The useful question is not whether AI can do a job. It is which half of the job. Where that line sits, why it belongs there, and how to build something so it holds on a bad week rather than only in a demonstration.
Almost every job in a business splits into two halves that look like one. There is the part that involves finding, sorting, assembling and drafting. And there is the part where somebody who is accountable looks at the result and decides.
Those two halves feel like a single task because the same person does them, usually in the same twenty minutes. But they are completely different kinds of work, and they should be treated completely differently.
The first half is fetching. It follows rules, the answer already exists somewhere, and doing it faster costs nothing. The second half is judgement. It draws on things that are not written down anywhere, it is what the business is actually being paid for, and doing it faster is not obviously a good idea at all.
Hand over the fetching. Keep the deciding. Almost every sensible design follows from that one sentence.
Take a document that has to be produced from a file. The fetching half is pulling out the relevant dates, names, figures and references, checking them against each other, spotting what is missing, and laying out a first version in the right shape. That is an hour of somebody’s afternoon and it is the same hour every time.
The deciding half is reading it and knowing whether it is right. That takes ten minutes, and it takes the twenty years that let somebody do it in ten minutes.
Handing over the first hour does not reduce the value of the second. If anything it protects it, because the ten minutes get done properly by somebody who is not already tired from the hour before.
We build everything so three things stay with a person, and we say so in writing.
Decisions. Advice, and anything that amounts to it. AI gathers, sorts, drafts and flags. What it never does is settle the question. If something is going to be relied on, a person decided it.
Approvals. Nothing leaves the business on its own. Letters, documents and messages are drafted and then held. You see it, and you press send. It never presses send for you — not because it could not technically, but because the moment it can, the accountability gets blurry and nobody quite knows who agreed to what.
Money. Anything with a figure attached. AI can prepare, check and reconcile all day. It does not move money and it does not authorise a payment, ever. Amounts get put in front of a person to confirm.
It would be easy to read that as timidity, or as hedging. It is neither. It is the design that actually works, for two practical reasons.
The first is that AI is confident when it is wrong. It does not hesitate, it does not hedge, and it does not say “I could not find this so I have guessed”. It produces something fluent and plausible either way. Anything built on the assumption that a mistake will look like a mistake is built on sand — so you build on the assumption that it will look completely fine, and you put a person where that matters.
The second is that responsibility does not transfer. If something goes out under your name, it is yours, and no amount of explaining which tool produced it changes that for a client, a regulator or an insurer. Given that, the only sane place for the sign-off is exactly where the responsibility already sits.
When somebody shows you an AI tool, the question that sorts the useful from the risky is simple enough to ask in a demonstration: where exactly does a person say yes?
If there is a clear answer — here, at this point, before this happens — then it has been designed by somebody who thought about it. If the answer is vague, or the point of the tool is that no one has to look, then what is being sold is the removal of the very step that keeps you out of trouble.
The good version of this is not slower. It is the same work with the tedious hour taken out of it and the important ten minutes left exactly where they were.
Two pieces go further on where that line actually sits day to day: three things AI should never do in a cashroom, and the one rule worth having before any others. Working out exactly where it sits for your own business, job by job, is what the audit does.
More on how we handle this, including what happens to your information and what we put in writing, is on the doing it safely page.
Twenty minutes on the phone, free, no obligation. We will tell you plainly where the line would sit — or whether that particular job is one to leave well alone.