Two Staff, Not a Data Team

Fifty percent of SMEs report their employees lack the skills to use generative AI effectively. Fewer than 30% offer any training. Those two numbers, sitting next to each other in OECD survey data, describe a specific kind of waste: tools running every day, producing output nobody fully trusts, while the people using them have never been taught what to ask for.

The confidence-competence mismatch nobody talks about

The Techaisle and AWS survey of SMBs found that 37% of small businesses name AI skills as their top barrier. Not tool access. Not cost. Skills. And the smallest firms in the survey struggled most with this, which makes sense: a larger company has someone whose job is to figure out why the AI output is wrong. You don't.

What makes this pattern strange is that the firms reporting the skills problem already use AI daily. They're not blocked from the tools. They're using them badly, without knowing it, because nobody has ever explained what a clear request looks like versus a vague one.

The research on requirement-oriented prompt engineering, abbreviated ROPE in the literature, tests this directly. Novices who receive focused instruction on expressing clear requirements in prompts produce significantly more accurate outputs than untrained peers using the same tools. The training doesn't change the model. It changes what the person feeds into it.

The objection worth taking seriously

A reasonable critic would argue that prompt training is a shrinking investment. Models are becoming more conversational. They ask clarifying questions now. They infer missing context. If the model keeps getting better at handling vague input, training staff to write precise prompts is building a skill against a moving target.

This is a real argument. It applies most directly to elaborate prompt construction: multi-step chains, role assignments, formatting instructions. Newer models absorb that complexity without explicit direction. The objection lands there.

It does not land on the ROPE finding. Requirement-focused training teaches something different from syntax. It teaches staff to form a clear mental model of what they need before they type anything. A more conversational model makes the exchange easier. It does not supply the missing clarity about what the output should accomplish. That gap exists in the person, not the interface, and a better model does not close it.

What two people, trained well, actually changes

The Copenhagen Business School barriers research found that lack of AI competence and little prior effective experience are documented even in firms with tool access. Not a few firms. Most of them. The fix the research points toward is internal training tied to real workflows, not generic AI literacy courses.

The practical version of this is narrower than it sounds. You're not building a training program. You identify two people who already touch AI daily, whose work the rest of the team depends on. You teach them one thing: how to express what they need in terms of output, purpose, and constraints, before they write a single word of a prompt. The ROPE research shows this instruction produces measurable output quality gains in controlled conditions.

Those two people become the firm's working standard for what good AI use looks like. Not a policy document. A person you can watch and copy.

The Canada SME AI Adoption Blueprint puts the scale of the skills problem plainly: half of SMEs in OECD surveys report employees lack the skills to use generative AI effectively. The blueprint calls for skilling, reskilling, and upskilling at both foundational and targeted levels. That framing is correct but too large for a founder without a data team. The targeted version is two people, a specific task category, and a clear definition of what accurate output looks like for that task.

Hiring a data scientist does not solve this. A data scientist builds pipelines and models. The problem here is that your staff do not know how to communicate with the tool they already run. Those are different problems, and the second one is cheaper to fix.