What to Ask Before Selecting an AI Consultant

Tools alone have a poor record of changing how an organization works. Moving a revenue number takes strategy, sequencing, adoption, and pricing decisions working together, which is why the consultant you choose carries as much weight as the technology they bring. Qualifying those partners is the harder job. With a tool, you can list the functionality you need and confirm the system delivers it. With a partner, every candidate can oversell their capabilities and underdeliver on your expectations.

So how do you properly evaluate potential AI partners before you commit to anything?

‍

What you need to know about yourself

No external consultant can tell you what your goals are, what costs you can absorb, or what result would make you walk away. Before talking to any partners you must first answer some important questions about yourself.

How does the return decompose?

Whatever target you've set, name the levers that move it and assign each one a share. A revenue-per-employee target, for example, only moves three ways: win more work, price the work better, or deliver it in fewer hours and redeploy those hours toward growth. Then decide how much of the target belongs to AI at all, next to organic growth, hiring, or acquisition. Until you've allocated it, every respondent will claim the whole number, and you'll have no way to score a proposal that honestly addresses one third of it.

What does the work cost today?

Establish current performance, specifically the time spent on the tasks the proposals claim to absorb. Whoever owns this baseline owns every ROI conversation for the length of the contract. If a vendor's proof of value rests on measurements taken after they arrive, the proof is circular.

What goes first?

Which division, which workflow, which team. Sequencing is your decision, not a vendor's, and it's determined by things they can't see: where IT maturity is uneven, which group has capacity for change right now, which manager will run the thing.

What would make you stop?

Name in advance the event that would push a proposal, or the whole initiative, to a stop or a one-year deferral. Left to the vendor, nothing ever justifies stopping. Setting the boundary before you start is what keeps the money already spent from becoming the argument for spending more.

‍

What an AI consultant needs to tell you

The sales pitch was the consultant's opportunity to put their best foot forward, while these six categories are your opportunity to stress test that. Use these categories to frame your questions and evaluate potential partners.

1. Contribution to the target

This category uncovers how much of the target a consultant is claiming and the mechanism they believe will get them there. Connecting to your objectives is the stated purpose. The second thing it surfaces is whether the candidate can back a projection with something that already happened.

Ask: Name a client of comparable complexity where your work moved a firm-level financial metric. What was the metric, how far did it move, and can we speak with them?

2. Scope, gaps and handoffs

Contributing to an objective this size takes strategy, software, adoption, and pricing working together, and not every vendor covers all of it. A good one knows which parts are theirs and says so before you ask.

Ask: What must we, or another provider, do in order for your contribution to materialize? Name the dependencies specifically.

3. Data, confidentiality and ownership

Consultants and vendors both touch your files, and both route work through model providers you have not vetted. You need to know where the data goes, who has access, and whether it trains anything outside your account. Ownership and retention of what gets generated during the project belong in writing too.

Ask: Is our data used to train models or improve services beyond our own account, and can that be disabled in writing?

4. Legal, liability and professional obligations

Regulated work carries obligations that consultants and vendors do not operate under themselves. Find out where liability sits when AI-assisted work causes a problem.

Ask: If AI-assisted work leads to a claim, where does liability sit among us, our licensed professionals, and you? Point to the contract language that says so.

5. Cost, contract and exit

The initial quote is rarely the invoice three years in. Scope creep arrives without an announcement, and extended service is not a donation from your vendor. Identify where costs can expand before you sign anything.

Ask: Walk us through year three. What costs more once we are a customer instead of a prospect?

6. Viability and references

Just like a job candidate, uncover a vendor's references, past experiences, and length of practice to understand their past applications. If your project is the first time a vendor has attempted anything like it, you are paying for their learning curve and carrying the risk if it doesn't work.

Ask: Can we speak directly with two or three reference firms of similar size and sector mix?

‍

How to score qualitative answers

Every respondent answers a friendly question favorably. That is the job of a sales conversation, which means the words themselves tell you little. Cut through the eloquence and weigh the evidence instead. This simple four-point scale is one way to score responses and can be applied to any of the categories.

3 – verifiable by a third party

Answers that come with a named reference that will respond to your inquiries, a document from a real client, or a clause your attorney can read. These are answers someone other than the respondent can vouch for.

2 – specific and falsifiable

No outside proof, but the answer contains numbers, dates, named dependencies, or named people. If it turns out to be wrong, you will be able to tell.

1 – general assurance

Fluent, relevant, and impossible to check. Anything in the shape of "we handle that," "we have done this many times," or "that has never been an issue."

0 – deflection

The vendor answers a question you did not ask, redirects to a demo, or explains why the question is the wrong one to be asking.

‍

Score the answer as it was given rather than as you understood it, so the assumptions you brought into the room stay out of the number. Flag every question that scored a zero and look at them together, since the pattern tells you more than any single gap.

Then line the sheets up and look at the categories rather than the candidates. Where nobody scored well, you have found a part of your target that no respondent is covering, and that changes what you do next more than any individual ranking.

‍

Evaluating AI partners FAQs

What is an AI partner?

Any outside provider you depend on to reach an AI outcome: software vendors, implementation firms, strategy consultants, managed services, staffing. A tool makes claims about its features and you can verify those in a trial. A partner makes claims about your outcomes, which you cannot verify until the money is spent.

What is an AI consultant?

An outside advisor who helps you decide where AI belongs in your business and how to get it there. The work usually covers some mix of assessing what your operations and data can support, identifying use cases worth pursuing, building the case for what they are worth, selecting tools, and handling the adoption problem once something ships.

How do you evaluate an AI consultants' expertise before hiring?

Badges, certifications, and fluency with model names tell you someone has been paying attention. Weigh the evidence behind the claims instead, and score every candidate the same way. The question to hold each answer against is whether anyone other than the consultant could confirm it. A named client who will take your call, a document from a real engagement, or a clause in a contract all clear that bar.

What questions should you ask an AI partner before engaging with them?

Work by category rather than by question list, so the same ground gets covered with every candidate. Six areas matter: what part of your target they are claiming and how, what their contribution depends on that they do not control, how results get measured and by whom, what happens to your data and your ownership of the work, what the full cost looks like over the life of the engagement, and who has done this work before, at your scale.

Ask open-ended questions that cover each category. Some will need more than others depending on your organization and the project, and a question that can be answered yes will get answered yes. "Can you handle change management across our teams?" produces a nod. "How have you handled change management at a company our size, and what went wrong?" produces something you can check.

‍