How to choose an AI consultancy

Every AI firm's website says the same things. Outcomes, partnership, deep expertise.

None of that survives twenty minutes of specific questions. Here is what to ask, in the order that saves the most time.

Stage one: does the incentive allow a no?#

Ask this first, because it disqualifies fastest.

The question. "What did you last talk a client out of building?"

Strong answer. A specific story, with the reason. It usually involves discovering the volume was too low, or that the problem was a data quality issue, or that a rule would do it.

Weak answer. A pause. Then a general statement about only doing work that adds value.

Why it matters. A firm paid out of the build has a reason to find things worth building. That is not dishonesty, it is structure, and it means the assessment you receive is optimistic before anyone writes a word.

Follow up with. "How is the assessment paid for?" If it is free, or heavily discounted against the build, the recommendation is not independent. Price that in.

Stage two: who actually turns up?#

The question. "Who is delivering this, by name, and what have they built?"

Strong answer. A name you can look up. In a small firm it is the person on the call, which is both the point and the risk.

Weak answer. "One of our senior consultants." A partner sells, a team delivers, and you meet the team in week two.

Why it matters. The variance between two people at the same firm is larger than the variance between two firms. You are hiring a person.

Follow up with. "Can we meet them before we sign?" A firm that says no is telling you something.

Stage three: read the estimate, not the price#

This is the stage most buyers skip, and it is where the money is.

Ask for the estimate split into lines. Then look for four specific ones.

• The evaluation set. Real inputs from your data with the answers you would accept. Without it nobody can tell whether a change improved anything.
• Failure handling. What happens when it is wrong. On anything that acts rather than writes, this is roughly half the work.
• Cost and rate controls. Caps, caching, and an alert before the bill rather than after.
• Handover. A runbook, and enough written down that whoever inherits it is not reading the git history.

If those four are absent, the firm has priced a demo. That is why one quote is half another. The cheaper number is not cheaper, it is a smaller scope, and the difference arrives later as change requests.

Follow up with. "Which of these four is not in your number, and why?" A good firm answers immediately.

Stage four: how will anyone know it worked?#

The question. "What will we measure, and when will you take the before?"

Strong answer. They propose a number, a definition, and a date to read it, before any code is written.

Weak answer. A promise of reporting. Dashboards are not measurement, they are a place to look at things.

Why it matters. Without a baseline taken beforehand, the closing conversation is an argument about whether it feels better. That argument is always won by whoever is more confident, which is usually the supplier.

Follow up with. "What would count as this not having worked?" A firm that cannot answer has not agreed to be judged.

The scorecard#

Five mistakes buyers make#

Comparing prices before comparing scopes. Two quotes for the same brief differ by several times because one included the four lines above. Comparing the numbers is comparing different products.

Weighting logos. A logo means somebody bought something once. Ask what changed at that company and listen for whether the answer has a number in it.

Buying strategy when you needed a decision. A document everybody agreed to, that nobody can act on, because nothing in it was costed against your systems.

Skipping the reference call. Ask for a client who did not renew, and why. A firm with nothing to hide will find you one. This is the single most informative twenty minutes available to you.

Letting the assessment be free. It feels like a saving. It buys you a recommendation from somebody paid only if you proceed.

What to do this week#

Send stage one and stage three to every firm on your shortlist, in one email. Two questions.

What did you last talk a client out of building? And which of these four lines is not in your estimate?

You will get a specific reply within a day from the firms worth talking to, and a brochure from the rest. That sorts a shortlist faster than any call, and it costs you ten minutes.

Seven firms a startup can actually hire are named and compared here, What an in-house team costs instead is here, our own bands are on the pricing page, and the work we turn down is listed on the services page.