Most of what you want to automate does not need AI. Nobody quoting for it will say so.

Somebody has written "AI" on a brief. It may have been you. The process is real, it eats hours every week, and three suppliers have quoted for an AI-powered version of it. The quotes differ on price and on very little else.

Before you sign one, it is worth asking the question none of the three will raise: does this need AI at all?

The conventional wisdom, and the catch

The rule everyone repeats is that if a task is repetitive, AI can do it. That is true. It is also the wrong test.

If a task is repetitive and you can write the rules down, AI is the most expensive and least reliable way to do it. Ordinary software has done that kind of work for forty years. It runs for pennies, gives the same answer every time, and when it is wrong you can find out why and fix it once.

AI earns its place at the other end: where the input arrives in a shape you cannot predict, or where the rules exist only in somebody's head and change on Tuesdays. Most briefs we see contain some of each. The mistake is buying an AI system for the whole thing because AI is the word that got the budget approved.

What the numbers say about buying AI you do not need

Gartner predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs and unclear business value. Its own follow-up this year says the real figure was more than half. "Unclear business value" is a polite way of saying nobody checked whether the problem needed this solution.

McKinsey's latest global survey found that 88% of organisations now use AI somewhere, but only 39% can point to any effect on earnings, and most of those put it below 5% of profit. The 6% McKinsey calls high performers did one thing differently: 55% of them redesigned the workflow rather than bolting a model onto the existing one, against 20% of everyone else.

The value was in fixing the process. The model was, at best, one component of the fix.

The UK figures are more sobering. The ONS reported in July 2026 that 35% of businesses with ten or more staff now use some form of AI, and that the most common use, at 18%, is generating text. The average adopter uses 1.6 AI technologies, one in ten describes their use as extensive, and the ONS's own reading is that this "would imply relatively limited transformative impacts to date". The barrier businesses cite most is difficulty identifying a use case.

That is not a technology gap. It is a large number of organisations who bought the tool and are still looking for the job.

When "AI" means something else entirely

The pressure to call things AI is now strong enough that regulators fine people for it. The US Federal Trade Commission has brought thirteen AI-washing cases since 2024, seven of them over claims made to business customers rather than the public. The best known is Presto Automation, which told investors its voice-ordering AI had "eliminated the need for human order-taking". The SEC found the vast majority of orders needed a human to step in.

The lesson is not that suppliers are crooks. It is that "AI" is a word that gets budget, so it attaches itself to things whether or not it is doing the work. Sometimes the work is done by a person in a call centre. More often it is done by ordinary code, with a model somewhere in the pipeline so the brochure is not lying.

You pay the AI premium either way.

What it looks like from the inside

We build AI systems for a living, so this is not an argument against them. It is an argument about where they go.

Bank statements. We built a system that turns clients' bank statements into structured transactions. The part that makes it work is not the model. It is a deterministic extraction pipeline per bank, because every bank lays its columns out differently and does so consistently. The model earns its keep on the long tail: ambiguous descriptions, a scanned page, a layout nobody has seen before. That is a minority of the work. The rest is parsing.

Support triage. We built a classifier that reads inbound support mail and files it by issue and urgency. The model is genuinely needed there, because email has no structure and people describe the same fault a hundred ways. But which team gets it, what counts as urgent, and what happens after four hours with no reply are rules. The model reads. The code decides.

Chasing overdue invoices. This is a request we hear often, always with AI in the title. Overdue is a date comparison. Who to chase is a field on the invoice. The email is one template with three variants depending on how late it is. There is nothing for a model to do, and the honest answer is a scheduled job and a fortnight of somebody's time. We have given that answer, and it is the least popular thing we sell.

Which is the uncomfortable bit. Everyone quoting for your brief is paid more to build the AI version, and an AI consultancy telling you that you do not need AI is not an obvious business model. That is not dishonesty. It is incentive, and it means their answer to "do we need this" cannot be trusted on its own.

How to tell which one you have

When should you use AI rather than rules-based automation? Use rules when you can write down every decision the process makes, the input arrives in a predictable format, and a wrong answer is expensive. Use AI when the input is unstructured text, images or speech, when the rules cannot be fully written down, and when a small error rate is tolerable because a person or a check catches it. Most real processes are mostly the first with a little of the second, and the model should be confined to that part.

A study of 127 tasks in financial-services IT operations published in April 2026 found the pattern in the wild: AI agents completed 95% of the simplest tasks unaided, about 70% of the middling ones, and about 40% of those with real compliance weight. The tasks that suit a model are the ones with the least riding on them. The ones that matter still need the rules written down, with or without a model in the loop.

Three questions to ask before signing any of those quotes:

• Write down what the process decides. If you can finish the list, you do not need a model for that part. If you get stuck on "it depends", that is where the model belongs, and only there.
• Ask what happens to the one in twenty. A model is wrong sometimes and you will not know when. Rules are wrong consistently, which is fixable. If a 5% error rate is unaffordable, the model cannot be the thing making the decision.
• Ask each supplier what they would build without the word AI in the brief. A good one has an answer, and it will be cheaper. Then ask what proportion of the quote is the model. It is usually less than the proposal implies, and the expensive part is the integration nobody scoped.

The cheapest possible outcome

The best result you can get from this is a day that ends with "don't build it, do this smaller thing instead". That is what a sanity check is for: £1,250, a look at how the process actually runs, and a written answer that is allowed to say you do not need us.

If the answer turns out to be that you do need AI, you will at least know for which part. The quotes get smaller when the word is only on the bit that earns it.