The AI Reality Check

What exact job is this AI meant to do — in one sentence?

Prevents buying "AI" as a vague capability instead of a concrete task you can check.

Where does your data go, and who else can see it?

Prevents quietly leaking customer or company data into someone else’s model.

What happens when it's confidently wrong — who catches it before a customer does?

The real failure mode isn't gibberish; it's confident, specific, wrong output at scale. Something has to catch it.

Are you buying something you should just rent — or renting something you should own?

Prevents paying monthly for a commodity, or rebuilding what you could buy for £20.

What does this actually cost at ten times the usage?

Per-use pricing looks cheap in a demo and stings at scale. Know the real number first.

How would you even know if it broke or drifted?

AI degrades silently. If nothing is watching the output, "working" is just an assumption.

What still needs a human — and did someone decide that on purpose?

Prevents automating away the judgement that should stay human by accident.

Is the "success" you’re measuring the thing the business actually gets paid for?

It's easy to optimise a number that looks good and means nothing. Measure the real outcome.

If the vendor vanished tomorrow, what would you be left with?

Prevents lock-in. Know who owns the data, the workflow, and the account.

What's the smallest version you can test before committing?

Prevents a big-bang spend before anything is proven. Validate cheap, then scale.