How to choose an AI consultant for a small or medium-sized business

The short answer: pick the consultant who asks about your business before they talk about AI. A good one starts with a specific problem that costs you money, builds inside the systems your team already uses, calculates every number you will act on with code instead of an AI model, tells you plainly where your data goes, and leaves you owning what they built. If a consultant can't explain those things in plain language, keep looking.

The rest of this guide walks through how to tell the difference, the questions to ask, the warning signs, what drives the cost, and how a good engagement runs from the first call to a tool your team actually uses.

What an AI consultant should actually do for you

An AI consultant is worth paying only if they solve a business problem. That sounds obvious, but a lot of AI work starts from the technology and goes looking for a use. You end up with a chatbot nobody asked for, or a demo that impressed everyone in the meeting and was never opened again.

A good consultant works the other way around. They start with what is costing you money or time, figure out whether software can fix it, and only then decide whether AI is part of the answer. Sometimes it is. Sometimes the right fix is plain code, a better report or a cleaner process, and an honest consultant will tell you that even when it means a smaller job for them.

For a small or medium-sized business, the useful work usually falls into a few areas:

• Finding where money leaks. Comparing what a job, a location or a product line should cost against what it did cost, and showing where the gap comes from.
• Cutting repeated manual work. The report someone rebuilds by hand every week, the data copied from one system into another, the follow-up that depends on someone remembering.
• Answering questions your data already holds. Which customers are worth the most, which products carry the business, where the schedule keeps slipping.
• Market and business intelligence. Reading what competitors are doing, summarizing what customers are saying, pulling the signal out of a lot of text.

Notice that most of these are about your numbers and your operations. AI is a tool for some of that work. It is not the point.

Before you call anyone, name the problem

The best thing you can do before talking to any consultant is write down the problem in a sentence. Not "we want to use AI." Something like:

• "Our jobs look profitable when we quote them and thin when we close them out, and we can't see why."
• "We spend most of Monday building the same report from the same spreadsheets."
• "We have customer data in our CRM, our accounting system and a pile of spreadsheets, and none of it talks to the rest."

A clear problem helps in several ways. It lets you compare consultants on the same question. It keeps the project from drifting into something bigger and more expensive. And it gives you a way to judge, at the end, whether the work did what it was supposed to do.

If you can't name the problem yet, that is fine. A good first conversation can help you find it. But be wary of anyone who skips that step and goes straight to a proposal.

Questions to ask an AI consultant

These are the questions that separate consultants who will build something useful from those who will build something impressive. For each one, here is what a good answer sounds like.

What problem would you solve first, and how would we know it worked?

A good answer names a specific problem in your business and a specific way to measure the result: a number that should move, a task that should take less time, a question you should be able to answer that you can't today. A weak answer talks about "transformation," "efficiency" or "unlocking insights" without saying what changes on a normal Tuesday.

Where will the tool live?

A good answer starts with the systems you already use, such as Microsoft 365 or Google Workspace, and explains when and why anything would run somewhere else. A weak answer assumes you will move onto their platform, or can't say clearly what runs where.

This matters more than it sounds. A tool your team has to log into somewhere new is a tool your team will stop using. A tool that shows up where they already work gets used.

How are the numbers calculated?

This is the most important question on the list, and the one most people never ask.

AI language models are good at reading, writing, summarizing and finding patterns in text. They are not reliable calculators. Ask the same model the same math question twice and you can get different answers, and a wrong answer looks exactly as confident as a right one.

A good answer is that every number you will act on, every total, variance and ranking, is calculated by ordinary code, and that tests check that code. AI may describe or explain the result, but it does not produce the figure. A weak answer is "the AI analyzes your data and tells you." If a number drives a decision about money, people or inventory, it should come out the same way every time you run it.

What happens to our data?

A good answer is specific. Which services will see your data, whether any outside AI service will store it or use it to train its models, and what agreements are in place to prevent that. If your data is sensitive, a good consultant can also run the AI on hardware you control, so it doesn't go to an outside AI service at all.

A weak answer is a reassurance with no detail behind it. "It's all secure" is not an answer. Ask what, specifically, makes it secure, and ask to see it in writing.

Who will actually build this?

In a small engagement, you want to talk to the people writing the code, not a salesperson who hands your project to someone you never meet. A good answer names the people and lets you talk to them. The person building your tool should understand your business before they write a line of it.

What do we own when you're done?

A good answer is that you own the tool, it runs in your accounts, and you are not locked into paying the consultant forever to keep it working. Ask what happens if you part ways. You should be able to keep using what you paid for.

What won't AI do well here?

This is a good test of honesty. Every competent consultant knows where AI falls short in your situation. If the answer is "nothing, it can do all of it," you are hearing a sales pitch.

How will our team actually use it, instead of going around it?

A good answer treats this as part of the job, not an afterthought. If the approved tool is slower or clumsier than the free app on someone's phone, people will use the phone. A good consultant builds the approved tool to feel as quick and easy as what your team would otherwise reach for, and pairs it with a clear policy. A weak answer is "we'll send everyone a policy."

Warning signs

Walk away, or at least slow down, if you see these:

• A proposal before a conversation about your business. If they can price it before they understand it, they are selling a package, not solving your problem.
• Promised dollar results before they have seen your data. Nobody can tell you what you will save until they have looked at how your business actually runs.
• AI producing the numbers. If the figures you would act on come out of a language model, they will be wrong some of the time, and you won't know which time.
• Vague answers about data. If they can't tell you exactly where your data goes, assume it goes further than you would like.
• Everything has to move to their platform. Some work does need its own platform, but that should be explained and justified, not assumed.
• You never meet the builder. Distance between you and the person writing the code is where misunderstandings turn into expensive rework.
• Buzzwords instead of specifics. If you finish the call and still can't say in a sentence what they would build, they haven't told you.
• A policy with nothing behind it. A rule that says "don't paste customer data into AI" does nothing on its own. If nothing enforces it and nothing better is offered, people will work around it.

Custom-built or off the shelf?

You can buy a suit off the rack and have it tailored. Or you can have the perfect bespoke suit built for you, the one that fits like nothing else and dominates the room.

Software works the same way. There are plenty of off-the-shelf AI products, and some of them are a good fit. If a product already does exactly what you need, inside the systems you already use, at a price that makes sense, buy it. A good consultant will tell you so.

The trouble is that most businesses run on off-the-rack software, altered until it almost fits. The product tracks most of what you need, in roughly the way you work, and your team fills the gap with spreadsheets and workarounds. That gap is often exactly where money leaks.

A custom build makes sense when:

• your problem depends on how your business specifically runs, not on a generic process;
• the data you need is spread across several systems that don't talk to each other;
• you have tried off-the-shelf tools and your team works around them;
• the numbers you need have to be calculated your way, not the way a product decided.

A custom build is the wrong choice when an existing product genuinely fits, or when the problem is small enough that a better spreadsheet would solve it.

Your accounts or their platform?

Where the tool runs affects how much your team uses it, who controls it, and what happens if you stop working with the consultant.

Built in your accounts. The tool runs inside your own Microsoft 365 or Google Workspace. Your team works where it already works, you control access, and the tool stays yours. This is the right starting point for most small and medium-sized businesses.

Built on the consultant's platform. Some jobs need more than your current systems can do: heavier calculations, more data, connections your tools don't support. In that case, a consultant may build on platforms of their own. That can be the right call, as long as they tell you exactly what runs where, who can see your data, and what happens if you part ways.

The question to ask is not "which is better," but "why this one, for this job?" A good consultant has a clear answer.

Keeping sensitive data safe

Many businesses hold data they can't afford to expose: customer records, financial details, employee information, pricing. Before any AI touches it, you should know:

• Which outside services will see it. Every AI service the tool calls is a place your data goes.
• Whether those services keep it or train on it. Many AI providers offer business agreements that turn off data retention and model training. Ask whether those agreements are in place, and ask to see them.
• Whether the AI can run on your own hardware. For the most sensitive work, the AI can run on hardware you control instead of a cloud AI service. It costs more to set up, but your data doesn't go to an outside AI provider.
• What gets logged. Logs are easy to forget, and they can hold more than anyone intended.

A consultant who answers these questions clearly, without being pushed, is one who has thought about them before you asked.

Your team is already using AI

Here is the uncomfortable part. Whether or not you have approved any AI tools, some of your people are probably already using them. They paste an email into ChatGPT to clean it up, drop a spreadsheet into a free AI app to summarize it, or ask a chatbot to draft a reply to a customer. Usually with good intentions, usually to save time, and usually without telling anyone. This is often called shadow AI.

The risk is not that your team is lazy. The risk is what goes into those tools. A customer's name, address and account details. Financial information. Pricing. Employee records. Once it is pasted into a free consumer AI app, you may have no control over where it goes, how long it is kept or whether it is used to train a model.

Banning AI rarely works. The tools are too useful and too easy to reach. What works is giving people something better:

• A clear AI use policy. What tools are approved, what data can and can't go into them, and why. Short enough that people actually read it.
• Systems that enforce it. Protection that catches personal and financial information before it reaches an AI service that hasn't been approved for it, and refuses rather than just warning.
• Approved tools that feel like the ones they would sneak. If the approved tool is as fast and easy as the free app, people use it. If it is slower, harder or locked behind extra steps, they go back to the phone in their pocket.
• The right agreements. Where an outside AI service is used, a business agreement with data retention and model training turned off.

When you talk to a consultant, ask how they handle this. A good one will have an answer that covers all of it: the policy, the protection and a tool your team actually wants to use.

What drives the cost

AI consulting prices vary widely, and a price quoted before anyone has looked at your business doesn't mean much. What actually drives the cost:

• How clear the problem is. A well-defined problem costs less to solve than a vague one, because less time goes into figuring out what to build.
• The state of your data. Clean data in one place is quick to work with. Data spread across systems, with gaps and inconsistencies, takes longer.
• How many systems are involved. Every system the tool has to read from or write to adds work.
• Where the tool runs. Building inside your existing accounts is usually simpler than standing up something new. Running AI on your own hardware adds setup cost.
• How much has to be calculated and checked. Numbers you act on need code and tests behind them. That is worth paying for, and it takes time.
• Ongoing support. Ask whether support is included, what it covers and what it costs after the build.

The useful question is not "what does AI consulting cost," but "what would it cost to solve this specific problem, and what is that problem costing me now?" If the leak is bigger than the fix, the work pays for itself.

How a good engagement runs

Every consultant works a little differently, but a good engagement for a small or medium-sized business usually follows this shape.

A conversation about the business

Before anything technical, the consultant learns how your business runs, where you think money or time is slipping, and what systems you use. You should leave this conversation feeling understood, not sold to.

A look at the data

The consultant finds out what your systems already record and what is missing. Sometimes the honest answer is that a problem can't be measured yet with the data you have. You want to hear that before you spend money, not after.

A clear plan

You should get a plain description of what will be built, where it will run, what it will and won't do, and how you will know it worked. If you can't explain the plan to your own team, it isn't clear enough.

The build

The tool is built inside your systems where possible. Numbers are calculated by code and checked by tests. You see progress along the way, not just at the end.

A check against your own records

Before you rely on the tool, the calculations should be tested against records you already trust. If the tool says a job lost money, you should be able to see why, and it should match what you know.

Handoff and use

Your team uses the tool where they already work. You own it. You know who to call if something needs to change.

A checklist for your first call

Bring this list to every consultant you talk to:

• Did they ask about my business before talking about AI?
• Can they name the problem they would solve first, and how we would measure it?
• Will it run in our existing systems, and if not, why not?
• Are the numbers calculated by code and checked by tests?
• Can they say exactly where our data goes and who can keep it?
• Can the AI run on our own hardware if the data is sensitive?
• Will I talk to the person who builds it?
• Will we own what they build?
• Did they tell me something AI can't do well here?
• Do they have a plan for the AI my team is already using without approval?
• Do I understand, in a sentence, what they would build?

If you can answer yes to most of these, you are probably talking to the right consultant.

Questions business owners ask

Do small businesses really need an AI consultant?

Not always. If an off-the-shelf product fits, or the problem is small enough for a better spreadsheet, you don't need a consultant. You need one when the problem depends on how your business specifically runs, or when your data is spread across systems that don't talk to each other.

Can AI calculate our numbers?

It shouldn't. AI language models are not reliable calculators. Totals, variances and rankings you act on should be calculated by code and checked by tests. AI is useful for describing results, reading text and market and business intelligence.

Is our data safe if we use AI?

It can be, if you ask the right questions. Find out which services see your data, whether they keep it or train on it, and whether the AI can run on your own hardware for sensitive work.

How do we stop employees from pasting customer data into ChatGPT?

Banning it rarely works. Write a short AI use policy, put systems in place that catch personal and financial information before it reaches an unapproved AI service, and give your team approved tools that are just as quick and easy to use, so they have no reason to go around them.

Should the tool run in our accounts or the consultant's?

Start in your own accounts, such as Microsoft 365 or Google Workspace, whenever possible. Build elsewhere only when the job needs it, and only when the consultant explains exactly what runs where.

How much does an AI consultant cost?

It depends on how clear the problem is, the state of your data, how many systems are involved and where the tool runs. Be cautious of a firm price quoted before anyone has looked at your business.

How do we start?

Write down the problem in a sentence, then talk to a few consultants using the checklist above.