Five Jobs You Should Not Give an AI Employee
Knowing where AI doesn't belong is what makes the rest of it trustworthy. Here are five jobs to keep firmly in human hands — and the reason each one fails in a way you won't notice until it's expensive.
Most articles about AI are about what it can do. This one is about the opposite, because in practice the businesses that get the most out of an AI employee are the ones that drew a clear line around it early.
An AI employee is very good at jobs with a settled shape and a known right answer. It gets into trouble on jobs where the right answer depends on judgment, relationship, or consequence. The trouble is that it doesn't look like trouble — it looks like a confident, well-written response that happens to be wrong.
Here are the five jobs to keep out of scope, and what usually goes wrong when they don't stay out.
1. Anything that commits your money
Approving an invoice. Issuing a refund above a trivial amount. Agreeing a discount to save a deal. Signing off a purchase order.
An AI employee can absolutely prepare all of these — pull the numbers, match the PO to the delivery note, draft the credit note, flag the discrepancy. What it shouldn't do is press the button at the end. The failure mode here isn't dramatic; it's a slow leak. A slightly-too-generous refund policy applied a hundred times, or a duplicate invoice paid because it looked legitimate, costs real money and takes months to surface.
The rule: AI prepares, a human approves. Set a threshold — say anything under £50 goes through, anything above lands in someone's queue — and you keep the speed on the small stuff without ever handing over the chequebook.
2. Anything with a legal or regulatory consequence
Tenancy notices. Employment decisions. Medical or clinical advice. Tax positions. Anything where being wrong means being wrong to a regulator, not just to a customer.
The problem isn't that AI is careless here — it's that these areas are full of rules that look general but aren't. The right answer varies by jurisdiction, by contract, by the specific circumstances of the person asking. A confidently phrased general answer is worse than no answer, because the recipient acts on it.
The rule: in regulated territory, the AI's job is to route, not to answer. "That's a question for Sarah, I've passed it to her and she'll come back to you today" is a perfectly good response, and it's the right one.
3. The conversation where someone is upset
A complaint that's escalated. A customer who's genuinely angry. Bad news about a delay that's going to hurt someone's plans. A cancellation someone is trying to talk you out of.
These conversations aren't really about information. They're about being heard by a person who can actually do something. An immaculate, empathetic-sounding automated reply in this moment reads as a brush-off — and it is one. People can tell, and the resentment sticks to your brand rather than to the technology.
The rule: train the escalation, not the reply. An AI employee that spots frustration early and hands over quickly, with the full history attached so the human doesn't ask the customer to repeat themselves, is worth far more than one that tries to smooth it over.
4. Anything that's actually a relationship
The call to your biggest client. The negotiation with the supplier you've used for fifteen years. The check-in with the customer who's gone quiet and might be leaving.
You could automate the scheduling of these. You should not automate the substance. A relationship survives on the accumulated evidence that a specific person is paying attention to you. Outsource that and you're not saving time, you're spending goodwill you built up over years, on a saving that's smaller than you think.
The rule: if losing this account would ruin your quarter, a human makes the contact. Let the AI make sure the human never forgets to.
5. Any job nobody can describe
This is the sneaky one. Sometimes a task fails the test not because it's risky, but because when you sit down to explain how it's done, nobody can. It lives entirely in one long-serving employee's head, and every case is a judgment call informed by twenty years of context.
You cannot automate what you cannot describe. Attempts to do so produce something that handles the easy 60% and silently mangles the rest — and the mangled 40% is exactly the part that needed the twenty years.
The rule: if two people in your business would explain the job differently, it isn't ready. Sometimes the useful project is writing the job down, and automation comes a year later. That's a real result, not a consolation prize.
Why the boundary is the point
There's a temptation to read a list like this as a list of limitations. It isn't. It's the thing that makes the rest work.
An AI employee with a narrow, clearly stated remit — answer these questions, log these jobs, chase these documents, escalate everything else — is something your team can trust and your customers don't resent. An AI employee with an unclear remit gets second-guessed by everyone, which means every output gets checked by a human anyway, which means you've added a step rather than removed one.
The businesses that do this well are usually the ones that wrote down the "not this" list before the "yes this" list. It takes an afternoon, and it saves the awkward conversation eighteen months later about why a customer got a refund they shouldn't have.
If you want a straight read on which of your jobs sit safely inside the line and which don't, our free AI-readiness audit is built around exactly that question. Our services page covers what we take on and, just as usefully, what we tell clients to keep.
At BuildPulse, we care as much about where your AI stops as where it starts.