"We want AI in the product" is a request shaped like an answer, and it usually arrives before anyone has written down the question it is meant to solve.
The questions we ask first
- What decision or task does this replace or speed up, specifically, for a specific user, today.
- What does it cost when the model gets it wrong, in money, time or trust, and who bears that cost.
- Would a simpler, deterministic feature solve most of it, because a well-designed form or a sorted list often gets most of the value at a fraction of the cost and none of the unpredictability.
- Who is accountable for the output, once it is live, in the org chart rather than in the abstract.
What this filters out
Roughly a third of the AI feature requests we get do not survive this conversation, usually because the honest answer to "what does it cost when it's wrong" turns out to be "quite a lot, to someone who isn't in this meeting" — a customer, a regulator, an employee whose work is being judged by the output.
What survives and gets built well
Features where being right most of the time is genuinely valuable and being wrong occasionally is cheap and visible tend to work well: drafting rather than deciding, summarising rather than acting, flagging for review rather than auto-approving. The common thread is a human still in the loop at the point where a mistake would matter.
The interesting AI conversation with a client is not about the model. It is about what happens when the model is wrong, and almost nobody arrives ready to have that part of the conversation.
Where we push back hardest
Requests to fully automate a decision that currently requires judgement and has a real cost of error — credit decisions, medical triage, anything touching eligibility for a benefit — get the most resistance from us, not because the technology cannot do it, but because the accountability question rarely has a good answer at the point someone first asks for it.
What this has done for our client relationships
A handful of clients have been annoyed by this conversation in the moment and have come back after it, once, specifically to say the version we talked them into was easier to explain to their own board than the version they originally asked for.