Manufacturing /Retrieval (RAG) live field guide · 9 min

Shop-Floor SOP Copilot

Answers shop-floor SOP, machine-setup and troubleshooting questions from your own approved manuals, work instructions and fault history — citing the source document and section on every step, and refusing when the documents don't cover it, so new starters stop queueing for the one veteran fitter.

Open the live demo

theater/demos/manufacturing_shop-floor-sop-copilot.html · sandbox · read-only

Open

FIG. 1

The live demo, running on fabricated data. Open it to step through the full flow — every output is shown for a person to approve before anything happens.

How it would work

Searches your own approved SOPs and work instructions, drafts a plain-language answer with the source document and section quoted, and refuses rather than guess — so the technician confirms it against the open SOP before acting.

Input 01

An operator's plain question

A shop-floor or new-starter question in plain language — die-clamp bolt torque, the lockout steps before clearing a jam, a guard-interlock reset rule — typed against your own indexed SOPs, work instructions, standards and quality manual.

Agent 02

Retrieves, quotes, scores

Searches the indexed corpus, gathers the verbatim source text for each matching section, drafts a grounded answer strictly from those chunks, and shows a grounded-confidence indicator naming how many of your documents it drew on.

Output 03

A cited answer, or an honest refusal

A cited answer — document, revision and section quoted — that the technician reads against the open SOP and approves before acting; or, when the corpus doesn't cover it, an amber "not in your approved documents" refusal that routes to the Maintenance Lead.

Where it works well

It pulls the controlled answer to the person at the machine in seconds, with the source shown.

Best where the answer genuinely exists in writing but is slow to reach the floor — a die-setup torque sequence, the lockout steps before clearing a jam, a guard-interlock reset rule.
Helps a multi-shift plant most, and the new starter or apprentice most of all — the people who currently can't act without interrupting someone senior.
The veteran fitter gets their day back for the diagnosis that genuinely needs judgement, instead of answering questions already written down.

The slow, invisible cost on a shop floor is the bottleneck of one veteran: the real answer to "what torque, what sequence, what's the reset rule" lives in a manual nobody can find, a maintenance log, or one fitter's head — so everyone queues for that person, and when they retire the fault-finding leaves with them.

Where it works badly

It is confidently wrong when it cites a superseded revision — and the citation makes a stale answer more convincing, not less.

Poor at anything that depends on what the technician can see, hear or feel — whether the tooling is seating, whether a noise is normal, whether this jam is the one the SOP describes. It retrieves text; it does not inspect plant.
If most of your fault-finding lives in one person's head and was never written down, there is little for it to retrieve yet — the valuable first job is capture, not a copilot.

The honest test

If there's any doubt about which copy of an SOP is current — if your work instructions aren't under real revision control — this amplifies that mess at the speed of search, and the document work has to come first.

If WI-12 has moved to v3 on the floor but only v2 is in the index, the copilot quotes v2's torque figure with the source shown and looks completely trustworthy. That is the trap.

What it doesn't do — and shouldn't

It surfaces and cites. A person approves before the setting is applied or the cell is run.

WHAT IT DOES

Quotes the source document, revision and section it drew the answer from

Shows a grounded-confidence indicator and how many of your documents it used

Refuses with an amber "not in your approved documents" notice when the corpus doesn't cover the question

WHAT IT WON’T

Decide that a machine is safe to run or that tooling is secure

Authorise resetting a guard interlock or returning a cell to service

Improvise a plausible figure when the document isn't there

In a plant this is a safety boundary, not a nicety. A wrong coolant pressure can wreck a spindle; a missed lockout step under the plant-isolation and guarding duties in the model WHS Regulations can hurt someone. The point where a person confirms against the open source document is where consequence concentrates — so that is where the human stays.

What your data has to look like

A defined, revision-controlled corpus — current SOPs and work instructions, the standards you work to, and the fault history that lives in one fitter's head.

Typical readiness

across orgs we see, before the first job

SOPs and work instructions as discrete, current revisions

Needs shaping

The standards you work to, as approved references

Usual weak point

Captured fault history

Needs shaping

Machine-readable document text

Usual weak point

Your quality manual under document control

Usually ready

The real first job

Most plants have some of this in good shape and a lot of it not — live and dead revisions mixed together, scanned PDFs that aren't searchable, the most valuable knowledge unwritten. Getting the corpus clean and under control is usually a matter of how documents are captured and revised, not buying a new tool — and it is typically bigger and more valuable than the AI layer on top.

Right fit if…

Your answers genuinely exist in writing but are slow to reach the person at the machine

A multi-shift plant with new starters or apprentices who can't act without interrupting someone senior

Your SOPs and work instructions are under real revision control, with superseded copies retired

Your documents are searchable text and each section can be identified to cite

Walk away if…

Most of your fault-finding lives in one fitter's head and was never written down

Live and dead revisions sit together with no clear "which copy is current"

Your manuals are scanned binder images that aren't searchable

You want it to sign off that a machine is safe to run or a guard is clear

Open questions

The worried-buyer questions, answered straight

It only answers from documents you have indexed and approved, and it shows the source document and the exact section the answer came from, so the technician can check it against the real SOP before acting. When a question falls outside the indexed corpus — say the commissioning manual for a new CNC cell hasn’t been loaded yet — it refuses with an amber “not in your approved documents” message rather than guessing, and routes the question to the Maintenance Lead. It accelerates the diagnosis; the fitter still makes the call and approves before anything is set or run.

Not until the corpus is sorted, and that is the honest first job. The copilot is only ever as good as the documents behind it, so superseded revisions, scanned binders that aren’t searchable, and tribal knowledge that was never written down all have to be dealt with before it can be trusted. We help you scope which SOPs, work instructions and standards go in, retire the old revisions, and capture the fault history that currently lives in one fitter’s head. That document work is usually bigger and more valuable than the AI layer itself.

It does the opposite of replacing them — it captures what they know so it doesn’t walk out the door when they retire, and it stops new starters queueing at their bench for answers already written down. The copilot surfaces and cites; the technician judges whether the situation actually matches the SOP, and the Maintenance Lead still owns anything the documents don’t cover. The capacity it frees goes back into the higher-judgement fault-finding only a person can do.

As current as your controlled-document process keeps it. The copilot answers from whatever revision is indexed, so if a work instruction has moved to v3 but only v2 is loaded, it will confidently cite the superseded torque figure. Re-indexing has to be tied to your document-control workflow — when an SOP, work instruction or a standard like the AS/NZS 4024 machine-safety series is updated, the new revision replaces the old in the index, not alongside it.

No. It runs as a private retrieval system over your own corpus — your SOPs, work instructions, standards and maintenance records stay in your tenancy (the demo pattern uses Microsoft Copilot inside your Microsoft 365 environment). Answers are generated only from documents you have loaded and approved; nothing is used to train a public model. We scope the data boundary with you up front so you know exactly what is indexed and where it sits.

What it takes to build

3–4 weeks · 4 phases

Reused from template~70%

Bespoke to this skin~30%

stack · Microsoft Copilot · private RAG · review UI

What it would cost

Fixed scope, fixed price, fixed dates.

Bite-sized first piece

One contained change, low risk

Pilot build

Most builds land here

Embedded support

Scale on proof

Considering this for your plant?

The honest place to start is the corpus — scoping which SOPs, work instructions and standards go in, and what's missing. Tell us where the floor queues for answers; we'll play it back, scope it, and show you what's possible.

Book a call How we work