How to choose an AI consulting firm
Seven questions that separate serious operators from slide-deck sellers, with the good and bad answers spelled out before you sign anything.
Austin Vu · Managing Partner, Qonnex
A mid-market company evaluating AI consultants today will meet firms whose pitch decks are nearly interchangeable: a maturity curve, a case study from a company ten times their size, and a projected saving with no method behind it. The decks converge because decks are cheap. What separates firms shows up only when you ask questions the deck was not built to answer.
Seven questions do most of the work. For each one, the shape of a good answer and a bad answer is predictable, which is exactly what makes them useful.
Typical selection
- Shortlist by reputation
- Sit through pitches
- Compare proposals
- Pick the best presenter
- Find out during delivery
Evidence-first selection
- Request a diagnostic first
- Review findings in your own numbers
- Agree baseline and measurement
- Meet the actual delivery team
- Commit with a defined exit
1. Can I see the diagnostic before the pitch?
Ask the firm to study your operation and show findings before any proposal is discussed.
A good answer sounds like: "We spend two to three weeks in your processes and your data, then present what we found in your numbers. If the findings are thin, we will say so and stop there." The findings should name specific processes, specific hours, specific dollar amounts.
A bad answer sounds like: "Our methodology is proprietary," or a proposal that arrives before anyone has looked at your business, quoting results other companies achieved. A projection built on someone else's operation is a brochure, not a diagnosis.
A firm confident in its diagnosis shows you the diagnosis. A firm confident only in its sales process shows you a deck.
2. How will results be measured, and who runs the measurement?
This is the question most buyers skip, and it is the one that determines whether you ever know what you bought.
A good answer: a baseline agreed in writing before work starts, results measured against your own ledger and system data, and your team reviewing the numbers before anything is settled. You should be able to hand the measurement to your finance lead and have them follow it line by line.
A bad answer: a dashboard the vendor builds and controls, "value narratives," or savings estimated from time-study assumptions nobody validated. If the seller is also the scorekeeper, the score will flatter the seller.
3. What happens if the results don't come?
Every firm projects success. Ask what the contract says about the other case.
A good answer names a concrete commercial consequence: payment is reduced or not owed, or a defined remediation period runs at the firm's cost. The firm should be able to point to the clause.
A bad answer is a reassurance in place of a term: "That has never happened to us," or a change order that quietly converts the shortfall into a second engagement. If failure costs the firm nothing, you are carrying all the risk while they carry the projector.
4. Who actually implements, the people I met or subcontractors?
Consulting has a well-worn pattern: partners sell, and a rotating bench delivers. It is not always fatal, but you deserve to know before you sign.
A good answer names the delivery lead, puts them in the room during evaluation, and lets you interview them. Ask how many other clients that person will carry at the same time; more than three is a schedule, not a commitment.
A bad answer keeps the team abstract: "our delivery organization," "our certified partners." If the firm will not name the people, assume the people change monthly.
5. What will my team have to do?
No automation project succeeds as a spectator sport, and any firm claiming otherwise is planning to fail quietly.
A good answer is specific and slightly uncomfortable: a named internal owner, four to six hours a week from the process leads whose work is changing, data access in the first week, decisions returned within days. Firms that ask for real commitment have usually delivered before.
A bad answer is "we handle everything." That promise sounds like service and works like distance: six months later a system nobody inside your company understands is quietly abandoned.
6. What won't you take on?
A firm's refusals reveal more than its references.
A good answer draws real boundaries: outcomes that cannot be measured in money within a reasonable window, projects where the data cannot legally leave your systems, timelines shorter than a fair measurement period. Firms that decline work have a model; the model is what protects you.
A bad answer is "we do end-to-end transformation across every function." A firm with no edge cases has no method, only appetite.
7. How does the engagement end?
Ask for the ending on day one.
A good answer describes a defined end state: your team operating the system, documentation in your hands, tooling under accounts you own, and any ongoing support priced as an option rather than a dependency.
A bad answer has no ending at all, an open subscription in which the consultant becomes a permanent organ of your company. Dependence is a revenue model; it should not be yours by default.
The pattern behind all seven
Each question forces the same trade: evidence before commitment, terms before trust. Firms built on delivery pass easily, because the diagnostic, the measurement, the named team, and the exit are how they already work. Firms built on selling stall, because every question moves the decision away from the pitch and toward proof.
This is the standard we hold ourselves to at Qonnex: the analysis and strategic review are free; beyond that, you don't pay unless verified savings exist. The full sequence is laid out on the how it works page. But the seven questions above are not ours, and they work on any firm you are evaluating, including us. Ask them everywhere, in the same order, and write the answers down. The comparison usually makes the decision for you.