Contents
There's a mistake we see a lot at the moment.
Someone discovers AI, it's amazing, and they try to make it run everything. Answer the enquiries. Book the appointments. Send the invoices. Decide who gets a discount.
Then one day it does something slightly different to what it did yesterday, and nobody can work out why. 🤦
Two kinds of work
Nearly every task in your business is one of two kinds.
Work that needs judgement. Reading a rambling email and working out what the person wants. Writing a friendly reply. Summarising a call. Sorting 200 enquiries into "hot", "maybe" and "timewaster". There's no single right answer, and a sensible guess is useful.
Work that has to be the same every time. Sending the invoice for the right amount. Booking the slot that's actually free. Not texting someone who unsubscribed. Deleting nothing you shouldn't. Here, "a sensible guess" is how you end up apologising.
The techy word for the second kind is deterministic. It just means: same input, same result, every time. A calculator is deterministic. Ask it 2 + 2 a thousand times, you get 4 a thousand times.
AI isn't. Ask it the same question twice and you'll get two slightly different answers. For writing a reply, that's fine, even good. For working out a bill, it's a disaster.
So split the job in half
The best systems we build use both, and keep them in their own lanes.
AI does the thinking. It reads the message, works out what it's about, pulls out the useful bits (name, date, what they want), and writes a draft.
Rules do the doing. A fixed automation takes what the AI worked out and acts on it. It checks the calendar properly. It applies your real prices. It sends only if the conditions are met.
Here's a real one.
An enquiry lands on a Saturday: "Hi, can you do a boiler service next week, not Tuesday, and is it still £85?"
- AI: reads it, tags it "boiler service", notes "next week, not Tuesday", spots it's a price question, drafts a friendly reply.
- Rules: look up the actual price from your price list (not whatever the AI remembers), check real free slots, drop them into the reply, and hold it for a human to send if it mentions money.
The AI never touches the price or the calendar. The rules never have to understand English. Each does what it's good at.
A rule we live by ourselves
We use AI heavily inside NotLuck. It drafts our emails, reads our inbox, chases our tasks.
But the thing that decides whether anything leaves the building isn't the AI. It's a plain, boring rule: no email, message or post goes out unless a human has said a specific password in their latest message. The AI can't argue with it, forget it or interpret it creatively. It's a lock, not a suggestion.
That's the pattern. AI suggests. Rules decide.
How to tell which is which
Ask one question about any task: "If this came out slightly different each time, would anyone mind?"
- A reply worded a bit differently? Nobody minds. Give it to AI.
- A summary of a call in slightly different words? Fine. AI.
- An invoice for a slightly different amount? A booking at a slightly different time? A message to someone who opted out? Everyone minds. That's a rule.
The point
AI is the brilliant new member of staff who can read anything and write anything. You still wouldn't give them the company card on day one.
Let it think. Let it draft. Let it sort. Then hand over to rules that do the same thing every single time.
That's how you get the speed of AI without the "why on earth did it do that" moments. 😀
