Software that sells, and software that just costs.

The adoption figures are settled. The Bank of England and the FCA surveyed 118 firms and found 75% already using AI, with another 10% planning to within three years. The ONS puts wider UK business adoption lower, at around one in three. Nobody needs another article telling them AI is here.

The interesting question is the one underneath: of the firms that bought, which ones can point at what it did?

The one-sentence test

Before buying anything, complete this sentence honestly:

Three blanks. Each one fails differently, and each failure has a smell.

No number. The purchase is defensive — bought because a competitor has one, or because a board asked. These are the tools that get renewed for three years and opened twice. They are not always wrong; sometimes the defence is real. But they should be bought with that named as the reason, and budgeted as marketing rather than as efficiency.

A number nobody counts yet. The most common failure, and the most avoidable. If you cannot state today's value, you will never be able to demonstrate the improvement — and in twelve months the renewal conversation will be a discussion of vibes. The fix is free: start counting before you buy. A fortnight of tallying is worth more than any pilot.

Nobody checks the output. Either the output does not matter, in which case reconsider the whole purchase, or it does and the supervision cost has been left out of the business case. For anything client-facing or regulated that cost is not optional — the safeguards that came into force in February require a route to human intervention on significant automated decisions, and a route is a rota.

Why case admin is the standard example

The clearest wins we see are boring, and they are boring for a reason: the number was already there.

Research published by Nivo in May 2026, from conversations with more than seventy specialist lenders and brokers, found that only half of submitted information is right first time, that a typical secured loan case takes fifteen to twenty rounds of messages, and the broker-to-lender leg alone consumes upwards of five hours of admin, with a 35% chance of coming back for corrections.

Every one of those is already counted by somebody, usually in their head and bitterly. That means the sentence completes on the first try: this moves the correction rate, which the case manager already tracks, and the case manager checks the output. A firm buying against that can tell in six weeks whether it worked. The full breakdown is here, and the system we build against it is deliberately narrow.

The two honest exceptions

Buying a capability you do not have. Answering a question you currently cannot answer at all has no baseline by definition. That is legitimate — but say so, and set a date by which the new number will exist.

Buying to remove a risk. An insurer's requirement, a regulator's expectation, a single-person dependency. The value is the tail you cut off, and you will never observe it. Also legitimate; also worth writing down as the reason, so nobody judges it on efficiency later.

What is not legitimate is buying against a number, failing to move it, and quietly re-describing the purchase as one of these two afterwards.

What this does not tell you

It does not tell you the tool will work. A perfectly framed purchase can still fail on the product, the data or the people. The test filters out the purchases that cannot be judged, which is a different and larger category than the purchases that fail.

It also does not settle whether the saving is big enough to be worth having. A tool can move a real, counted number and still lose money, because adoption costs eighty hours before the subscription — that is the £200 floor. Run this test first to see whether the purchase is judgeable, then run the floor to see whether it is worth judging. The triage tool does both in six questions, and it says no more often than yes.