Worked Industrial AI acceptance pack: fictional unplanned-stop triage on one packaging line

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

A fictional mid-sized plant that packs consumer goods. One cartoning line runs two shifts, five days a week. The workflow is the triage of unplanned stops longer than five minutes: from the moment the line stops to the moment a technician has confirmed the cause.

• Maintenance manager. Operating owner. Sets acceptance with the technicians, takes the decision and runs whatever is kept.
• Plant manager. Sponsor. Agrees the value of a downtime hour and the two economic thresholds before the trial.
• Line maintenance technicians. Users. Take part in the trial, record each stop and say when a suggestion was wrong.
• IT and operational-technology security lead. Approves what the assistant may read, where it runs and who can open it.
• Safety officer and quality manager. Keep their authority. Either can suspend the trial without giving a business reason.
• Integrator. Installs and connects the candidate under the plant's purchase order.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

Invented measurements over 8 weeks, with the way each would be collected in a real plant.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

Fix the process

Standard cause codes on the line terminal and a one-page fault-finding card for each of the five most frequent causes, written by the maintenance engineer with the technicians.

The five causes cover 62.98% of stops; the other 77 stops in the window have no dedicated card. This is a coverage limit, not by itself a reason to buy AI.

Use a tool the plant already has

The text search of the maintenance system the plant already licenses, opened on a tablet at the line, over past work orders.

This search reaches less frequent causes, but past work orders are written inconsistently and it does not search the machine suppliers' manuals.

Both together, measured

The two were run together for 4 weeks before any new technology was selected. It cost the plant 10 staff days (4,000 EUR at the internal day cost).

Mean diagnosis time fell from 14 to 10.5 minutes over 104 stops, and first fixes that held rose to 73.08%. On the annual stop count that is worth 62,790 EUR a year. This is the result the candidate has to beat, not the original baseline.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

A triage assistant on the line terminal. The technician enters the stop code and a short description; the assistant searches the suppliers' manuals, past work orders and the shift log, and returns up to three likely causes, each with the page or record it comes from and the plant's own check procedure for that cause.

• It reads documents. It is not connected to the line controller and cannot start, stop or change anything.
• A suggestion is never a command. The technician decides what to check and what to do.
• When it finds no supporting document it says so and suggests nothing.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

The thresholds were agreed with the operating owner before the trial. The observed values are invented.

In this fictional trial, each of the 71 consultations produces one response: 66 contain up to three proposed causes and 5 abstain. The top-three rate is counted per non-abstaining response, and 49 of them contain the confirmed cause. The citation check and the review for contradictions with procedures cover every proposed cause in every non-abstaining response.

The automated citation check establishes that references open. It does not establish source relevance, currency, support for each claim or correct abstention; those require additional evaluation before operational use.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

At every stop

1. Make the machine safe with the plant's existing procedure. The assistant plays no part in this step and is not opened until it is done.
2. Enter the stop code and a one-line description on the line terminal.
3. Read the suggestions and open the cited source for the one you intend to act on.
4. Decide. You may ignore every suggestion.
5. Carry out the check and the repair under the plant's normal rules.
6. Close the stop record: the confirmed cause, whether the first action held, and whether the suggestion was right, wrong or not used.

When the assistant is wrong

1. Mark the suggestion wrong on the stop record. This takes one tap and is never held against the technician.
2. Carry on with the fault-finding card or the maintenance system search, as before.
3. Call the maintenance engineer under the existing escalation rule. The assistant does not change who is called or when.
4. The maintenance engineer reviews every suggestion marked wrong each week, distinguishing source errors from retrieval, interpretation and configuration errors. The document owner approves any source correction; valid source documents are retained. The integrator corrects assistant defects, and affected examples are retested before the change is accepted. Safety contradictions trigger immediate suspension under the stop conditions.

When the assistant is unavailable

1. Do not wait for it. Use the fault-finding cards and the maintenance system search, which stay at the line for this reason.
2. Tick the unavailable box on the stop record so the minutes are counted.
3. IT restores the service outside the stop. No stop is extended to bring the assistant back.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

Data

• Sources are the suppliers' manuals the plant is licensed to hold, the plant's work orders and the shift log. Nothing else is indexed.
• The index and the query log stay inside the plant network. Whether any text may leave it is decided by the IT and operational-technology security lead, in writing, before the trial.
• Shift logs and work orders contain technicians' names. The plant's data-protection contact decides what is removed or masked before indexing and how long query logs are kept, and the technicians are told what is recorded.

Access

• Read-only accounts. The assistant cannot write to the maintenance system or the shift log.
• No connection to the line controller or to any safety function.
• Only the line's technicians, the maintenance engineer and the maintenance manager can open it.

Safety and quality

• Lockout and the plant's safety procedures are unchanged and come first.
• The safety officer reads every suggestion issued in the trial and can suspend it at once.
• Quality release decisions stay with the quality manager. The assistant gives no opinion on product release.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

These are costs the plant itself carries. No consulting fee and no Hyperion price is part of this pack: a fee for leading such work is quoted separately in writing and would be added by the plant as its own line.

The payback of 3.46 years in the measures above is a total-project calculation. At the post-trial decision, distinguish expenditure already incurred from avoidable future costs; do not treat spent trial effort as a new purchase cost. The comparator's 62,790 EUR is gross modelled annual benefit before any ongoing upkeep. A real comparison must identify recurring costs for both options and separate shared costs from incremental assistant costs.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

• One suggestion that contradicts a safety or lockout procedure: the trial is suspended the same shift and does not restart without the safety officer's written agreement.
• Text from the plant's documents or logs found outside the boundary the security lead approved: the assistant is switched off and the index removed.
• First fixes that hold fall below the comparator for two weeks running: the trial stops and the line returns to the comparator procedure.
• Availability below 98% of scheduled minutes over the trial: no decision to go is possible, whatever the other measures say.
• The comparator and the candidate both fail to improve on the baseline: the work stops and the plant keeps its original procedure.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

The rule, agreed before the trial:

• Go: every acceptance and economic measure is met. The assistant goes into daily use on this line with the procedure above.
• Fix the process: the comparator beat the baseline and the assistant did not clear every measure. The plant keeps the cause codes, the cards and the search, and does not buy the assistant now.
• Stop: a safety contradiction occurred, or the assistant did not clear every measure and the comparator did not beat the baseline either.

Outcome in this example: Fix the process.

In this fictional example, all eight listed acceptance measures meet their thresholds: mean diagnosis time is 1.6 minutes shorter than with the comparator, technicians consulted the assistant at 71 of 100 stops, and the review recorded no contradiction with the plant's safety or lockout procedures. This does not establish that the assistant or its suggestions are safe. A safety assessment by the plant's safety officer would still have to establish that acting on a proposed cause cannot lead a technician to skip or shorten isolation and lockout, that a wrong or outdated proposal is caught before anyone acts on it, and that the procedures used for the review are complete and current. The two economic measures are not met. Beyond the process fix it is worth 28,704 EUR a year against a running cost of 21,000 EUR, a ratio of 1.37 where 1.5 was agreed, and the one-off cost of 26,650 EUR would take 3.46 years to repay where 2 was agreed. The process fix alone is worth 62,790 EUR a year for 4,000 EUR of staff time. The maintenance manager's decision is to fix the process: keep the comparator procedure on this line and not put the assistant into use now.

The following changes trigger a fresh review, not approval. The first two address only the running-cost ratio; the payback gate of at most 2 years must also pass. With all stated costs unchanged, both economic gates require at least 34,325 EUR of annual value beyond the comparator: a gain of at least 1.92 minutes per stop at 1,196 stops a year, or at least 1,431 stops a year at a gain of 1.6 minutes. Expansion costs must be included before applying these thresholds.

• The gain over the comparator reaches 1.76 minutes a stop at the present stop count, approximately the threshold for the value to reach 1.5 times the running cost.
• The same licence and upkeep serve at least 1,313 comparable stops a year at the observed gain, for example because further lines with the same machines are added. That would have to be measured on those lines, not assumed.
• The running cost changes, or the plant's finance controller revises the value of a downtime hour.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

Every derived figure above, with its formula. Results are rounded to two decimals.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

• Nothing here was measured. The plant, the line, the stops, the times, the costs and the trial results are invented to show the shape of an acceptance record.
• The three periods run one after the other, so the stop mix differs between them. A real comparison has to show that the periods are comparable, or alternate the procedures by shift.
• The comparison rests on means over about 100 stops. A real review needs the spread of the times, not only the mean, before treating a difference of 1.6 minutes as real.
• A minute of diagnosis saved is counted as a minute of line running regained. The calculation assumes diagnosis is on the critical path to restart and that earlier diagnosis brings restart forward by the same amount, without increasing repair time, repeat faults or quality losses. The plant must verify that relationship using total stop-to-restart time and establish whether the recovered capacity creates the assumed contribution.
• The illustrative first-fix measure uses the remainder of the shift as follow-up, so exposure differs between stops. A real trial should agree a common follow-up period in operating hours, distinguish recurrence of the same fault from unrelated faults, and identify incomplete follow-up before comparing first-fix rates.
• The candidate combines an assistant with document clean-up and broader source access. The example does not isolate the contribution of AI; a real comparison should also consider making the same cleaned manuals and records searchable without an assistant.
• The annual figures apply the baseline stop rate to a full year. They are not a forecast of stops, and they say nothing about a different line, plant or product.
• The thresholds are the fictional plant's choices. A real operating owner sets different ones for a different workflow and risk.
• This pack gives no legal, data-protection, safety or conformity opinion. Those decisions stay with the people named for them.

Fictional worked example: invented inputs and calculated illustrations; no customer evidence.

The Markdown file carries the text and figures of the worked pack. The JSON file holds the same inputs, calculated figures and checks as structured data.

• Markdown file
• JSON file

Payload SHA-256: f63794bb10931e8191d85e3ffc1ceddd0c6c835bb0fc56a1b35e3720d1f14d5e