How to Choose an AI Consulting Firm: Seven Tests for One Call

Who are the best AI consulting firms? Search the question and you get a dozen ranked lists that share almost no names, and most of them are written by a firm that put itself at the top. That is not a criticism of the firms. It is a reason to stop shopping from a list and start testing whoever is in front of you.

This guide is written to be useful whichever firm you hire, including none. It covers what an AI consulting firm actually does, the kinds you will meet, what to settle before the first call, seven tests you can run in that call, what the work costs, and the red flags. Facts first; my own answers to the seven tests are at the end, for the record, and the details of how I do this work are on my AI consulting page.

Who this is for: the CEO, CFO or operating partner of a company that wants AI working in the business, not another strategy deck. That covers a 30-person firm and a PE-backed company with a thousand people. The tests are the same; what changes is how much help you need.

Key takeaways

• Most “best AI consulting firm” lists are written by firms ranking themselves. Test the firm; do not trust the list.
• Before the first call, write down the result you want and the number that proves it, where the data for it lives, and who inside will own it after the consultant leaves.
• The seven tests: have they shipped AI that is still running, who does the work, what the first milestone delivers, whether they ask about your data before naming a tool, how they will measure it in your P&L, what you own when they leave, and whether they can get your people to use it.
• No neutral body publishes AI consulting rates. Firms’ own guides put it at roughly $150 to $300 an hour and $5,000 to $500,000 or more a project. Price the shape of the work, not the hour.
• A good firm will tell you when you do not need a firm.

What an AI consulting firm actually does

An AI consulting firm separates where AI pays in your business from where it does not yet, and then gets the first wins into production. The label covers very different work, and most disappointments start with buying one kind when you needed another.

The work comes in four shapes:

• Strategy and assessment. Map the work, put a number on each candidate, and hand you a ranked plan. The output should be decisions, not a maturity score.
• Build and integration. Connect models to the systems your business already runs on: the ERP, the CRM, the data warehouse, the ticketing system. This is engineering, and it is where most of the cost and most of the risk sit.
• Adoption and change. Redesign the work with the people who do it, train them on the new flow rather than on the tool, and measure whether the volume actually moved.
• Fractional leadership. Someone who owns the AI agenda at the level where trade-offs get made, without a full-time hire.

And the sellers come in several kinds:

None of these is better in general. The right one depends on which of the four shapes you need, and the fastest way to find out is the work below.

Settle three things before the first call

A firm can only be judged against a problem. Write these down first, even roughly.

1. The result, and the number that proves it. Not “use AI in customer service.” Something like: take a defined share of support volume off the human queue within a quarter, measured against this month’s cost per ticket. One workload, one owner, one number to beat.
2. Where the data for it lives, and who owns it. A workload that looks simple on a process map can sit on three systems, two of them a vendor’s, with no clean history. You do not need the answer. You need to know the question, because it sets the timeline.
3. Who inside will own it after the consultant leaves. A workload with no internal owner is a pilot, and pilots are where most AI money goes to wait.

There is one more thing to settle, and it is about you. You do not need to become a technologist, but you need enough fluency that an expert cannot hide behind language you do not understand. If a proposal uses a word you cannot explain to your CFO, ask for the plain version in the meeting. A firm worth hiring is happy to give it.

Seven tests, one call each

Ask these of each firm on your list, including the one a friend recommended. Each test says what it checks, what to ask, and what a good answer sounds like.

Test one: have they shipped AI that is still running?

What it tests: whether they have built and deployed AI, not studied it or advised on it. Building and launching is the only way to learn what breaks in production.

What to ask: what did you build, what metric moved, and is it still running? Can I talk to a client who is more than a year past go-live?

A good answer: a named outcome with a number, a system that is still in use, and a reference who can describe what went wrong along the way. A weak answer is logos, a count of pilots, and “we can’t share details.”

This test matters more than it used to. Fluency is cheap: someone can talk well about agents and models on a call without having shipped anything, and the label costs nothing to claim. There are guides that teach people to start an AI consulting business with a free 30-minute “opportunity call.” A system still running at a client you can phone is the evidence that cannot be borrowed.

Test two: who actually does the work?

What it tests: whether the people in the pitch are the people on the project. Senior people sell the work and junior people often learn your business on your fee.

What to ask: will the people in this room be on the delivery team? Name the person who will be in our building each week.

A good answer: names, and the same names in the statement of work. If the person who scoped it will not be the person doing it, you want to know before you sign, not in week three.

Test three: what does the first milestone deliver?

What it tests: whether production is a term of the engagement or a hope. Most proofs of concept succeed, because they are designed to: they run on clean sample data with someone watching. The failure comes afterwards, when the model meets live data and there is no owner and no path for the cases it gets wrong.

What to ask: what will be running, for real users, at the first milestone, and on what date? What happens to the cases the model gets wrong?

A good answer: one workload in production at a defined share of volume, an internal owner, a named person who handles the exceptions, and a baseline measured before the switch is flipped. Better still, the new flow runs beside the old one for a few weeks so the results can be compared before the old way is switched off. A first milestone that is only a document is a strategy engagement; buy it as one, knowingly.

Test four: do they ask about your data before they name a tool?

What it tests: whether the technology follows the problem or comes before it. Three vendors can pitch three different technologies for the same problem, and none of them may have asked what you actually need. A firm that quotes a build before asking where your data lives is selling a pilot, not an outcome.

What to ask: what would you need to know about our data and systems before you could recommend anything? Why this approach, and what would you have proposed instead?

A good answer: questions first. Where the records live, who owns them, how clean the history is, what the system of record will accept. Then a reason for the recommendation that you can repeat to your board. Two technical questions separate a real integrator from a demo: will the system respect the permissions people already have (someone who cannot see a record today should not see it through the AI), and will your own engineers be able to maintain it when a field name changes?

Here is what asking first looks like. The CEO of a $25M company asked me how to use AI without an enterprise budget. We skipped the strategy deck and went looking for the boring wins: the phone calls, texts, emails and browser work the team was re-typing into systems day after day. Those flows did not need a platform. They needed AI reasoning in the middle of known inputs and outputs. A one-time $50,000 investment took $250,000 of annual cost out.

Test five: how will they measure it, in your P&L?

What it tests: whether they think in business results or in activity. An engagement measured in logins, licenses or “use cases identified” can succeed on paper while the cost line does not move.

What to ask: what number will we report to the board at the end of this, and how will you measure it?

A good answer: a line the CFO can audit. For automation work that is usually the share of volume that no longer touches a human queue, times what a unit of that volume cost before, against a baseline measured before the change. Adoption counted in work moved, not in seats. Bonus points if the person answering has run a P&L of real size, because otherwise the plan tends to be technically elegant and commercially beside the point.

Test six: what do you own when they leave?

What it tests: whether you are buying a capability or a dependency. Dependency models look cheap in the first year and expensive by the third.

What to ask: what will my own team be able to change six months after launch without calling you? Do we own the code, the prompts, the validation rules and the documentation? Is the deliverable an application we own, or a configuration that lives inside someone else’s product? Do you take fees or referrals from any vendor you might recommend?

A good answer: you own the work product, it is documented well enough for your engineers to extend, and the firm tells you plainly which vendors it has relationships with. Add two more questions here. Who will own the record of which models run, on what data, what each is allowed to read, and what each decided? And what does it cost to keep it running, including retraining and the model bill, once the project is over?

I have seen what skipping that question costs. An enterprise bet heavily on one frontier model vendor, then noticed the bill: tokens burning with no clear proof of value. When we mapped the workloads, open models were plenty smart for a large share of the tasks, which cut some workload spend by 80 percent, and moving the speed-critical paths to newer inference hardware such as Cerebras brought responses back roughly 18 times faster. Same outcomes, a fraction of the bill.

Ask one more thing about control: where does the model’s output become a decision, and who signs off at that point? Name that person before anything goes live, check that they can actually reject the output, and log the overrides. A log with no overrides after months of use usually means the review is not happening. And if you change models next year, do those controls survive the swap?

If your business is regulated, press harder on that record. Under the EU AI Act, systems used to judge a person’s creditworthiness, to price life and health insurance, or to recruit, promote or monitor employees count as high-risk. They must be able to log their own operation automatically, and the company using them has to give oversight to people with the competence, training and authority to override the output, and keep the logs for at least six months. A firm that cannot explain in the first call how it will record what the model saw and decided, and who can override it, is not ready for that work.

Test seven: can they get your people to use it?

What it tests: the part most AI projects fail on, which is adoption, not technology. Licenses get bought, a kickoff email goes out, and three months later the work has drifted back to the old routine.

What to ask: tell me about a time you got an operating team to adopt something it did not ask for. What did you change in how the work was done, and how did you know it stuck?

A good answer: a specific story about redesigning the work with the people who do it, training on the new flow rather than the tool, and leaders using it visibly. (More on that side of the work in what change leadership is.) If the only answer is training sessions and a change-management slide, the build will be fine and the volume will not move.

Most firms pass one or two of these well. A firm that passes the first three is worth a second conversation. A firm that cannot answer the first two will not be rescued by its strategy capability.

What AI consulting costs

No neutral body publishes a rate card for AI consulting. There is no professional body surveying fees. The ranges below come from firms publishing their own numbers, often inside lists that rank the author first, so read them as a sense of scale, not a quote:

The spread is the point. The same label covers a $5,000 pilot and a $500,000 build, and large firms sit at the top of it. A low hourly rate can still cost more once you add the minimum project size and the rework.

What you can compare is the structure, because each one fits a different kind of work:

Two costs rarely appear on the first quote. The first is the data work: in most builds, getting the records clean and connected is the longest phase and the most underestimated one. The second is the running cost after launch: model fees, monitoring and retraining. The pilot bill does not predict it: an agentic workflow that loops, retries and checks its own work can burn many times the tokens the demo did, and the first signal is usually the invoice. Ask for both in writing before you compare proposals, along with the cost in the second and third year: platform add-ons and usage fees are where a business case quietly changes after the first year.

One question tells you more than the price. Ask: what would make you tell us this engagement is not worth it, and would you say so? A firm that has stopped engagements that were not working will answer without flinching.

If you need to get the spend approved, the useful number is not the fee. It is what the target workload costs you today, per month, and how long it has cost that.

Big firm, boutique or operator?

Hire a large firm when the program spans many functions and countries, the budget is measured in many millions, and you need delivery capacity more than speed. Expect timelines in quarters and a team you will meet in stages. The large firms are excellent at producing evidence and staffing a big program; they are not built to own an outcome. If the problem is a decision rather than an analysis, that matters.

Hire a boutique or an operator-advisor when the job is to find the first workloads that pay and get them into production inside a quarter, and you want the person who scoped it in the room. Mid-market companies often sit in this gap: too complex for a freelancer, below the engagement minimums of the global firms, and rarely with a machine learning team of their own. At that size, the deciding question is whether the firm can work inside the systems you already run, because you do not get to start clean.

When you do not need an AI consulting firm

• When an off-the-shelf tool already does the job. If your CRM or support platform has the feature, turning it on and training people may be the whole project. A good firm will say so in the first call.
• When no one inside will own the result. Without an owner, a build becomes a pilot. Fix the ownership first.
• When the decision is already made. If you know the workload, the data is clean and you have engineers, you may need a builder for a few weeks, not a consulting engagement.
• When the problem is not AI. If the process is broken, automating it makes a broken process faster.

Red flags

• A ranked list as the main evidence, especially one the firm wrote.
• Logos and pilot counts instead of a system still running and a client you can call.
• A partner at the pitch and unnamed staff in the statement of work.
• A tool recommended before the firm asks where your data lives.
• A first milestone that is only a document, presented as progress toward production.
• Success measured in licenses, logins or “use cases identified.”
• Governance, permissions and security deferred to “phase two.”
• A deliverable you cannot take with you: no code, no documentation, or a configuration inside the vendor’s product.
• Undisclosed vendor fees or referral arrangements.
• A promised return on investment with no baseline behind it.

Where to look

Start with the people who have already paid for this: operating partners, CFOs and CEOs you respect, and the board members who have watched an AI program from the inside. Ask who they used, for what, and whether it is still running. Lists and review directories are a second source, read with the caveat above. Then run the seven tests on three firms, not ten, and pay for a small, defined first step before you sign a large one.

If you have not yet worked out which workloads to start with, read what an AI readiness assessment should hand you first; it makes the conversations with firms much shorter.

My own answers to the seven, for the record

• Shipped and still running: I built and launched AI products as CEO of Product Genius, COO of Blameless and EVP and GM of the New Relic platform (New Relic AI, anomaly detection across two billion events a minute).
• Who does the work: me. The person on the scoping call is the person in the room, and your team does the work with me, so the capability stays when I leave.
• First milestone: a fixed-fee assessment that ends in a ranked plan with a number on each workload, then one or two workloads in production inside a quarter.
• Data before tools: I sell no software and take nothing from a vendor.
• Measured in the P&L: the share of volume that no longer touches a human queue, times what a unit of it cost before, against a baseline.
• What you own: the work runs in your systems, and one person in your company owns the record of which models run and what they decided.
• Adoption: I led a cloud and DevOps transformation across more than 1,000 engineers at Cox Automotive.

The first step has a published price: from $3,500 for a small business and $15,000 per company for a mid-market or PE-backed company, credited in full toward an Adoption Sprint that starts within 90 days. How it works is on the AI consulting page; if you want workloads in production and a smaller bill, let’s talk.

Frequently asked questions

What does an AI consulting firm do?

It finds where AI pays in your business and gets those workloads into production. The work comes in four shapes: strategy and assessment, build and integration, adoption and change, and fractional leadership. Most disappointments come from buying one shape when you needed another. An AI strategy consultant covers the first shape: which workloads to start with, in what order, and what each should return. Ask whether the strategy ends with a workload in production or with a document.

How much does an AI consultant cost?

There is no neutral rate survey. The ranges firms publish about their own work run from $150 to $300 an hour, $5,000 to $15,000 for a pilot, $5,000 to $50,000 for a first small-business project and $50,000 to $500,000 or more for a custom build. Compare the structure as well: a fixed-fee assessment to find where AI pays, a scoped project fee for a build or sprint, a retainer for standing advice or a fractional seat, and hourly rates for narrow technical work. Ask for the data work and the running cost in writing, because neither tends to appear on the first quote.

Who are the top AI consulting firms?

There is no neutral ranking, and most published lists are written by firms that rank themselves first. The best firm for you is the one that passes the seven tests for your problem: it has shipped AI that is still running, names who does the work, puts something in production at the first milestone, asks about your data first, measures in your P&L, leaves you owning the work, and can get your people to use it.

Are boutique AI consulting firms better than big firms?

They are better at different jobs. A large firm fits a program that spans many functions and countries, where you need delivery capacity more than speed. A boutique or an operator-advisor fits when the job is finding the first workloads that pay and getting them into production inside a quarter, with the person who scoped the work in the room. The large firms are excellent at producing evidence and staffing a big program; they are not built to own an outcome.

Can a small business afford an AI consultant?

Usually, if the first step is small. Firms’ own published ranges put a pilot at $5,000 to $15,000 and a first small-business project at $5,000 to $50,000. Start with one workload that costs real money each month, such as the calls, emails and data entry your team re-types, and one person inside who owns it. If an off-the-shelf tool already does the job, turning it on may be the whole project.

How long does an AI consulting engagement take?

An assessment usually takes two to four weeks, depending on the size of the company. Putting one or two workloads into production takes about a quarter. Data readiness is what adds time, which is why a good firm asks about your data before it sets a date.

Should we hire an AI consulting firm or build an internal AI team?

Often both, in that order. A firm can find the first workloads and get them running faster than you can hire; an internal owner is what keeps them running. If an engagement ends without someone inside who owns each workload, you have bought a pilot.

Do we need a fractional chief AI officer?

When the AI agenda needs an owner at the level where trade-offs get made and you are not ready for a full-time hire, yes. Most companies at that point need someone to build the machine, not to drive it, and building is finite work that rarely justifies a permanent seat. Decide what you want the role to do first: set the strategy, govern an existing roadmap, or build the data foundations. They are different jobs.

What should be in the contract with an AI consulting firm?

The result and how it is measured; the people who will do the work; what runs in production at the first milestone; who owns the code, prompts, rules and documentation; how new scope is priced; the running cost after launch; and any vendor relationships the firm has.

Should we start with an AI readiness assessment?

If you do not yet know which workloads to start with, yes, as long as it ends in a ranked plan rather than a score. Here is what a good one should hand you.