AI Consultant for Small Business: What to Ask

A useful AI consultant for a small business can name the job, the data, and the person who is allowed to stop the work. If a ready-made product already does that job, the honest advice is to buy the product and skip the custom project. We would rather lose the work than invent a build to justify a fee. That is how the studio works. If you are torn between a consultant and a firm that wants to own the install, read the agency comparison. If a pretty first test run never became a real process, read the stall note.

Cited in this note

40%

Bring the person who does the job into the first meeting. An owner alone will be sold a platform. The office manager, the bookkeeper, or the dispatcher will ask where the file lives. That question is worth more than any slide with a diagram. If the consultant looks annoyed by it, you have learned something cheap.

We feel strongly about this because we have watched the other way go wrong. A pitch that starts with a tool name has skipped the work. You can always get back to a tool later. You cannot easily get back the month you spent building the wrong promise.

What gets canceled

Gartner said in June 2025 that over 40% of AI agent projects will be canceled by the end of 2027. It named cost, unclear business value, and weak risk controls as the reasons.¹ In a poll of 3,412 people, 19% said their group had made big investments in AI agents. Another 42% had made small ones, 8% had made none, and 31% were waiting or unsure. Gartner also said only about 130 vendors sold what it would call a real AI agent. Think about that last line before you let a proposal use the word agent as decoration. The label is everywhere, but the real thing is rare.

McKinsey’s 2025 survey shows the same gap in other numbers. About 88% say they use AI regularly in at least one part of the business, up from 78% a year earlier.² Only 23% are growing an AI agent system somewhere, and 39% are still trying it out. A consultant who quotes only the first number is describing use and not a finished system. Ask them to say the second pair of numbers out loud. If they cannot, they are selling the feeling of being current.

The red flags are easy to spot once you know them. Guaranteed rankings go against Google’s own guide. It says the new AI features still depend on ordinary search work.³ Nobody can promise a spot in an AI answer. A person who does will probably bluff on the rest of the job too. If there is no data plan, nobody has named the one place the facts live. If there is no handover, your team cannot run the thing after the last invoice. We treat those three as reasons to end the meeting. We do not treat them as issues to talk through.

A worked example

Sit in on a hiring talk at a small dental clinic. This is a walk-through we use, and it is not a scoreboard. The owner is a dentist who wants help. The office manager runs the reminders and the front desk, and she is in the room because we asked her to be. The bookkeeper is not there yet. She will matter when the invoice arrives, and she should see the proposal before anyone signs.

The first consultant opens a laptop and names a platform. He talks about AI that will “handle the whole patient journey.” He has not asked where the appointment book lives. He has not asked about the paper list of people who do not want texts. When the office manager asks who checks a message before it goes out, he says the system has safety rails. A safety rail is not a name. She writes nothing down, and that is the right instinct.

The second consultant does not open a tool. She asks the office manager to describe last Monday, including the part that went wrong. A reminder went to a patient who was already in the chair, because the reminder list and the treatment list do not talk to each other. The consultant writes one sentence. It says to draft cleaning reminders from the appointment record only, and the office manager sends them. She asks where the do-not-contact names are kept. She asks what happens when the manager is out sick. She also says, out loud, that if a ready-made product already drafts that reminder inside the clinic’s software, they should buy it and stop talking to her.

We would hire the second person, and we would still make her put a stop rule in the proposal. The work goes on only if the manager can clear the queue in the time she has set aside. It stops if the drafts keep picking the wrong patients. A consultant who will not write the stop rule is planning to stay forever. That is how a small fee turns into the kind of project Gartner expects many AI agent efforts to cancel.

Notice who did the real interview. It was the office manager, and not the owner. The owner felt the pain, but the manager knows the drawer with the paper list. A good consultant talks to the owner about money and risk. Then she talks to the manager about the Monday. A weak one charms the owner and avoids the manager, since the manager asks questions that shrink the proposal. Let it shrink. A smaller job that is true beats a big job that is foggy.

How we’d run it

We run a hiring talk like a filter. The consultant is not owed a second meeting.

First, we send our questions ahead in writing, so nobody can say they were ambushed. What job will change? Where does the data live? Who checks the answer, by name? How will we know it failed? What will you refuse to build? If the reply is a calendar link and a slide deck with none of those answers, we skip the meeting. Being polite is not a reason to give away an hour.

Second, we hold the meeting with the person who does the work, and not only the owner. We ask the consultant to say the job in one sentence that the staff could repeat. If the sentence has a tool name in it and no verb the staff would use, we send them back. “Roll out the platform” is not a job. “Draft the cleaning reminder and wait for a person” is a job.

Third, we ask the buy-or-build question straight out, and we watch the face. If a ready-made product already does this, will you say so? The honest answer is yes. It comes with a short list of what they could still help with, such as the wording and the checking step. The dishonest answer is a custom build that copies a feature the clinic’s software already sells. It pads the consultant’s hours and leaves the manager with two systems to look after.

Fourth, we ask for a handover plan before we talk about price. Who can run this when you leave? Where are the rejected drafts written down? What will the bookkeeper see on the invoice, in words and not only in a total? A price with no handover is really a subscription to one person. That can be fine for advice, but it is a bad way to buy a system. The agency comparison is the page we use when an owner is unsure which kind of help they even want.

Fifth, we check the proposal itself for red flags. Any promise of guaranteed rankings gets a no. Google’s guide says the new AI features still depend on ordinary search work. Any slide that quotes use without the gap between use and finished systems gets a question. Any plan that cannot name a risk control gets a no. That is one of the reasons Gartner gave for cancellations. We would rather be called difficult in the first meeting than be stuck in a project that gets canceled after the staff have rebuilt their week around it.

Common mistakes

The most common mistake is hiring the best talker, and it is a very human one. The owner is busy, and a clear voice feels like skill. In this work, skill sounds slower. It sounds like questions about a drawer full of paper. If you leave a meeting excited and unable to repeat the job in one sentence, you were entertained.

A second mistake is leaving the person who runs the work out of the room. The office manager will inherit the queue. If she first meets the system on launch day, she will find ways around it, and she will be right. Invite her while the proposal can still change. If the consultant objects, that objection tells you what you need to know.

A third mistake is accepting “agent” as a product type, which is how the poll gets ignored. In that Gartner poll of 3,412 people, 19% had made big investments in AI agents, 42% had made small ones, 8% had made none, and 31% were waiting or unsure. Only about 130 vendors met Gartner’s bar for a real AI agent product. A proposal that calls a reminder draft an agent is borrowing a word. Ask what the software does on a Monday. If the answer is that it drafts and a person sends, call it a draft. The plainer word will get you a plainer price.

A fourth mistake is the search fantasy, and it can arrive even when you never asked for marketing help. A promise of a top spot in an AI answer tells you the other claims bend too. Google’s guide is public. A person who sells certainty against a public guide will sell certainty against your files as well.

The last trap is signing because the quote is small. A small quote with no stop rule has a way of becoming a long one. Cost is one of the reasons these projects get canceled. Get the stop rule in writing while the number still feels harmless.

What we’d do

We would tell many owners to buy the ready-made product and send the consultant home. That opinion costs us work, and we still hold it. Custom work makes sense when the job depends on your own files and no product will touch them. It also makes sense when the value is the thinking and the handover, and not the code. It does not make sense just because a platform demo was impressive and the office manager was not there to object.

We would also bring the awkward numbers into the proposal review, on purpose. Over 40% of AI agent projects canceled by the end of 2027 is not a reason to avoid all help. It is a reason to ask for a value you can point at and a risk control with a name. The McKinsey gap belongs in the same talk. The consultant will happily say that 88% use AI in at least one part of the business, up from 78% a year earlier. You should be the one to say that 23% are growing an AI agent system and 39% are still trying it out. If they only brought the first number, they are describing a hobby and pricing it like a system.

If the proposal mentions search, rankings, or “being the answer” in an AI result, we stop and open Google’s guide with them.³ The new AI features still depend on ordinary search work, and no serious person promises the outcome. And if the first test run they show you could not survive a normal week, read the stall note before you pay for a sequel. The studio will tell you to buy the product when that is the truth. We think you should ask anyone else you pay to say the same thing.