How to Hire for AI Skills in a Small Business (Without Hiring an AI Specialist)

The advice going round small-business circles right now is that you should hire an "AI person". A prompt engineer, an AI lead, someone whose job is to work out what all this means for you. For a business of eight people paying £45,000 for that role, it is almost always the wrong move — you end up with one person who understands AI and seven who still do not.

The better framing is this: you are not hiring for AI. You are hiring people who can work out how to do their own job faster, and AI is currently the biggest lever available for that. That changes what you write in the advert, what you ask in the interview, and what you do with the team you already have.

You probably do not need an AI specialist

A dedicated AI hire makes sense in a narrow set of cases, and small businesses almost never fall into them. The honest test is whether you have a technical AI problem or an adoption problem.

A technical problem looks like: you are building AI into a product customers pay for, you are fine-tuning or hosting models, or you are processing data at a volume where the engineering is genuinely hard. That needs an engineer.

An adoption problem looks like: "we've got four tools, and nobody knows what any of them actually do". That is the situation in most SMBs, and hiring a specialist does not fix it. Adoption problems are solved by the people doing the work changing how they work — which no single hire can do on their behalf.

There is also a simple economics argument. A specialist at £45,000–£60,000 costs roughly the same as giving forty people a paid AI licence for six years. If the constraint is capability across the team, the licence plus structured training wins every time. Our guide on how to train your team to use AI covers the training half of that.

What "AI skills" actually means for a small team

"AI skills" is a useless phrase in a job advert because every candidate will claim it. What you are actually screening for breaks into four things, and only the first is about the tools themselves.

• Task decomposition. Can they look at a piece of their own work and see which 30% of it is pattern-matching a machine could draft? This is the skill that matters most and it has nothing to do with knowing which model is best.
• Verification instinct. Do they check the output? A candidate who says "I use AI for research" without mentioning how they confirm a claim is a liability, not an asset.
• Judgement about what to keep human. Knowing that the client apology email, the redundancy conversation and the pricing negotiation are not AI tasks is a skill, not a limitation.
• Tool fluency. Actual familiarity with a general assistant and whatever is standard in their function. Genuinely the easiest of the four to teach, and the one most job adverts over-weight.

Notice that three of the four are judgement, not software. That is deliberate. Tool fluency has a half-life of about eighteen months; the other three compound.

Rewriting the job description

Most SMB job adverts have gone one of two unhelpful ways: either no mention of AI at all, or a bolted-on line reading "experience with AI tools a plus", which screens for nobody.

Replace that line with something that describes the work. For a marketing executive role:

You will be expected to use AI tools as a normal part of the job — drafting first versions, summarising research, analysing campaign data. We will give you paid licences and time to learn them. What we care about is that you can tell the difference between a good output and a confident-sounding wrong one, and that you know which parts of the job should never be automated.

That paragraph does four jobs at once. It sets the expectation, removes the fear that the role is about to be automated away, signals you will actually pay for tools rather than leaving people on free tiers, and tells the candidate you value verification. It also quietly filters out anyone whose position is that AI is beneath them or that it does everything.

Three things to avoid in the advert:

• Naming a specific tool as a requirement unless it genuinely is one. "Must have ChatGPT experience" narrows your pool for a skill you can teach in a fortnight.
• "AI-first" as a culture claim. Candidates read it as "we are planning redundancies".
• Asking for "prompt engineering" by name. It attracts people who have done a course, not people who have done the work.

How to test for AI skill in an interview

Do not ask "do you use AI?" — everyone says yes. Ask questions that only someone who has actually done it can answer specifically. Four that work:

1. "Walk me through the last thing you used AI for at work. What did you ask, and what did you have to fix?" The fix is the whole question. Someone who has genuinely used it will describe a rewrite, a wrong figure, a tone that was off. Someone who has not will stay vague.

2. "Where in your last role would you have refused to use it?" You are testing for the judgement in point three above. Good answers are specific: anything going to a regulator, anything where a client's numbers would be pasted into a third-party tool, anything where being wrong is expensive.

3. Give them a real output and ask them to critique it. Take an AI-drafted version of something from your own business — a product description, a client email, a summary — with a plausible error left in. Ask what they would change before sending it. Ten minutes, and it separates verification instinct from tool familiarity better than any other question.

4. "If we gave you an AI budget of £100 a month for your own role, what would you spend it on and why?" This tests whether they think in terms of outcomes or subscriptions. "Six AI subscriptions, no coherent workflow" is the failure mode you are screening against.

Two practical notes. First, expect AI-assisted applications — most CVs and cover letters now are, and treating that as disqualifying will cost you good candidates. Screen on the interview, not the paperwork. Second, if you are using AI to filter applicants yourself, read our guide on using AI for hiring and candidate screening first, because the discrimination risk there is real and the EU AI Act treats recruitment as high-risk.

What to do with the roles you already have

Hiring is the slow lever. Most of the capability you need is already sitting in your existing team, usually unevenly distributed and usually invisible to you — half of them are already using tools you never approved, which is the shadow AI problem.

A workable sequence for a team of five to fifty:

1. Find out who is already good at this. Ask, without judgement, what people are using. You will usually find one or two quiet power users. They are worth more to you than an external hire because they already understand your work.
2. Give those people two hours a week, formally. Not a title, not a pay rise, not "AI lead" — time. Two hours to document what they have worked out and show one colleague. This is the single cheapest intervention available to a small business.
3. Update role descriptions before the next review cycle, not after. If you expect AI use in the job, it needs to be in the job description before it appears in someone's performance review.
4. Set the floor. Every role gets a licence, an hour of onboarding, and three worked examples relevant to their function. The floor matters more than the ceiling — a team where everyone is mildly competent beats one with a single expert.

Budget guidance: for a ten-person team, expect roughly £200–£300 a month in licences and about two working days per person in the first quarter. That is the real cost. Anyone quoting you less is assuming your staff will learn on their own time, which they will not.

When a specialist genuinely is worth it

There are three situations where a dedicated hire earns the money.

You are shipping AI in a product. If customers pay for a feature that calls a model, that is engineering with uptime, cost and liability attached. Hire an engineer, not a strategist.

You are past about fifty people with several functions adopting at once. At that size the coordination cost is real and someone needs to own tooling decisions, data policy and vendor contracts. Below it, the owner or an ops lead can carry it.

You are in a regulated sector and the compliance load is genuine. Financial advice, healthcare, legal, recruitment at scale. Here the hire is often a compliance or risk person who understands AI, rather than a technologist.

In every other case, the fractional option is better value: a few days of external help to set direction, then your own people executing. A £2,000 engagement that produces a tool decision and a written policy beats a £50,000 hire who spends six months learning your business.

The bottom line

The businesses getting real value from AI are not the ones who hired for it. They are the ones where a finance person, an account manager and an operations lead each found two hours a week they no longer spend on drafting, and where nobody was punished for admitting the first three attempts were rubbish.

So write AI into the roles you were already hiring for, screen for verification and judgement rather than tool names, give your existing quiet power users formal time instead of a title, and keep the specialist hire in reserve for the day you have an engineering problem rather than an adoption one. That order — team first, hire last — is what separates the small businesses compounding an advantage from the ones still asking which tool to buy.