AI Automation Agency vs Consultant: Who to Hire

Hire a consultant to name the job, and hire an agency only after that naming is done. An AI consultant names the job, the data, and the stop. An automation agency builds the system once those three are written down. When the agency comes first, the build has no owner. We’ve watched that happen in the wrong order more than once. The proposal arrives with staff, a toolset, and a start date. Nobody in the room can say which output a person still has to check. The build goes ahead anyway, because the calendar was the only decision anyone made.

Cited in this note

25%

The consultant note is the list of questions we wish owners had before the first sales call. The buy-or-build note covers whether to build at all. The decision guide gives you a matching score when the choice itself is the thing you keep putting off. Hiring is a decision, so score it before you sign. A signature on a vague scope is not a plan. It is a hope with a payment schedule.

Two training figures, two populations

Microsoft’s 2024 index says only 25% of companies were planning to offer generative AI training.¹ Statistics Canada’s 2025 report says 38.9% of Canadian businesses that already use AI had trained staff.² Those are different groups of people, so do not average them. One is a plan among companies. The other is training that already happened at businesses using AI. A pitch that blends the two into “about a third of firms train people” has invented a fact. Planning and doing are not the same thing. Companies and AI-using businesses are not the same group. The two surveys also cover different countries and different years. If a consultant or an agency treats them as one number, they are not reading the reports. They are decorating a slide.

McKinsey’s 2025 survey says 23% are rolling out an AI agent system somewhere, and 39% are still experimenting.³ An agency pitch that quotes only a usage figure is selling the experiment. Experimenting is the larger group, and rolling out somewhere is the smaller one. The word “somewhere” matters here. It does not mean the job you care about or a firm the size of yours. We bring this up because the word agent shows up in proposals early. Nobody admits they are still testing. An experiment can be the right buy if it is labeled as one, with a stop and a person. An experiment sold as a finished operation is how an owner ends up maintaining a demo.

Pay a consultant to refuse work that a product already does. Pay an agency when the process is written and the checker has a name. The refusal is the valuable part, and it is the part a build contract is bad at. A builder is paid to build, but a consultant can be paid to say don’t. If you only hire people who get paid when the project starts, you will get a project. You may not get a job worth doing.

A worked example

Take a clinic that almost hired an agency to put an AI agent on the front desk. This is a walkthrough, not a scoreboard. The owner had three calls in one week and a folder of proposals. The front desk still books by phone and by a form on the site. The practitioner still decides which visits are in scope. Billing still waits for a note that the visit happened. Ask what the agent would be allowed to send and the room goes quiet. Ask what happens when the draft is wrong and the room gets quieter. That quiet is the consultant’s whole job. It is to name the work, the data it may see, and the stop.

In that room we would send them to a consultant before any build. The consultant’s output is a page, not a login. The page might say that the front desk drafts reminders and a person sends them. It might say the practitioner is the stop whenever a reminder includes a clinical instruction. It might say billing is not part of this job at all. It might also say the booking product they already pay for sends a plain reminder. Then they should not build one. That refusal is worth more than a sprint of work. An agency that starts before the page exists will build the reminder anyway. A reminder is easy to demo. Easy to demo does not mean it was missing.

A trades desk is the other shape, and it is where an agency can be the right call. The quote steps are already written. The office manager logs the address, and the lead tech files measurements before leaving the driveway. The owner sets the price, and that price is the stop. The bookkeeper does not open an invoice until the yes is written down. A product they tried does not keep site photos and supplier notes in one place. The checker has a name, so this is a real build. Even then we would want the agency to repeat the stop back to us. They should do it before they give an estimate. If they talk about agents and never say the owner’s name next to the price, they are selling an experiment. It is the same one that McKinsey’s larger group is still running.

A small shop is usually neither a big consulting story nor a custom build. The owner wants someone to do AI because a neighbor did. The clerk already knows the reorder, and the bookkeeper already keeps a late-invoice list. A consultant who earns the fee will point at one of those jobs, refuse the rest, and go home. An agency will only earn a fee here if one job truly needs a system that no product covers. Most weeks, it does not. Paying either of them to rediscover the clerk’s list is a polite way to avoid deciding.

How we’d run it

First, write the job in one sentence that a new hire could read. Do not write “AI for the front desk.” Write “The front desk sends appointment reminders after a person reads them.” If you can’t get to a sentence like that, you are hiring a mood. Moods get expensive when they come with a team.

Second, write down the data the job may touch and the data it may not. Reminders can see the time and the place. They can’t see the clinical note unless a named practitioner says so. Quotes can see the site measurements, but they can’t invent a price. If the data line says “whatever the model needs,” you don’t have a boundary. You have a hope that the vendor will be careful. Vendors are careful about their own scope, so you have to be careful about yours.

Third, name the checker in the same sentence as the job. That means a person, a place where the output waits, and what happens if it is wrong. A checkpoint nobody owns is not a checkpoint. We would rather delay the hire than sign a scope where “someone will review it” is the only human. Someone is not a name, and when everyone is responsible, everyone goes home.

Fourth, ask whether a product already does this work. The consultant is paid to refuse it. If the booking tool already sends the plain reminder, stop there. If the books already flag the late invoice, stop there too. Bring in an agency only for the gap that remains after that refusal. The buy-or-build note is the longer version of this step. Don’t skip it because an agency proposal looks nice.

Fifth, keep the training story apart from the build story. The 25% figure is companies planning to offer training. The 38.9% figure is AI-using businesses that had trained staff. Neither one promises that your clerk will be shown the new way of working. Put the training in the scope. A launch with nobody to run the job next Tuesday just moves the work back into the owner’s memory. That is where it already lived.

Sixth, if the proposal mentions agents, ask whether they are rolling out or experimenting. Make them pick one. The 23% and 39% figures are not invitations. They are a warning about language. If the honest answer is experiment, write “experiment” at the top of the scope. Add a date when you will decide to keep it or turn it off. If they can’t say, you are buying a word.

Common mistakes

Hiring the agency because the proposal had more pages is the mistake with the nicest paper. Pages are not a named job, and a bench of people is not a checker. We’ve sat with owners who chose the thicker folder because it felt serious. Seriousness is a person who will tell you not to start. The thinner conversation is often the one that saves the quarter, and we don’t mean a cheap one. We mean a conversation that is allowed to end in no.

Averaging the training figures is the mistake that makes everyone sound informed. Someone says firms don’t train. Someone else quotes the other survey as if it were the same study, and the room nods. They are not the same study. One is a plan in a 2024 index of companies. The other is training already done among Canadian businesses that use AI, in a 2025 report. Use either number alone with its group, or use neither. The fake average will be repeated to your staff as if it were a law. It is not one.

Buying the experiment under a rollout headline is how agent talk gets into a clinic that only needed a reminder. McKinsey’s split helps if you say it out loud in the meeting. Most of the people in that survey were not rolling out an agent system anywhere. If your seller quotes only a usage number, ask what happened to the people who were still experimenting. Then ask which group the proposal belongs to. Silence is an answer, so treat it as one.

Leaving the refusal out of the contract is how you pay twice. You pay the agency to build the thing the product already did. Then you pay the product to keep doing it. A consultant who won’t refuse is a salesperson with a notebook. We would rather pay for a no on a single page. A yes that copies the booking tool costs more. Put the right to stop in writing before the build gets a channel and a weekly call.

Skipping the owner of the output is the last mistake, and it survives all the others. The scope names a platform but does not name the office manager. On the first real day the draft is ready and nobody’s job includes looking at it. The agency did what it was paid to do. The consultant, if you had one, either wasn’t heard or wasn’t asked. The draft waits in a queue that counts as a success on paper. Customers never see the success. They see whatever got out.

What we’d do

We would hire the namer first, almost every time. We would pay them to be willing to end the work with a refusal. That is our opinion. A clinic that can’t say what the front desk may send doesn’t need an agency yet. A shop whose late invoices already live in the bookkeeping product doesn’t need either one. A trades desk is the rare room where we would call a builder. The steps are written, a named person owns the price, and no product covers the gap. Even there, the builder does not get to rename the job.

We would not let a single training statistic into a proposal without its group sitting next to it. We would not let “agent” into the scope unless someone can show the step that branches. We would also want to meet the person who stops it. We would not sign a start date that arrives before the page. Dates are easy, but pages are the work. If the seller needs the date to keep their staff busy, that is their problem. They can solve it after you know what you are buying.

If you only do one thing before the next call, write the sentence with the job and the checker. Then ask the seller to repeat it without adding a platform. If they can’t, you are not late. You are early, which is the right time to be early. The experiment will still be there next month. Your name on a vague scope is much harder to take back.