Most AI pitches are built to survive one meeting. They rarely survive five direct questions.
Everyone is moving fast, everyone has a deck, and it is genuinely hard for a business owner to tell a considered answer from a confident one. The fix is not to become an AI expert overnight. It is to ask the same five questions of anyone who wants to sell you AI, including us.
Each one comes with why it matters, and with our own answer. Not because ours is the only correct one, but because showing the standard is more useful than telling you we meet it.
Who owns the work after the supplier leaves?
This is the question that separates a tool from a trap.
Some AI work quietly depends on the supplier staying involved forever: a prompt only they can edit, an account only they control, documentation that lives in someone's head rather than in your hands. You rarely notice this at the start. You notice it the day you want to change agency or hire in-house, and discover you cannot.
Our answer: you own it. The code, the prompts, the workflow documentation, the accounts: handed over as we go, not bundled up at the end as a bargaining chip. If we stopped working together tomorrow, you would have everything needed to keep running what we built, or to hand it to somebody else entirely. We do not build in dependency by design, because a tool that only works while you keep paying us is not a tool. It is a subscription wearing a disguise.
What will the monthly running cost include?
The number in the proposal is rarely the number that shows up in month four.
AI running costs stack up from several things: the model calls themselves, any hosting or storage the workflow needs, and the ongoing time somebody spends checking the output. A supplier who only quotes the build fee has told you a third of the story.
Our answer: we name the running costs before you commit, not after the first invoice arrives (what the model calls are likely to cost at your volume, what if anything needs hosting, and how much review time the workflow will realistically need each week). What a first AI project usually costs sets out that pricing in more detail. Ask for the monthly bill before you ask for the build.
What happens when the AI is wrong?
Every AI system gets something wrong eventually. The question is not whether that happens, but what happens next.
A supplier who has not thought about this will talk about accuracy percentages and move on, a sign the failure mode has been hoped away, not designed. The useful answer names a specific moment: where a person checks the work, what they are checking for, what happens if the answer looks off.
Our answer: anything customer-facing, anything that changes what someone pays or is told, and anything irreversible gets a person's approval before it goes anywhere. The AI drafts, sorts, or flags; a named person checks the result at a clear point before it becomes final. That review step is not a workaround for AI being unreliable. It is how you get the benefit of its speed without handing over judgement it was never meant to carry.
Can you show something similar running today?
This is the question that is worth pushing on.
Plenty of AI suppliers can describe what they would build. Fewer can point at something equivalent already running, in daily use, with a real name attached. A demo built for the pitch shows the supplier can make something work once, on their best day. It does not show the thing surviving a Tuesday.
Our answer: how we run our own business on AI (the scheduled agents, review systems, and daily tools we actually use to run Stuntrocket, not a curated highlight reel built for prospects). A supplier who will not show you their own working AI is telling you something, whether or not they mean to.