Five jobs you should not give to AI.
Knowing what not to automate with AI matters as much as knowing what to hand over. The honest list, from people who automate for a living.
We spend most of our time telling businesses what to automate. This post is the other half of the job. Knowing what not to automate with AI matters just as much, because a bad automation does not just waste the build cost. It makes mistakes at scale, in your name, while you are not looking.
So here are the five jobs we tell clients to keep human. Some of them will surprise you, because they are exactly the jobs the AI industry keeps promising to take.
1. Anything you would have to apologise for
Complaints. Refund disputes. The email from a customer whose delivery ruined their week. AI can read these, summarise them and even draft a response, and we build exactly that. But the send button belongs to a person.
The maths is simple. A routine reply got slightly wrong costs nothing. An apology got slightly wrong costs the customer, and sometimes a one-star review that outlives the mistake by years. Draft with AI, decide with a human.
2. Pricing judgement calls
We are happy to build a system that drafts your quotes from your price list and past jobs. We wrote about the pattern in our own dashboard story. What we will not build is a system that sends quotes without you reading them.
Every trade and consultancy has jobs that look standard on paper and are not. The awkward access, the customer who was slow to pay last time, the work you quietly do not want. That knowledge lives in your head, and pricing is where it earns its keep.
3. Hiring and people decisions
This is the one with legal teeth. Using AI to screen candidates or score employees puts you into territory the regulator watches closely, and the ICO's guidance on AI and data protection is clear that automated decisions about people carry specific obligations.
Beyond the law, there is the practical problem. AI screening learns from your past decisions, including the bad habits. Use AI to write the job advert and schedule the interviews. Sit across the table yourself.
4. The final sign-off on money leaving the building
Automated invoice chasing brings money in, and we build it every month. Money going out is different. Supplier bank detail changes, unusual payment requests and new payees are precisely where fraud lives, and fraudsters now use AI too.
Our rule: automation can prepare a payment, match it and queue it. A human approves it. That single manual step costs you seconds a day and is the cheapest fraud insurance you will ever buy.
5. The thing your customers actually pay you for
If you are a plumber, the plumbing. If you are an accountant, the judgement. If you run a restaurant, the food. Whatever sits at the centre of your reputation should keep human hands on it, because that is the part customers notice and the part reviews are written about.
Everything wrapped around it is fair game. The bookings, the invoices, the reports, the retyping between systems. Automate the wrapper, protect the core.
The test we use
Before we automate anything, we ask one question. If this goes wrong once a month, does the cost of that mistake outweigh the hours saved? For reports, reminders and retyping, the answer is almost never. For the five jobs above, the answer is almost always.
That question is also step one of how our pilots work. We map what should be automated, and we are just as clear about what should not. If you want that map for your business, book the free conversation, or start with the checklist below.