Electrical wholesaler supplying contractors from one warehouse, next-day delivery by own vans. 46 staff. Team: Warehouse manager, two pickers, the packing lead, a customer service adviser.
Wrong items delivered to trade customers: a different product, or the right product in the wrong quantity, reaching the customer. Seen across all routes and both shifts; not seen on counter collections, which are checked by the customer at the counter.
Verdict: 412 occurrences a year cost £28,889, with £27,000 a year of customer margin at risk. 3 of 6 possible causes were confirmed by evidence, explaining 98% of occurrences. The biggest root cause: slotting rule ignores look-alike risk, and nothing checks the item before it leaves. The actions should stop 96% of occurrences reaching the customer (63% prevented outright), saving £20,743 a year after £5,000 of running costs for £5,600 one-off (payback 3 months). Not yet checked after the fix: count again, month by month, once it has run.
Each credit note was assigned to its main cause, so the confirmed causes don't overlap; they explain 98% because every count comes from the same credit notes. Poor lighting in aisle 7 was ruled out as a cause of wrong deliveries (its error rate was no higher) but is worth fixing for safety.
None of the roots is "the picker made a mistake": each is something in the process that let a mistake through. That is where fixes last.
| Countermeasure | Type | One-off; running | Owner and date |
|---|---|---|---|
| Scan every item against the order at packing, with the handhelds already owned For: All causes | Error-proofing: detect before it leaves removes about 90% | £3,500 £4,200 a year | Warehouse manager 27 Nov 2026 · Under way |
| Re-slot look-alike pairs at least one bay apart; colour-coded bin labels For: Look-alike products stored next to each other (e.g. 16A and 20A breakers) | Process or design change removes about 70% | £1,800 £200 a year | Warehouse manager 13 Nov 2026 · Not started |
| Standard picker induction covering look-alikes; first-month picks double-checked For: New pickers unfamiliar with part codes | Standard work removes about 50% | £0 £600 a year | Packing lead 30 Oct 2026 · Not started |
| Print the unit of measure in bold on the pick list For: Unit of measure confused (pack of 10 against single) | Process or design change removes about 70% | £300 £0 a year | IT contractor 23 Oct 2026 · Not started |
Scanning at packing catches mistakes from every cause before they leave, so it is combined with the other fixes rather than added to them; it saves the credits and re-deliveries, but a wrong pick caught at packing still has to be re-picked, so only the prevention fixes save the hours. Scanning every line adds about £4,200 a year of packing time. The saving is an estimate until the credit notes are counted again, month by month, after the fixes have run.
Occurrences are counted from credit notes coded 'wrong item' in the accounts system, October 2025 to September 2026 and priced from a sample of real cases: cash costs, plus overtime only for the share of hours that really causes it. Causes count only when confirmed by evidence, each occurrence assigned to one main cause. How much each countermeasure removes is an assumption by type (error-proofing 90%, training 20%), not a forecast; fixes on all causes are combined, not added, and checks that catch mistakes before they leave save the cash cost but not the time to put them right. Running costs are taken off. Lost customers are shown but not counted. The saving is proven only by counting again after the fix.
At handover the managing director agreed the four countermeasures, their owners and dates. The agent was then set up in Tarnbeck's own AI assistant account, with read-only access to the accounts system's credit notes. On the first Monday of each month it counts the month's wrong items from the same records and the same definition as the analysis, adds the month to a new dated copy, checks the actions' dates and sends the managing director a short note. It uses the rules on Inputs: a month counts as above when it is more than double the expected rate, and two such months in a row are a flag. Nothing left Tarnbeck, and no decision changed, without one of its own people approving it. These events continue the fictional example.
| Date | What the agent found | Flag? | What people decided |
|---|---|---|---|
| 9 Nov 2026 | First run. October's count was 29 wrong items. Two countermeasures were done; re-slotting and scanning at packing were under way. | Note only | Nothing needed. The managing director read the note. |
| 7 Dec 2026 | Re-slotting the look-alike pairs was past its due date of 13 November, with 25 of 85 pairs still to move. | Flag | A new date of 18 December, met. Scanning at packing went live on 27 November as planned. |
| 8 Feb 2027 | January, the first month after all the fixes: 2 wrong items against 1.3 expected, within the margin set. | Note only | Nothing needed. |
| 5 Apr 2027 | The problem was returning: 4 wrong items in February and 5 in March, more than double the expected rate two months running. Seven of the nine were rush orders picked by office staff and not scanned at packing. | Flag | Rush orders go through the packing scan from 12 April. The warehouse manager tested the rush-order cause, the one left untested in the analysis. |
| 10 May 2027 | April: 1 wrong item. The reduction so far was 91% a year, against the 96% the actions should give: above the line where the method says look again. | Note only | The rush-order cause was confirmed and added to Causes with its five whys. |
| 7 Jun 2027 | The accounts system had split the 'wrong item' code in two, so the count no longer matched the definition on Inputs. The note gave no figures. | Flag | Count both new codes together, as before; the wording on Inputs was updated and May's count (1) entered. |
At the check-in after the first monthly run, on 30 November (about six weeks after handover), Geoff and the managing director read the first run together and confirmed the two rules on Inputs; Geoff's upkeep of the set-up ended there. No rule needed changing; every decision in the log that followed was Tarnbeck's own.
What the agent added. The method is unchanged: the same count, the same definition, the same countermeasures. The agent made sure the count happened every month, so the fix that was being bypassed was found within two months and the untested cause was tested. The benefit here is the result holding and problems caught earlier; no £ figure is claimed for it.
Fictional and labelled: these events continue the worked example. The same 6 events are the sample rows in the Agent log of the worked-example spreadsheet. The monthly count table on the Agent sheet shows each month to May 2027; the other sheets, including the actions' statuses, stay as at October 2026. The agent is part of the guided and done-for-you routes; the self-implement kit does not include it.