A Responsible AI-Assisted Operating Model for Manufacturing

01Observe the real work before the paperwork

Compare standard procedures with actual setup, inspection, and exception handling.

Capture conditions, signals, decisions, actions, and outcomes rather than producing a simplistic answer book.

Observation is required because the shop floor itself is thinning. According to Japan's 2026 White Paper on Manufacturing Industries (METI, MHLW and MEXT, approved by Cabinet on 29 May 2026), the share of manufacturers using digital technology rose from just under 50% in 2019 to over 80% in 2024. Over roughly the same two decades, employment in manufacturing fell by 1.57 million. Holding on to the judgement criteria that live in experienced workers' heads is getting harder as an organisational matter.

Start by filling in the observation sheet below, one event per row. The last column, whether the step appears in the standard procedure, turns the gap between documents and practice into a list. Processes with many "no" rows are the first candidates for knowledge cards.

If you are unsure where to start, run a process inventory first. Good candidates are processes that generate many questions to experienced staff, or changeovers where the steps vary from day to day.

02Make tacit knowledge verifiable

Structure knowledge by equipment, product, process, symptom, possible cause, and verification method.

Attach source, date, and approver. When evidence is insufficient, the assistant should escalate rather than improvise.

Use the knowledge card below as the container. Keep one symptom per card. Both "first response" and "escalation trigger" must be filled in, and a card with either field empty is not approved. If the card does not say how far the operator may act alone, the assistant cannot say it either.

The approver should be the role accountable for quality or equipment in that process. The experienced operator supplies the content but is not the approver. A next review date stops outdated knowledge from staying in use after a material or machine change.

03Worked example: first-article check after a mould change

The following is a fictional example. It shows the first response to a defective first article after a mould change in injection moulding, written up as a knowledge card. Machine names, part numbers and dates are illustrative only and do not come from any real plant.

Note two things. The possible causes are ranked, and the operator's scope stops at waiting and taking another sample. The card does not tell the operator to change process conditions, because that affects quality and belongs to the team lead beyond the escalation trigger.

The source field points back to an observation record and an interview date. Only cards that trace back to an observation record go into the assistant's reference set. A card based on interview alone is approved after the event has been observed the next time it happens.

04Separate assistance from quality decisions

Search, checklists, and record formatting are suitable early support cases.

Product disposition, equipment shutdown, and safety decisions remain with named owners and explicit standards.

Draw the line with the decision table below. "What AI may do" stops at gathering and arranging inputs. Rows for disposition and stopping say "not delegated", and the assistant may only point to the criteria and the owner. Write owners as roles, not personal names.

Back-office functions need the same line. For the internal-control view, see control checkpoints before using AI in accounting.

Once the table is agreed, write the same boundaries into the assistant's instructions. The prompt below uses approved cards as the only source and escalates instead of answering when evidence is missing. Recording why a suggestion was not used reveals operational problems that model accuracy alone does not show.

You are a work-support assistant for a manufacturing floor. Follow these rules.

# Allowed information
- Base every answer only on the "knowledge cards" and "work records" pasted below.
- Do not fill gaps with general knowledge or guesses.

# How to answer
1. State the number of the knowledge card that matches the symptom the operator entered.
2. List that card's "how to check" and "first response" steps in order.
3. After each step, add the card or record number you relied on in brackets.
4. If an "escalation trigger" applies, start with "Contact the owner now" and name the role to contact.

# Do not answer
- Do not answer product disposition, stopping or restarting equipment, or safety decisions. Escalate to the role listed as "final decision owner".
- If no card matches, reply only: "No matching knowledge. Please check with the owner."

# Knowledge cards
(paste here)

# Operator input
(describe the symptom here)

05Run a controlled parallel pilot

Compare the assisted process with the current process in one bounded operation.

Measure response time, recurrence, record completeness, and expert interruptions while protecting safety and quality thresholds.

Fix the measures before the pilot and record the baseline the same way. Choosing measures afterwards leaves only the flattering ones. During the pilot, log in one line whether each suggestion was used and, if not, why.

Decide on expansion by looking at whether escalation happened when it should have, not only at the measures. If the assistant answered where it should have escalated, fix the cards and the prompt before extending to similar equipment or adjacent processes.

06Create reusable plant-level building blocks

Separate shared procedures from site-specific exceptions.

In card terms, keep symptom, possible causes and checks in a shared card, and put condition values and contact roles in a site-specific appendix. When a shared card changes, you can then see which sites' appendices need checking.

Version knowledge, workflows, and permissions together so teams can see the impact of every change. For the conditions that let the floor lead this work, see operator-led AX.