"What does an AI consultant do?" Owners ask it after a board meeting where somebody said "get someone in", or after a vendor's demonstration that was impressive and answered nothing. The honest answer is that a good one spends most of the engagement listening and reading, writes down what the company should not build, names the one or two people inside who should own the subject, puts a number and a date on the first thing to try, and leaves a written plan the board can read in twenty minutes. This article is for the owner or chief executive of a company with fifty to five hundred people, often with a private equity sponsor or a family on the board, who has been told to find help and wants to know what the help is supposed to look like. Where money is handled it is not accounting or tax advice.
I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them. Every one of those companies had been advised by somebody, and the advice that lasted was always the same shape: short, written, specific about who did what next, and silent about things the company did not need. I am a lawyer by training. Now I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies through 013 Labs, so this is the work described from the inside, including the parts a buyer should refuse.
The short answer
An AI consultant is paid to turn a vague instruction, "we should be doing something about AI", into a decision the company can act on and a board can read. The product is a written plan, not software and not hours. Most of the work is finding out what is already true: who uses these tools quietly and what they paste into them, which processes run on paper and memory, and where the data actually lives. The technology is the smallest part; the judgement is in what to leave out. If the first conversation starts with tools, you are talking to a seller; how to choose an AI strategy consultant covers how to tell before you sign.
What a consultant does, in order
Listens before proposing anything. The first days are interviews: the owner, the finance lead, the operations lead, the people who answer the phone, the person in the warehouse who built the spreadsheet everyone uses. The question is not "what do you want AI to do" but "what takes your week, and what breaks".
Goes through the company's own material. Process documents, if any; the systems list; the last three board packs; the contracts that say who owns which data; the reviews customers leave. A consultant who has not seen the board pack cannot write for the board.
Finds the quiet users. In almost every company of this size, a handful of people already use these tools, and they know more about what works for this company than any outsider. A good consultant finds them in the first week and treats them as the first draft of the answer. Spot your AI champion is the one-page version of how to find them.
Writes down what not to build. This is the half that earns the fee. A list of the things the company will not do this year, with a reason each: not a chatbot on the website, not a rewrite of the quoting system, nothing that touches payment data until the data rule exists. A plan without a never-list is a wish list.
Names who owns it inside. One or two people, by name, with a sentence each on what they decide, what they check and who they tell. Not a committee and not the consultant. Who owns AI in a company covers why the owner is rarely the right person.
Writes the data rule. One page: which tools are approved, under which terms, what may go into them, what never may, and who may say yes to a new one. Short enough to be followed, specific enough to be enforced.
Puts a number and a date on the first thing to try. One process, one measure, one date on which a person will say whether it worked, and a stop rule written before it starts.
Leaves a plan the board can read in twenty minutes. Plain language, no vendor names where none is needed, the never-list, the owners, the data rule, the first test, and the three questions the board should ask at its next meeting. If it takes a second meeting to explain the plan, the plan is not finished.
What a consultant should never do
Sell software they are paid to recommend. If a tool is proposed, ask who pays the consultant when you buy it. A referral fee turns an adviser into a channel. Evaluating AI vendor claims covers the questions to put to any product that is recommended.
Run a pilot nobody asked for. A "quick proof of concept" that appears in week two, built on the consultant's own account with your data, is scope creep with your customer records inside it.
Promise savings before the data has been seen. Any number given before the interviews and the reading are done is a sales figure. The honest version is a measure and a date, not a percentage.
Take the logins. The consultant works in your accounts, under a named user you can remove, or works from copies you provide. Nobody outside the company holds the only key to anything.
Stay forever. The engagement has an end written into it, and the people named in the plan can run it without the author. A consultant whose plan needs the consultant has written a retainer, not a plan.
The data rule
Before any outsider sees a customer list, a payroll file or a contract, write the sentence that governs it: what the consultant may see, where it may be stored, that it is deleted or returned at the end, and that nothing is pasted into a public tool to speed up the work. A consultant who resists that sentence has told you how they treat data. What happens to your data explains what to ask any provider before the first file moves, and the same questions apply to a person.
Five checks before you sign
First, ask for a sample plan with the names removed. Twenty minutes with it will tell you whether this person writes for a board or for an engineer. Mostly diagrams of systems means the wrong plan for an owner.
Second, ask what they will tell you not to do. A good consultant can answer in the first conversation, from what you have already said. A seller will say it depends.
Third, ask who pays them when you buy a tool. The answer is either "nobody but you" or a reason to stop.
Fourth, ask who inside the company will own the plan when they leave. If the answer is "we can discuss ongoing support", the plan has no owner yet.
Fifth, agree the end before the start. A date, a deliverable, and what you will be able to do on your own the day after.
Check yourself
Seven yes or no questions. Each no is something to settle before anyone is engaged.
- Could you name the people in the company who already use these tools, and what they use them for?
- Do you know where the company's customer, pricing and employee data actually lives today?
- Is there one person inside who would own an AI plan if you handed it to them tomorrow?
- Have you written down what the company will not do with AI this year?
- Would your board accept a plan with no vendor named in it?
- Could you state the one process you would test first, and how you would know it worked?
- Do you know what you want to be able to do without the consultant on the day they leave?
Mostly no? That is normal, and it is what the first weeks of a good engagement are for. The AI readiness scorecard asks ten questions like these across the whole company.
What it costs, in plain terms
There are two honest shapes. The first is a fixed fee for a fixed deliverable: a written plan by a date, with the interviews and the reading included, and nothing sold alongside it. The second is hourly or monthly work, right for a company that already has a plan and needs a hand. Hourly work should make an owner ask one question before anything else: what does the end look like, and who decides when it has arrived. If the answer is vague, the cost is open. Not accounting or tax advice.
How to tell after thirty days whether it was worth it
Thirty days after the plan lands, five things should be true. The never-list exists and somebody has used it to say no to something. The named owners know they are named and have met once without the consultant. The data rule is written, short, and the quiet users know it. The first test has started, with its measure and its stop rule written down before the start date. And the board has had the plan in front of it and asked at least one of the three questions. If none of those is true, the engagement produced a document, not a change, and that is worth saying so.
Where it goes wrong
The platform in disguise. The strategy arrived with a licence agreement attached; the plan was a sales deck for a partner product, and nothing in it survived the first renewal.
The pilot with the customer list in it. A proof of concept built in week two on an outside account with real customer records. It showed what was possible, and it put the customer list where nobody in the company could reach or delete it.
The plan nobody owned. Sixty pages, well written, delivered to the owner, who filed it. No name inside the company appeared in it except the owner's, and six months later the board asked for the AI plan again.
The savings promised in the proposal. A figure quoted before anyone had seen the data became the number the board remembered, and every honest finding afterwards was measured against a guess.
Starting safely
Start by writing down, before any first call, the one thing you want to be able to decide at the end: which process to test first, who owns the subject, or whether to buy a tool a vendor is pushing. Ask for a sample plan. Agree the data rule and the end date in writing. Give the consultant the quiet users and the board packs in the first week, then let the work be mostly listening. Picking your first AI project covers how the first test is chosen, and Can AI replace an employee covers the question most boards actually mean when they ask about savings.
Picture it going wrong
A made-up example, to show the shape. A family-owned packaging distributor with about a hundred and fifty people and a minority investor on the board. The investor asks for an AI plan. The owner engages a firm that arrives with a team of four and a partner product that will "unify the data". In the second week the firm builds a demonstration on its own account using a year of customer orders. In the sixth week it presents a plan: a platform subscription, a phased rollout, a monthly retainer to manage it. The board approves it because it is the only plan in the room. A year later the platform is half connected, the customer order history sits in an account the company does not control, and the person in the warehouse who had quietly built the only working forecast has left. Nobody lied. The company bought software and hours when it needed a decision. The fix: a short written plan by someone with nothing to sell, the order history retrieved and the outside account closed, two people inside named as owners, the retainer ended, and one test, with a measure and a date, on the forecast the warehouse had already proved.
Mistakes that repeat
- Engaging a consultant before writing down what you want to be able to decide at the end.
- Accepting a proposal that names a tool before anyone has interviewed your people.
- Letting a proof of concept be built on an outside account with real company data.
- Taking a savings figure quoted before the reading was done as the measure of success.
- Receiving a plan with no name inside the company in it, and filing it.
- Paying hourly or monthly without a written answer to what the end looks like. Not accounting or tax advice.
Frequently asked questions
What does an AI consultant do?
A good one interviews your people and reads your material, finds the people already using these tools quietly, writes down what the company should not build, names who inside owns the subject, writes the data rule, puts a number and a date on the first test, and leaves a plan the board can read in twenty minutes.
What does an AI strategy consultant do differently from an AI developer?
A strategy consultant decides with you what should exist and who owns it; a developer builds what has been decided. If one person offers both, ask which you are paying for first, because the plan should not depend on the builder being hired.
What does an AI advisor do for a small or mid-sized company?
For a company of fifty to five hundred people the advisor's job is mostly subtraction: ruling out the projects that would consume the year, finding the one that fits, and making sure a named person inside can run it afterwards.
How long should an engagement take?
Long enough to interview the right people and go through the material, short enough that the plan arrives before the next board meeting. Weeks, not quarters.
Should the consultant build the first project?
Usually not. The first test should be small enough for your own people to run with the tool they already use; that is how you learn whether the plan works without the author in the room. Outside building, if needed, is priced separately, after the plan.
How do I know whether the advice is independent?
Ask who pays them when you buy a tool, ask for the never-list, and ask for a sample plan. Independent advice names fewer products, says no more often, and ends.
Where should I start?
Write down the decision you want to be able to make at the end, find the people already using these tools inside the company, and ask any consultant you speak to for a sample plan and their never-list before you discuss a fee.
Not sure where you stand? Take the 3-minute AI readiness scorecard.