The How to series · Post 4
How to hand a job to AI without losing control of it
For any UK owner about to give a job to AI: the three checks that decide which job goes first, and the one rule that keeps you in control of it.
A few weeks ago, on a Tuesday afternoon, a tool started drafting the chase emails. It was quick. It was also wrong once: last month's prices, straight out the door to a client. Now the drafts sit in a folder and the owner retypes them by hand anyway. The tool didn't fail. The hand-over did.
That's the loop this post is about: the hand-over itself. Hand a job over well and the week gets quieter. Hand it over badly and the tool goes in the drawer, and the job goes back on a person, plus the checking, plus the distrust.
The maths (a sketch, not a client figure)
Take one job: drafting the chase emails, about 30 minutes a day by hand.
- × 5 days = 2.5 hours a week
- × £30 an hour = £75 a week
- × 46 working weeks = £3,450 a year
Now the same job with a check step. Software drafts, a person reads and approves in 5 minutes a day:
- 5 minutes × 5 days = 25 minutes a week
- × 46 weeks = about 19 hours a year of human time, instead of 115.
The gap between those numbers is only real if the check actually happens. Skip the check and you're back to reading every draft from scratch, which is the worst of both: the machine's speed with your doubt on top.
Score it: the three checks
Pick the job before the tool, and score it:
- Hours. Four or more hours of someone's week, every week. Under that, the hand-over costs more than it saves.
- Repeat. Could a new starter follow the job from one written page? If you can't write the page, nobody can automate the job.
- Risk-if-wrong. If it slips once, how bad? A recoverable slip scores high. A slip that costs a client scores low.
Our example job scores Hours 4, Repeat 4, Risk 3. Verdict: two checks at 4 or more, so start narrow. One kind of email, drafted and approved, for a fortnight. Then widen. All three at 4 or more, hand it over. One or less, it stays human. Most owners run this backwards: they start with the one job that must never go wrong, get burned, and spend a year not trusting any of it.
Try one thing this week (free)
Write the page for the job you'd hand over first. What starts it, the steps in order, what finished looks like, and one line at the bottom: what must never go out unchecked. Then run the page test: give it to someone else and have them follow it once. By Friday you'll know whether the job is writable. If it isn't, it isn't ready to hand over, and knowing that early is worth the hour it took.
Where a person stays
The whole design is one line: software prepares, a person decides. AI drafts, a named human approves, nothing sends unchecked. Financial, compliance and legal work stays approval-only, always. Every approval is logged, so nobody argues later about who checked what. That's how you hand over the typing without handing over the judgement, which is the only hand-over worth having.
FAQ
Then the hand-over was designed wrong. Every draft lands in front of a named person with the source attached, and nothing sends unchecked. Wrong drafts get caught in two minutes. Automation with no named approver isn't a hand-over. It's hoping.
One page, one job: trigger, steps, what finished looks like, what never goes unchecked. The page doubles as the training doc for whoever covers the job when the person who owns it is on holiday.
The jobs worth handing over are the ones people hate. Say plainly which hours come back and where they go next. The approval step keeps a person on every decision that matters, so the tools make more of your team, not less.
One. Watched for a fortnight, then the next. Handing over five in one go is how nobody trusts any of them by Friday, and the checking lands back on you.
A queue, not a flood. Drafts wait in one place with what they came from attached. A named person taps approve, edit or reject, and the log keeps the receipt. On a quiet week that's minutes a day, not a new job.