"It can't touch the money or the customer without me saying so."
I want exactly that too, so every gate on this page runs in my own businesses. The system prepares the work. A human presses send.
The fear is concrete. An automation wired to your customer list or your card machine doesn't make small mistakes. It emails everyone, or charges the wrong amount, at speed, while you're asleep. So most owners I speak to either automate nothing and stay drowning, or hand over the keys and hope.
I build the bit in between: a machine that can read, draft, sort and queue everything, and structurally cannot touch a customer or a pound without a named human's yes. The yes lives in the code itself, so it can't get skipped.
Four gates, all running in my own estate
The only cron that could email customers is switched off
In the Skool of Drool codebase there's exactly one scheduled job that could email real customers. It's disabled behind three independent guards, and the schedule line is still in the code, crossed out rather than deleted, so switching it back on is a visible edit somebody makes on purpose. The re-enable checklist ends with "confirm with Dylan and Viviana first".
Exam results can't send until a named human says so
On my exam platform, a candidate's result can't be emailed until a human green-lights that specific candidate. Until then the release route refuses, and there's exactly one place in the code that can flip it, sitting behind the check.
A review wall where nothing publishes itself
Customer reviews on Skool of Drool are moderated before they appear. Only reviews marked approved ever reach the page, so nothing user-submitted goes up unreviewed. And check.
AI suggests. A human marks.
The machine marks the points it can be certain about. The written answers go to a person, who sees the AI's pre-marking as a suggestion and decides whether to apply it. And it's pinned to a local model with a double provider check that refuses to run otherwise, so candidate essays never leave the machine.
Same pattern all the way down. Report cards generate as drafts and wait for a person to send them. And no AI I've built has ever spoken to a real customer unsupervised.
What I'd tell you not to build
The autonomous agent. If someone offers you an AI that replies to customers, moves money or posts as your brand with nobody in the loop, ask them who rings your customers when it sends the wrong email to the whole list. The boring version wins: the AI does the reading, the drafting and the sorting, ordinary code that does the same thing every time does the moving, and you do the approving. That costs you one click per send, and I would say the click is the cheapest part of the whole build.
First piece is fixed price, typically £750 to £1,500. Then we extend or stop. How pricing works →