How to Choose an AI Consultant: Questions, Red Flags and a Scoring Sheet

What kinds of AI consultant are there?

Decide the kind before you take a call. The title covers three different jobs, and each one ends with something different in your hands on the last day.

The deliverable defines the kind, not the title. Ask every candidate what you will hold when they leave.

• The strategy advisor. Studies your business and recommends where AI fits. Delivers a plan. Fits a business that cannot yet name the task.
• The builder. Builds the workflows and hands over a working system. Delivers software. Fits a business that knows the task and its number.
• The operator. Builds it, runs it and owns the result in your numbers. Delivers a measured outcome. Fits a business that wants a result, not a project.

Strategy only or build: a three-question test

Answer three questions about the task, not the technology. Each wrong answer has a known cost.

Strategy without a task gives you a deck nobody runs. A build without a baseline gives you a prototype nobody can prove. RAND's 2024 study, based on 65 interviews with data scientists and engineers, names the root cause: people misunderstand or miscommunicate what problem AI should solve.

Questions one and two attack that cause directly. Three yes answers mean you hire a builder or an operator. A no on question one means strategy work, with a fixed fee and an end date. Score the task first with the AI readiness assessment.

1. Can you name the task in one sentence? If no, buy strategy work first. If yes, skip it and go to question two.
2. Do you know how that task performs today, as a number? If no, buy a short baseline before any build. If yes, go to question three.
3. Who runs it after launch? If nobody on your side can, hire an operator or price run support into the deal. A document will not run it.

Why do most AI projects fail to pay?

Most AI projects stall before they pay. The studies below use different samples and different definitions, so read each one as its own number. Do not add them into one failure rate.

The cause is mostly approach, not technology. MIT NANDA's 2025 report says the gap does not seem driven by model quality or regulation. RAND's 2024 interviews point the same way: teams that chase the newest advance for its own sake fail more often than teams focused on one problem.

• 95%of organizations get zero return from generative AI spending, preliminary findingsMIT NANDA, July 2025
• 42%of businesses scrapped most AI initiatives, up from 17% a year earlierS&P Global Market Intelligence via CIO Dive, March 2025
• 60%report minimal revenue and cost gains from AI, 1,250 senior executivesBCG, Sept 2025
• 30%of generative AI projects forecast to be abandoned after proof of concept by end of 2025Gartner, July 2024

Does buying beat building, and what does it mean for your hire?

In MIT's interview sample of 52 organizations, outside partnerships reached deployment about 67% of the time. Internal builds reached it about 33% of the time.

The report warns that this link does not necessarily prove causation. The figures are self-reported, and the sample may not represent the wider market. Read it as a signal, not a rule.

The more useful finding is how the best buyers behaved. MIT says they treated providers less like software vendors and more like business service providers, held to operating outcomes. Do the same. Hire on a measured number, not on a demo.

A wrong hire costs you the fee plus the months until you find out. This week, ask each candidate which number they will move and who will measure it.

What to ask on the first call, and what a good answer sounds like

Ask every candidate the same eight questions. Use the cards as your script. A good answer is specific and checkable. A red-flag answer is vague, or moves the risk to you.

Question one decides most of the rest. For example, the number can be the share of new leads called back inside five minutes. Pull it from your CRM before the call, and write each answer down word for word. You score them on the sheet near the end.

How much does an AI consultant cost?

No neutral, audited price survey exists. The figures below are published by consulting firms on their own sites, so treat them as vendor figures. They show the spread, not a fair price.

Hourly rates fall as the firm gets smaller. Project fees spread wider, from $10,000 to $500,000 in one firm's range, because scope varies. Ask what the rate excludes, and whether senior people do the work.

What should the price include, and what does it cost your business?

Compare the fee with the cost of the manual work it replaces, plus the run cost after launch. Example only: a task takes 20 hours a week at $40 an hour, which is $41,600 a year.

Say a $15,000 build removes half of it and run costs $400 a month. Year one costs $19,800 and saves $20,800. That is close to break-even in year one, and the gain comes after. Without the run cost in the quote, you would have seen a clear win that is not there. The small business AI adoption statistics show what other firms spend.

A small business can afford this if it buys narrow: one task, one owner, a short pilot with an end date. Ask for the quote to state each item below.

• The baseline, measured from your records.
• The build, with the acceptance test you can run yourself.
• Handover: documents, workflows and credentials.
• The first 90 days of run cost.
• Who pays for third-party subscriptions, and in whose name they sit.
• What each change request costs, before you sign.

Red flags inside the proposal

The sales call is friendly. The proposal is where the risk sits. Read the document with this list and a pen.

Two flags come from outside the document. Be wary of a pitch that starts with a tool: RAND (2024) lists chasing the latest advance as one of the most frequent paths to failure. Be wary of relabelled products too. In June 2025, Gartner estimated that only about 130 of the thousands of vendors selling autonomous AI agents were real, and expected more than 40% of such projects to be cancelled by the end of 2027.

• Success is a delivered document or system, not a number that moves.
• Discovery has no end date and no fixed fee.
• The rate is silent on who does the work, or leaves out senior time.
• Change orders are unlimited, or priced after you sign.
• The contract renews itself.
• There is no acceptance test you can run yourself.
• No people are named.
• Nothing says your data, workflows and credentials come back on exit.
• Software licences are resold inside the fee with no line item.
• There is no run-cost estimate for month four onward.

How to build the contract around one measured number

Pick one number from your own records. For example, the share of leads called back inside an hour, or the hours a week spent on one task. The whole contract serves that number.

MIT's 2025 report found that the buyers who got results held providers to operating outcomes, not model scores. RAND (2024) advises committing a team to one specific problem for at least a year. The pilot is the first step of that commitment, not the whole bet.

The five phases below are our method, not a sourced standard. Money moves only at a gate. Add two terms: you hold audit rights over the number, and the contract says who measures it.

1. Baseline. Short, paid, and measured from your records. The result belongs to you.
2. Pilot. One task, a written pass or fail target and an end date.
3. Decision gate. You expand, revise or stop. The fee to stop is capped in the contract.
4. Production. A monthly run fee, a named owner on your side and a report on the number.
5. Exit. Data, workflows, documents and credentials come back to you without the provider's help.

Score two or three consultants on one sheet

Score every candidate on the same ten criteria. Give each one 0, 1 or 2. Points are the weight times the score, divided by two, out of 100.

Two rules sit on top. A zero on baseline, ownership or exit ends the process, whatever the total. Treat a total under 70 as a no. Both rules are ours, built for owners who compare proposals.

What to check this week

Pick one task. Write down how it performs today as a number. Name the person on your side who will own it.

Then send the same eight first-call questions to two or three consultants and score the answers on the sheet. To see how an operator-led engagement runs, read AI implementation. The AI implementation roadmap lays out the phases after you hire.