The decision of who builds your AI isn’t settled in the 40-page proposal. It’s settled in a one-hour meeting, in how they answer five uncomfortable questions. A vendor who ships answers with mechanics: names, screens, clauses, numbers. One who does theatre answers with marketing: "transformation", "ecosystem", "it’s all covered". This is the script for that meeting. You don’t need to be technical: just listen for whether they answer with how, or with how nice it sounds.
If what you want is the underlying criteria to compare market blocks —Big Four, boutique, freelance— that’s in the best AI consultancies: the criteria, not the ranking. This isn’t about comparing vendors: it’s the concrete questions and how to read the answer live.
What to ask an AI consultancy before signing
Five questions. Each one is built so the answer tells you more than any case study in a PDF. They’re not trick questions: they’re the ones a vendor who delivers answers without breaking a sweat, and the ones a vendor selling smoke tries to dodge with a brochure line. Ask them to their face and time how long it takes them to change the subject.
1. Who in this room today is going to build the system?
The sector’s classic pattern: the partner who charms you in the sale vanishes at signature, and the work is inherited by a junior team with no supervision while the senior sells the next project. You pay senior rates and get intern execution.
2. Will you show me one of your systems running in production, now, live?
Two trades hide behind the same name. One sells diagnosis —tells you what to automate and disappears before touching a keyboard— and the other builds the system and leaves it running. They charge about the same; only one can be checked. The acid test is asking to see shipped work, not slides: a real flow with its logs and its weekly metric. If you want the same test from the client side, the guide on which processes to automate with AI says it plainly: a system in production has a number that moves every week; a diagnosis only has a nice recommendation.
3. Where does my data end up and which vendors touch it?
Ask where your data is processed, which subprocessors are involved, whether anything leaves your region and what stays with the model providers. It’s the question that separates whoever has deployed AI for real from whoever ran a demo. If you’re going to put customer data into a model, cross-check the answer against what using ChatGPT in your company without leaking data actually takes: minimum permissions, contracts that exclude training on your data, and traceability of every action.
4. Is there an SLA and what does maintenance cost once you leave it running?
This is the question almost nobody asks and the one that hurts most to skip. An AI system isn’t a deliverable you sign and forget: models change, integrations break, edge cases show up in month three. If nobody has talked to you about what happens after "delivered", they’re selling you opening night, not the season. Ask about the service-level agreement, the response time when something fails, and the monthly cost of keeping it alive, and get it in writing before you sign.
5. Will I depend on you forever, or do you leave my team able to run it?
Many a consultancy’s business model is dependency: the less you understand what they built, the more you need them. A vendor who trusts its work does the opposite: documents, trains your people and hands over the keys. Not because it’s short of clients, but because its next job comes from reputation, not technical hostage-taking. Ask whether the system is documented, whether your team gets training to run it, and what happens the day you decide to carry on without them.
The script in one table
Take this into the meeting. Five questions, and in each one you listen for whether they answer with mechanics or with marketing:
| The question | The answer you want | The smoke signal |
|---|---|---|
| Who builds the system? | A name in the room who shows their work | "A team assigned by availability" |
| Will you show it in production? | Shares a screen with something of theirs measuring today | A case study in a PDF, nothing live |
| Where does my data end up? | Providers, regions and concrete clauses | "It’s all GDPR-covered" |
| SLA and maintenance? | Written agreement and predictable cost | "Once it’s delivered it’s yours" |
| Will I depend on you? | Documentation, training and keys | Opaque architecture, zero handover |
None of these questions requires you to know tech. They require you to listen. The vendor who delivers answers all five without getting nervous, because they live off their work running; the one doing theatre dodges them, because they live off you not looking too closely. If you’re after someone who answers all five without missing a beat, that’s how we work in operations automation: whoever sells you is whoever builds it, it’s shown in production, and it’s measured every week.