When to Hire an AI Consultant, and When Not To

I sell AI consulting, so weigh this accordingly: most small businesses don't need a consultant to get started with AI. A ChatGPT or Claude subscription and a free afternoon will take you further than most owners expect.

Some do need help, though, and waiting too long has a cost of its own. The trick is telling which situation you're in before you spend money, or before you spend another quarter not getting to it. Here's how I'd decide.

The gap between using AI and running on it

Early this year, Goldman Sachs surveyed 1,256 owners in its 10,000 Small Businesses program. 76% said they were already using AI, and only 14% said they had fully integrated it into core operations.

That gap is the whole question. Three out of four owners have typed something into a chat window. About one in seven has AI doing a real piece of the company's work every day, without someone remembering to open a tab.

The same survey asked the owners who use AI what was getting in the way. About half named each of three things: data privacy and security concerns (50%), a lack of technical expertise (49%), and difficulty choosing the right tools (48%).

Getting into the 76% takes no outside help. Getting into the 14% sometimes does. So the first thing to work out is which side of that gap your problem is on.

When you don't need a consultant

The task lives in one person's hands. Drafting emails, writing quotes, summarizing a long contract, cleaning up a customer list, turning notes into a procedure. If one person can do it by pasting something into a chat window and reading what comes back, nobody needs to be hired. I wrote up a quote template that takes about twenty minutes to set up, and it's a fair picture of how far you can get alone.

You don't have a specific problem yet. "We should be doing something with AI" isn't a project. Hire someone at that stage and you're paying them to find a problem, and they'll find one whether or not it's your biggest. Spend a week on a tally sheet first: what gets done over and over, and how many minutes it takes. The five tasks I'd check first are a reasonable list to tally against.

Nobody can say how the job is done today. If two people in the office would describe the same process two different ways, automation will lock in whichever version happens to get built. Settle the process on paper first. That's a conversation between you and your staff, and it's free.

The task is rare. Something that takes two hours once a quarter is eight hours a year. It isn't worth anyone's project fee, mine included.

Four signs it's time to bring someone in

1. The work crosses systems. A chat window can write a follow-up email. It can't notice that a quote went out nine days ago with no reply, look up the customer in your job software, send the follow-up from your address, and log that it did. Once a task has to move between your inbox, your scheduling tool and your accounting system on its own, you're dealing with connections, permissions, and what happens when something fails at 2 a.m. This is where doing it yourself usually runs out, and I suspect it accounts for a lot of that 48% who have trouble choosing tools.

2. You've stalled. You tried it twice and it half works. Or it's been on your list since summer. A project that depends on your spare evenings competes with everything else that does. A stall isn't a character flaw. It is a cost, though, and the next section puts a number on it.

3. A mistake would be expensive. A draft that a person reads before sending is low risk. Anything that sends, bills, schedules or touches customer records with nobody looking is not. Half the owners in that survey named data privacy and security as a barrier, and they're right to. If you can't say where your customer data goes when the automation runs, get someone who can.

4. The math is bigger than the fee. This is the one that should decide it.

The math

Take a task that eats six hours a week across your office. Say it's chasing open quotes and unpaid invoices. At a loaded cost of $30 an hour, that's $180 a week, or $9,000 a year at 50 working weeks.

Now three cases.

If you can set it up yourself over a weekend, do that. You keep the whole $9,000 and you learn how it works.

If you've been meaning to set it up for a quarter, the delay has already cost 13 weeks at $180, which is $2,340. Another quarter makes it $4,680. At some point the waiting costs more than the help would have.

And if the same kind of task takes forty minutes a week, it's worth about $1,000 a year. Very few outside projects pay for themselves on that. Do it yourself or leave it alone.

Those numbers are invented, and yours will differ. The method holds: hours per week, times what the hours cost, set against a fixed quote for the work. If a consultant can't give you a fixed quote for one workflow, that tells you something too.

What the research says about building it yourself

There's one piece of evidence worth knowing, limits included. In 2025, researchers at MIT's Project NANDA reviewed more than 300 public AI projects and interviewed people at 52 organizations. In that interview sample, tools brought in through an outside partner reached full deployment about 67% of the time. Tools built internally got there about 33% of the time.

The caveats matter. These were bigger companies than yours, the outcomes were self-reported, and the authors say plainly that the correlation doesn't prove the partner caused the success.

What I take from it is narrower than "hire a consultant." The report's explanation for the failures was that the tools didn't fit day-to-day work: brittle workflows, and nothing that adjusted to how the job was really done. I read that as a warning against building custom software from scratch when something proven can be fitted to your operation. I'd read it as a warning about consultants, too: the ones who deliver a tool and leave before it fits.

A 30-minute test

Before you call anyone:

1. Write down the task, who does it, how many hours a week it takes, and which systems it touches.
2. Spend 30 minutes trying the core of it in ChatGPT or Claude, using a real example from last week.
3. Notice where you stop.

If you got most of the way there, keep going alone. You're on the easy side of the gap.

If you stopped because the model couldn't reach your data, because the result has to land in another system, or because you wouldn't trust it to run unwatched, you've just written the scope for outside help. It's one workflow with a known number of hours attached. That's a much better starting point than "help us with AI."

If you do hire

Keep it to one workflow, a fixed price, and one number you'll both check a month after it goes live. Make sure the accounts and automations are in your company's name when it's done. Walk away from anyone whose first deliverable is a strategy deck.

If your 30-minute test ended at a wall, that's what a fit call is for. You describe the task and where you got stuck, and I tell you whether it's a project, a setting you missed, or something to leave alone. Sometimes the answer is "not yet," and that's a fine answer. If you'd rather have the whole operation looked at before picking a workflow, the AI Audit does that for $995.

Either way, run the test first. It costs half an hour, and it tells you which side of the gap you're on.