How do you choose an AI consultant?
Ask every candidate the same ten questions and write down what you can verify. Start with scope, accountable people, data handling, proof, and the cost after launch. A missing answer is something to resolve before signing, not something to fill with optimism.
They are drawn from the same standard we publish against ourselves — the “what we don’t promise” list on our method page. Print this page and bring it to every meeting, ours included. If our answers and this list ever disagree, that is your red flag.
The ten questions
Each comes with what a good answer contains — and what a bad one sounds like.
Run the meeting the same way every time: same questions, same order, answers written down. Comparing firms only works when the inputs match, and taking notes changes the answers you get — vague firms get specific or get visibly uncomfortable, and both are information.
1. What will this cost, and when do I learn the exact number?
Ask for a useful range before buying discovery, and find out what must be learned before the final quote. Our audit is $3,500–$8,500 and our sprint is $18,000–$60,000. We fix the scoped price in writing before work begins. Guide 01 shows how to compare deliverables, exclusions, and ongoing costs.
Bad answer: “We cannot tell you what discovery delivers or what it costs.”
2. Will you guarantee an ROI number?
This one is a trap, and the right answer is no. Real results depend on your execution, so a serious firm quotes ranges and shows its assumptions; a firm that guarantees “300% ROI” is quoting fiction with confidence. We publish the assumptions behind our estimates precisely so nobody has to take a promise on faith.
Bad answer: “Clients typically see 10x returns in 90 days, guaranteed.”
3. How long until something actually works?
Ask for the stages, dependencies, and acceptance criteria behind the date. Our sprints typically take 4 to 12 weeks; a separate audit takes 2 to 4 weeks when needed. A simple integration may be faster than a multi-system document workflow. The useful answer explains what has to be ready, not just when the demo will run.
Bad answer: “We can have you fully AI-powered in two weeks.”
4. Who scopes the project, and who builds it?
Ask who is accountable for the scope, who implements it, who reviews the work, and who signs off. Those can be different people; the responsibilities should be clear. At Main & Machine, Christopher owns scoping and sign-off, supported by the build team. Meet the founder and see the delivery roles.
Bad answer: “You will find out who owns delivery after signing.”
5. Will you tell me if AI is wrong for us right now?
“Wait a quarter” must be an answer the firm can afford to give, or every diagnosis will be “build.” Ask when they last told a prospect not to buy. A firm whose audit can conclude “don’t build yet, and here’s why, in writing” is scoping for your benefit, not their pipeline.
Bad answer: “Every business needs AI today — the risk is waiting.”
6. What tools will you recommend, and what do you earn from them?
Ask which alternatives were considered and whether commissions, referral fees, or platform requirements influence the recommendation. A useful answer explains why the tool fits your workflow and what it will cost to leave. Our method includes recommending a free tool when it is sufficient.
Bad answer: “Our partner status is the reason this is the right tool for you.”
7. Where does our data go?
Ask which records leave your systems, where they are processed, who can access them, and how long they are retained. Our published MARCUS case processes borrower documents locally and can use external reasoning on filtered text for selected tasks. The security page explains filter and log limitations. Request the same specificity for your proposed build.
Bad answer: “It’s all encrypted in the cloud, totally secure.”
8. Who can overrule the system?
Name a person with authority to stop the system, and agree which actions may run automatically. Consequential decisions need an explicit approval path; routine replies or data movements may use defined rules. Ask how exceptions are surfaced and how the team resumes manual work when needed.
Bad answer: “The beauty is it runs itself — no oversight needed.”
9. What do we own when you leave?
Ask exactly what transfers: custom code, configuration, prompts, documentation, credentials, and data. Identify third-party licenses and services that remain subscriptions. Our sprint includes ownership of what we build and team training; the scope should make the handoff and ongoing dependencies clear.
Bad answer: “Most clients stay on our platform long-term.”
10. Show me something you built.
Ask for a working example, its starting problem, the delivered scope, and the evidence behind the results. Our public case is MARCUS, built for B:Side Capital where our founder is CEO. That affiliation matters. The results page explains the measurement window and limits; the illustrative sample audit shows the planning artifact. Ask each firm what is actual client evidence and what is an example.
Bad answer: “We’ve delivered AI transformation for leading brands across verticals.”
What are the red flags, in one list?
Pause on guaranteed returns without a defensible basis, unidentified delivery owners, unexplained data handling, hidden recurring costs, and unverifiable proof. A short timeline, a paid discovery phase, or a platform partnership is not automatically wrong; an inability to explain its fit is the problem.
A pilot needs a decision at the end: expand, revise, or stop. Ask which cases it will test, who accepts the result, and how the production costs differ. Our sprint is a scoped implementation, not an open-ended sequence of pilots.
The practical defense is a written comparison. Keep the same questions and evidence standards for every candidate. The Ampersand essay How to Smell the Hype develops the same skeptical habit.
Why publish questions we could fail?
Because we would rather be disqualified by a sharp buyer than hired by a confused one. A client who chose us against a real standard stays; a client who was dazzled churns — so the standard is good business, not virtue.
It is also the standard we already signed. Our method page lists what we promise and what we refuse to promise, and this guide is that list turned into questions anyone can ask any firm. Use it on us first: book the free 30-minute assessment and ask all ten. If you are still comparing, Guide 01 compares scope and total cost and Guide 03 runs the hire-or-engage math. Once you choose, two companions keep the engagement straight: a scope a vendor can’t inflate, and exactly what an audit should hand you.