The AI talent war is one SMEs cannot win. Good news: you do not need to
The frontier AI labs, the investment banks and the big consultancies are paying machine learning specialists the kind of packages that make a director of a twelve-person firm laugh out loud. You will not outbid them. You will not get close, and every month spent trying to recruit "an AI person" at SME money is a month of nothing happening.
Here is the part that should cheer you up: the job those salaries buy is not the job you need done. Frontier labs hire people to build models. You need someone to do the plumbing: connecting models that already exist, and already work, to the specific processes your business runs on. Enquiries, quotes, site records, invoices, chasing. That is not research. It is careful, slightly unglamorous integration work, and the person best placed to do it may already be on your payroll.
Why can you not win the bidding war?
Because you are bidding for the wrong role in the wrong market. The scarce, expensive skill is building and training large models. A handful of organisations do that work and they recruit globally. Everyone else, including most large corporates, is a consumer of those models, not a producer.
An SME competing for that talent is like a groundworks contractor trying to hire a bridge designer to lay kerbs. Even if you landed one, the work would bore them and they would leave. The ONS construction industry data shows an industry made overwhelmingly of small firms; none of them are going to staff a research team, and none of them need to.
What do SMEs actually need instead of researchers?
Plumbing, in three specific forms.
First, process knowledge: someone who knows that the quote goes out after the site visit, that the diary gets written (or does not) on a Friday, that invoices stall because the completion photos live on someone's phone. No model has this knowledge. It is the raw material of every useful automation.
Second, connection work: getting information out of the places it lands (inboxes, WhatsApp, spreadsheets) and into the model, then getting the output back into the tools people already use. This is configuration and judgement, not computer science, which is why we describe the work as automation rather than software.
Third, checking: someone accountable for reading what comes out before it matters, and tightening the process when it is wrong.
What are your realistic options?
Three, and they are not mutually exclusive.
| Option | Cost profile | Speed | Where it fits |
|---|---|---|---|
| Buy configured tools | Subscription, low setup | Fast | Common problems: transcription, document drafting, site record capture |
| Use a partner | Project fee plus support | Medium | Processes specific to your firm that off-the-shelf tools do not cover |
| Upskill an operations-minded employee | Salary you already pay, plus training time | Slower to start, compounds | Everything, forever, because they stay and keep improving things |
The third row is usually the best return on investment, and it is the one most firms never consider because they are busy scanning job boards for a unicorn. The office manager who built your spreadsheet system, the QS who actually reads the contract, the admin who fixed the filing structure without being asked: that person, given tools and time, becomes your AI capability. They already have the process knowledge that an external hire would take a year to acquire. The CITB makes the wider point in its research on construction skills: the industry's gap is filled by developing the people it has at least as much as by recruiting new ones.
Buying tools and using a partner still matter. Sensible firms buy the common stuff, partner for the specific stuff, and grow one internal person to own the lot.
What does "AI skills" actually mean at SME level?
Not Python. Not model architecture. Three habits, all learnable.
Process thinking: the ability to describe a task as steps, inputs and outputs. "Quoting" is not a process description; "enquiry arrives by email, someone extracts the scope, checks rates, drafts, director reviews, sends within three days" is. People who can do this can automate; people who cannot will automate the wrong thing.
Data hygiene: keeping information where a system can reach it. Named files in shared folders rather than attachments in personal inboxes. Consistent job references. This is dull and it is half the value.
Prompt-and-check discipline: writing clear instructions for a model, then verifying the output against something you trust before it goes anywhere. This is a quality assurance mindset, and construction people are better at it than they think, because snagging someone else's work is the industry's native skill.
Notice that none of these require a technical background. They require exactly the temperament described in building production capability as a set of habits: patient, systematic, mildly suspicious.
How do you develop the person you already have?
Give them a real problem, protected time, and permission to get it wrong for a while.
Pick one process that annoys everyone. Give your candidate half a day a week to work on it, with a tool budget smaller than one month of the salary you were going to offer the unicorn. Expect the first attempt to be clumsy. Run the rollout the way you would run any change with a wary workforce, which is to say deliberately: training sceptical teams is its own discipline and it rewards patience over enthusiasm.
Within a few months you will know whether it is working, because the annoying process will either be less annoying or it will not. AI Metric spends a fair amount of time helping firms structure exactly this arrangement, usually alongside training for the wider team.
The talent war will carry on without you, salaries climbing in a market you were never really in. Let it. The firm that wins locally is not the one with the researcher; it is the one where somebody ordinary was given the time to connect boring tools to boring processes and quietly compound the result.