The How to series · Post 14
What we won't automate, and why
For UK owners who keep hearing what AI can do: the shorter list nobody says out loud, what we won't automate and the reason for each line.
It's 4pm on a Thursday and someone is pitching you the version where the machine just handles it, no person in the loop. It sounds restful. Ask one question back: which parts would you still want a name on? This post is our own answer, written down before we sell anything to anyone.
Anyone selling automation should answer what they won't automate in one breath. Our list, with the reasons:
Financial work, compliance and legal. Approval-only, always. A machine can prepare the invoice, the filing, the summary; a person signs. If it moves money or goes to a regulator, the signature has a name on it, ours included. The rule doesn't bend for a faster demo.
Anything with no right answer. If two good people would write the reply differently, it's a judgement call, and judgement calls need a human name on them. A complaint with real damage in it, a price you haven't settled: those get written by a person from the first word, because the cost of a wrong version isn't measured in minutes.
A process that's a mess. Automating a bad process makes the mess faster, not better. Tidy the loop first, then hand over the typing, in that order. Anyone promising to automate your chaos without tidying it is selling you the chaos on a timer.
So the honest limit, said out loud: it only works where work repeats and there's a right answer. The rest stays with you, where it belongs. And because we're pre-launch with no testimonials, the proof on offer is the design itself: AI drafts, a person approves, nothing sends unchecked.
The maths (a sketch, not a client figure)
What the boundary costs, directionally:
- 3 hours a week kept human on principle: the hard replies, the sign-offs, the messes tidied by hand first
- × £30 an hour = £90 a week × 46 weeks = £4,140 a year of work we deliberately keep.
That's the price of the boundary, paid on purpose. Against it sits the cost of one automated promise you didn't mean to make, which doesn't come in pounds: it comes as a client who stops calling. At sketch rates the boundary is cheap. Swap in your own numbers. (A sketch, not a client figure.)
Score it: the three checks
The checks find the boundary. Two tasks from the same week:
- The Monday debtor chaser. 20 minutes every Monday of chasing late payers (3). Writable on one page (4). A slip is caught at approval, because nothing sends unchecked (4). Three at 4 or more: hand over the draft.
- Writing off a bad debt. Rare (1). Not writable on one page; it depends on the client, the history and the conversation (1). A wrong call costs real money (1). One or less: human-only, and it's financial on top, so approval-only forever.
Same week, same tools, opposite verdicts. That's the boundary in practice: where the rules are written, hand it over. Where they aren't, don't.
Try one thing this week (free)
Write your own won't list. One page: the work you'd never hand to a machine, one line each with the reason. Pin it above the desk. Any AI pitch you take this quarter gets judged against it, including ours. Judge it by Friday: with the page written, the parts that are never for sale stop being debated.
Where a person stays
Approval-first is the pattern underneath everything we build: software prepares, a person decides. On our own system, every write is proposed, then approved, and a person taps. Financial, compliance and legal never leave approval-only. The pattern is the product, not a policy page: it's what makes the rest of the automation safe to run. We automate the typing; the deciding stays with you.
FAQ
Yes, on purpose. The sketch maths puts the kept work near £4,000 a year at one rate, and we think it's cheap for what it protects. The typing we hand over pays for the boundary many times over; the judgement we keep is what clients are actually paying for.
Anything that moves, commits or reports money: invoices, payments, payroll runs, tax filings. A machine can prepare any of it, and a person signs. The test isn't how hard the task is; it's whether a person can stand behind it later.
No. The mess gets automated too, just faster. Tidy the loop first, then hand over the typing. The tidying is usually a week of writing down what actually happens, which is the first useful look most processes have had in years.
The categories are. The edges move as tools improve and as a loop earns trust. The rule underneath doesn't: a person decides what goes out the door, and financial, compliance and legal work keeps a signature on it. Any pitch that removes the person is one we'd decline for ourselves.