By the end you’ll have a way to list what a task really contains, sort every step into one of three piles, and test the handover on real cases before anyone’s hours are cut.

6 MIN READ · FICTIONAL WORKED EXAMPLE AND CHECKLIST

On day six of the trial a tenant reported a broken boiler. Her third message mentioned a newborn in the flat. The assistant had graded the request routine after the first message, and routine it stayed. That one miss is why the agency in this example still has a full-time repairs coordinator.

Marlow & Reid, a 14-person lettings agency looking after 420 rented homes, is fictional, and so is everything that follows, the boiler included. Its plan was ordinary enough. Tenants’ repair requests took one coordinator about 22 hours a week. An AI assistant would read each request, grade its urgency, book a contractor and update the landlord. The role would drop to part-time.

Before touching the rota, the office manager wrote two rules. First: no step moves to the assistant until every step is listed, written or not, and sits in one of three piles: hand over, hand over with a check, keep. Second: a single missed emergency in testing blocks the switch for that step.

The step nobody had written. Two landlords always pay contractors late, and the good plumbers won’t take their jobs. The coordinator routes those to a firm that invoices the agency instead. It’s in no procedure. Tagged knowledge, written down that afternoon, and moved to hand over with a check.

The two misses. Both followed the boiler pattern: the fact that mattered arrived after the first message. Under the second rule, grading couldn’t move.

The rule that would have caught it. Checking downgrades wouldn’t have. The boiler request was never downgraded. It started routine and stayed there. So the new rule has three parts. Every new message on an open request triggers a fresh grade from a summary of the whole thread. Any request that began as routine and then receives another message reaches the coordinator within 30 minutes in office hours, whatever the new grade says, and goes to the emergency line outside them. And the re-test is scored on that deadline: an emergency that isn’t in front of a person inside 30 minutes counts as a miss, and one miss sends grading back to keep.

Reading the re-test. The 80 repeats only show that the known failures are fixed, because the rule was written with them in view. The 12 new cases are the real test, and 12 is a small number. That’s why the review date is three months, not twelve.

Where the assistant was better. Three times it drafted the landlord update she’d forgotten on a busy day. The test cuts both ways, and the write-up says so.

What stayed with a person. Deciding whether a repair is the tenant’s fault. Calming a tenant who’s rung four times. Choosing which job waits when two emergencies land together. Signing off any spend over the landlord’s limit.

The expected saving. This is a projection, not a result. During the shadow run she did everything herself, so no hours changed. Timing the four keep steps and the five checks over that fortnight came to about 9 hours a week, roughly 2 of them checking, against 22 today. The agency will measure the real figure over the first four weeks live. On the estimate it saves 13 hours a week, not the half-salary it planned on, and it keeps the person who knows which landlords pay late.

The record. ‘Repairs handover, 12 October to 8 November. 14 steps, 6 previously unwritten. Shadow run on 80 requests: 61 agreed, 19 differed, 2 missed emergencies, both where a later message changed the picture. Rule changed: fresh grade on every message, 30-minute escalation. Re-test: 80 repeats, both misses caught; 12 further cases, all escalated in time. Final sort: 5 hand over, 5 with a check, 4 keep. Hours to be measured for four weeks from the switch. Review in three months.’

Last week’s companion traced a number to its source. This is the same habit pointed at a job: find out what’s there before you rely on the summary.

Hand over tasks, not roles. A role is a bundle. Some of it is written down. Unbundle it first.

Ask about last Tuesday. A procedure is what should happen. Three real cases tell you what did.

Unsure means keep. Moving a step out of keep later is cheap. Rehiring for it isn’t.

Write the blocker first. Decide what result stops the switch before you see any results. Name the checker and the moment too.

Test on live cases, as they arrive. Tidy examples flatter the tool. Real requests come in pieces, out of order, with the key fact in message three.

Ten cases start a test. They don’t finish it. Add the rare, costly exceptions on purpose, because a normal fortnight may not contain one.

Measure the saving after the switch. A shadow run tells you about quality. It says nothing about hours, because the person is still doing the work.

‘This is process mapping. Consultants have sold it for decades, and it’s slower than letting the AI have a go.’ Both true. For work you can undo, a draft or a first-pass sort, let it have a go. What’s changed is who notices the gap. A new hire asks when the procedure runs out. An assistant doesn’t ask. It carries on, and the gap shows up at a customer. The inventory is for those tasks.

The newborn example is invented. The pattern behind it isn’t. In a preprint posted on 6 November, researchers ran 13 chatbots through ten-turn conversations with simulated users built on 112 mental-health case files. The chatbots missed 20.3% of emergencies. Shown every detail in one message, they missed 5.6%. By the last turn the decisive facts had nearly always been disclosed. The early judgement simply held. Re-presenting the facts disclosed in the conversation as one message cut misses to 12.1%. It’s one study, not yet peer reviewed, with simulated patients, and applying it to repair requests is my inference. What it supports is narrow: reassess the facts in a fresh context, and test on conversations, not on summaries. arXiv · path-dependent triage

Pick one task you plan to hand to AI this quarter. List the steps from three real cases, write down what would block the switch, and run ten live examples in parallel as a starter sample. Add two or three rare cases where a miss would be costly. Reply with one line: how many steps were on paper and how many you found. Leave out names.

By 30 June 2027, at least three companies on the 2026 Fortune Global 500 will say on the record that they have rehired for, or re-created, roles they cut citing AI. Robert Half’s 32% suggests reversals are common among firms that cut. Gartner predicts half of the companies that cut customer-service staff because of AI will rehire by 2027, under different job titles. So the hard part is the admission, and I may well lose this one.

Counting. It counts if the company says so in a statement, a filing or an earnings call, or a named executive is quoted saying so, after 11 November 2026, and the company itself attributed the original cut to AI. Anonymous sourcing, surveys, earlier cases and hiring for work the company never cut don’t count. Robert Half · 32% · Gartner · prediction

New to iPrompt? Every Wednesday, the week’s AI news turned into one thing you can do. Subscribe free