer can see what is being replaced and why.

How does an AI-proposed change work in practice?

In Vexum, operators work each procedure with a voice agent on the station tablet or their phone. Every reading, photo, decision and question is recorded against the step and compared with what is written. When the answer doesn't match the step, the agent asks why and asks for proof.

When the real work keeps differing from a step, Vexum drafts the change and attaches the runs that show it. It also looks at the outcomes of past runs, so a change is proposed because the evidence supports it, not because a deviation happened once.

The proposal shows every procedure, location and training it touches, side by side with what it replaces. It follows your approval route, with every signature recorded. It can roll out to one line or location first. Old versions are kept, and everyone trained on them is flagged for retraining.

The AI proposes. A person decides. The procedure gets better with use instead of decaying.

What should an AI-proposed change contain?

A reviewer should be able to approve or reject a proposal in a few minutes without opening another system. That means it needs:

  • The step as written today, and where it came from in the source documents.
  • The proposed wording, side by side with the current version.
  • The runs behind it: how many times the work differed, on which shifts and lines, and what operators said when asked why.
  • The outcomes: whether the runs done the new way had fewer problems, or more.
  • Everything it touches: other procedures that share the step, the locations that run it and the training built on it.

For example: operators on two shifts keep checking a seal with a go/no-go gauge before the visual inspection, not after, because a failed gauge check makes the visual pointless. Over six weeks, every run done that way passed downstream inspection. The proposal moves the gauge check up one step, shows the runs, and flags the inspection training for an update. The quality engineer approves it in one sitting.

Where does AI-written work instruction go wrong?

When it publishes on its own. If the floor is doing a step wrong, an AI that learns from the floor will write the wrong way into the procedure. A deviation should trigger investigation, not become the standard. That decision belongs to a person.

When it has no plant data. General AI tools can draft a plausible instruction from a prompt. They have no idea how your line actually runs, and they leave no audit trail of where the content came from or who approved it.

When it can't show its sources. A reviewer who can't see why a change was proposed either rubber-stamps it or rejects it. Neither is review.

What should you look for?

  • Data from real runs, not only the document.
  • Proposals with evidence: the runs behind each change, and the source each step came from.
  • Proposes, never publishes: a person approves every new version.
  • Your approval route, with every signature recorded and old versions kept.
  • Gradual rollout, one line or location first.
  • Training assignments when procedures change, for everyone trained on the old version.

See how changes move from the floor to an approved version in management of change, or read a worked example of one change.

Frequently asked questions

Does the AI change procedures on its own?

No. Vexum proposes changes and a person approves them through your approval route. Nothing is published without a recorded approval.

How is this different from using ChatGPT to write SOPs?

General AI drafts a document from a prompt, with no plant data and no audit trail. Vexum proposes changes from what actually happened in your runs, cites the evidence, and records who approved the new version.

What if the floor is doing it wrong?

Then the procedure should not change. The evidence helps a person decide whether to retrain or to change the procedure.

Is our data used to train models?

No. Nothing a customer shares is used to train models.

If you run procedures at a manufacturer and want to see where your written procedures and the real work differ, request a demo.

Get new posts by email

Product news and notes on procedures, change and training.