How to buy AI without being sold to

We build AI systems for a living, which means the sales pitches aimed at our own industry land on us too. Four claims come up often enough to be worth naming plainly, with the question that exposes each one — including the version of these questions you should be asking us.

"It learns your business"

Retrieval over your own documents is real and useful: an assistant that answers from your policies, your pricing, your catalogue. A model that quietly absorbs how your company works and improves on its own, with no defined process for what changes or how you'd know, is not what most vendors are actually selling, however the phrase is used.

Ask instead: what exactly gets better, and how would I know if it didn't? If the answer is not a specific, checkable mechanism, treat the phrase as marketing rather than architecture.

Fully autonomous, and customer-facing

The technology can do it. Whether it should is a different question, and it turns on one thing: do you want to hear about mistakes from a customer, or from a review step you control?

For most service businesses, the honest version — a draft is prepared automatically, a person sends it — captures nearly all of the time saving with none of the exposure. See approvals that stall the work for the fuller argument, and the same admin work, every day for where full automation is actually safe.

Ask instead: what happens the first time it gets something expensive wrong, before it happens? If the answer only exists after the fact, the review step is missing, not just undocumented.

Anything that can't show its source

If an assistant answers a question about your own policies, prices, or contractual terms, you need to see which document and which page. Without that, a wrong answer is indistinguishable from a right one — which makes the whole system unusable for anything that actually matters. This is not a nice-to-have feature; it is the difference between a tool and a liability. See the page is the proof for how we build this in.

Ask instead: show me a wrong answer, and show me how I would have caught it. A vendor who cannot produce this on request has not built it.

"It replaces a person"

In a team of twelve, nobody is doing exactly one job. Removing a task rarely removes a headcount; it usually gives someone their afternoons back, which is valuable in a different way and should be priced differently.

Ask instead: price the task, not the salary. What does automating this specific task save in hours, and what does it cost to build and run? If a vendor's pitch leans on headcount reduction for a 12-person business, be wary — that is a sign the pricing is aimed at a narrative, not the actual arithmetic.

The through-line

None of this means don't buy. It means buy the boring, checkable version: the one where you can see the source, see the failure mode, and see the arithmetic. A system that cannot survive these four questions is not a system you should trust with anything consequential, regardless of how good the demo looked.

Turn it on us

Every question above is one you should ask when we quote you a project. If we can't show a source citation, name a failure mode, or price a task instead of a headcount, that is a reason to say no to us specifically — not just to "AI vendors" in the abstract.

Honest limits

• Scepticism is not the same as refusal. The point of these questions is to buy correctly, not to avoid buying. Most of the projects that pass all four questions are worth doing.
• Some vendors answer these questions well and still build something wrong for your business. Good answers to due-diligence questions are necessary, not sufficient — they don't replace a proper scoping conversation about your actual workflow.

← Back to AI for owner-led service businesses