Watch someone new to AI at work and you will see the same thing almost every time: a one-line request, a bland answer, and a shrug. "It's not that good." The tool was fine. The brief was thin.
Yes, prompting is a job skill, and a small business can teach it. But teach it as briefing, not as a bag of tricks. The people who get the most from AI are rarely the most technical. They are the ones who can explain a job clearly to someone who has never seen it before. That is a management skill every business already has somewhere, and it can be spread across a team in a month.
Is prompting really a job skill?
Yes: research now shows that how people write their requests accounts for a large share of what AI delivers at work, not just which model they use.
The clearest evidence comes from a study led by Eaman Jahani and David Holtz, Prompt Adaptation as a Dynamic Complement in Generative AI Systems (latest revision January 7, 2026). Across two preregistered experiments, 3,750 people wrote about 37,000 prompts. When participants were given a stronger model, performance rose. But on structured tasks, with fixed criteria and a clear goal, roughly half of that gain came from people changing how they wrote their prompts, not from the model itself. On open-ended creative tasks, the model did most of the work and prompt changes mattered far less.
Two details matter for a business owner. First, the researchers found that automatic prompt rewriting could not stand in for people adapting their own requests, and could even undermine the gains. You cannot buy your way out of the skill with a tool that "improves" prompts for you. Second, the people who did best were not the technical ones. As Holtz told MIT Sloan (August 4, 2025): "The best prompters weren't software engineers. They were people who knew how to express ideas clearly in everyday language, not necessarily in code."
My reading: prompting at work is not a new technical craft. It is the old craft of giving a clear brief, applied to a new colleague who reads fast, never gets tired and knows nothing about your business.
Why are so many small businesses stuck at the shallow end?
Most small businesses are using AI but getting small returns, and the gap is skills and confidence, not access to tools.
The NatWest AI Adoption Report, published September 25, 2026 from a survey of 1,400 UK SMEs and mid-market businesses, found that "44% of businesses use AI today, but only 6% of users have reached transformational adoption." Firms at the earliest stage typically report saving 1% to 10% of working time; firms at the transforming stage often report average savings of 61% to 70%. NatWest's own diagnosis: "Many businesses still lack the skills and expertise needed to deploy AI confidently."
The US picture rhymes. In the Goldman Sachs 10,000 Small Businesses survey (March 17, 2026; 1,256 business owners), 76% said they currently use AI, yet only 14% said it is fully embedded in their core operations, and 73% said they would benefit from additional access to training and implementation resources.
And time is the quiet killer. Workera's 2026 State of Skills Intelligence Report (September 23, 2026) surveyed staff at much larger US organizations, so treat it as a signal rather than a small-business measure, but the signal is loud: the share of companies offering AI skills training rose from 25% to 58% in a year, and 56% of employees said no time was set aside during work hours to build the skills.
Put those together and the shallow end makes sense. People have the tool. Nobody has shown them how to brief it, and nobody has given them the hour to practice.
The New Hire Test
The New Hire Test is the way I suggest a small business teaches prompting: as briefing, with one question asked before any prompt is sent.
The New Hire Test is how I suggest a small business teaches prompting: teach it as briefing, not as a technique. Before anyone sends a prompt, they ask one question: could a capable new hire, on their first morning, do this job well from these words alone? If not, the brief is missing one of four things: the job, the material, the standard or the limits. Add what is missing, then treat the answer as a first draft to be managed, not a result to be accepted.
The four things are worth spelling out, because they are what a good manager already gives a new starter without thinking about it.
The job. What you want, who it is for and why it matters. Not "write an email to a customer" but "reply to a customer who has ordered from us for six years and just received the wrong item; she needs to know what happens next and feel that we noticed."
The material. Everything you would hand a new hire: the customer's email, the returns policy, last year's version of the report, the notes from the call. AI cannot know what is in your filing cabinet. Most weak answers are weak because the facts were never supplied, so the tool filled the gaps with plausible filler. Material also comes with a boundary: nothing goes in that your AI policy says must never go in, which is one of the five questions in the One-Page Five.
The standard. What good looks like. Length, tone, format and, best of all, an example: "Here is a reply we sent last month that the customer thanked us for. Match its tone." One real example does more than a paragraph of adjectives.
The limits. What must not happen. Don't promise a refund date. Don't mention the competitor. Keep it under 150 words. For anything that goes to a customer, I suggest adding the standing instruction from the Receipt Rule: "Keep every detail I have given you. Do not add any claim, number, quote or promise I have not given you."
Then comes the half most people skip. A new hire's first draft is rarely perfect, and nobody fires a new hire over a first draft. You tell them what to change. The same goes for AI: reply, don't restart. "Shorter. Warmer in the first line. You've said we'll call her tomorrow; we haven't decided that, so take it out." That second exchange is where most of the quality comes from, and it is exactly the adapting that the Jahani and Holtz study measured.
If the New Hire Test sounds familiar, it is a cousin of the Errand Test, which asks whether a job is clear enough to hand to an AI agent at all. The Errand Test decides whether to delegate. The New Hire Test decides how well you brief.
What does a good brief look like in a real small business?
A good brief is usually four or five sentences longer than a bad one, and those sentences are almost all facts only the business knows.
Take a hypothetical 18-person landscaping company. The office manager needs a follow-up email after a site visit to a new commercial client.
The thin version: "Write a follow-up email after a site visit."
The result is a polite, generic email that could have come from any company in the country. She deletes most of it and writes her own, and concludes AI saves her nothing.
The briefed version runs to five sentences. It names the client and what they manage (a business park with three buildings). It pastes in the surveyor's site notes. It says what the email must do (confirm the three problem areas they walked, say the proposal will follow by Friday, ask who signs off). It supplies one earlier follow-up that won a contract as the model for tone. And it sets the limits: no costs, no dates beyond Friday, nothing about work they did not discuss.
The result needs two edits and goes out in ten minutes. Same tool, same person, same afternoon. The difference was the brief.
Won't better AI models make prompting unnecessary?
Partly, for some tasks, but not for the structured, checkable work that fills most of a small business's week.
The argument has something to it. The Jahani and Holtz study found that on open-ended creative tasks, improvements came mainly from the model, and the way people prompted mattered much less. If your team mostly asks AI for brainstorms and first ideas, better models will carry more of the load.
But look at what a small business actually runs on: quotes, follow-ups, schedules, reports, replies, summaries of calls. These have a right answer, or at least a wrong one. That is exactly where the study found people's prompting accounted for roughly half the gain from a better model. And no model, however capable, knows your client's name, your returns policy or what you promised on Tuesday unless someone tells it. A better model makes a good brief go further. It does not write the brief.
This is also why "prompt engineering" as a list of magic phrases has aged badly while briefing has not. Tricks are tied to one model and go stale. A clear brief works on any model, and on any person.
What I See When I Train Business Teams
What I see most often is not people who can't use AI. It is people who have never been shown what a good request looks like, and so decide the tool is the problem.
Across the business teams I train, the first prompt most people write is a single line. When I ask them to brief a colleague sitting next to them on the same job, out loud, they suddenly supply the customer's history, the deadline, the thing that went wrong last time and what "good" means here. They knew all of it. They just didn't think the machine needed to be told.
That moment is familiar to me from a previous life. Before I worked with businesses I was a teacher, an assistant headteacher and a Director of Digital Strategy in further education, and a large part of teaching is exactly this: explaining a task so clearly that thirty different people can do it well without asking you again. The skill transfers almost perfectly. The best prompters in a business are often its best trainers and best managers, and they are rarely the people who were given the AI project.
The principle underneath is the one I come back to in every session: outsource the doing, not the thinking. The brief is the thinking. The draft is the doing.
How do you teach prompting to a small team?
Teach it on real work, in pairs, with a shared record of briefs that worked, and protect a little time every week for it.
A caution first. In my article on AI training for small business teams I argued that training that ends with "a folder of prompts" fades, because it skips what makes training stick:
The Four Rungs are what I suggest any workplace AI training must give people, in order, if it is to change what they do rather than what they know: a real job to use it on, clear permission, practice within the week, and a return. People climb only as far as the first missing rung, however good the session was.
Prompting lessons are no exception. Here is how I would run them in a 20-person business.
1. Start from one real job per person. Each person picks one recurring task they do every week. If the business has already chosen its first AI job using the One Job Rule, begin there.
2. Show the difference live. A manager takes one real task and briefs AI twice in front of the team: once in a single line, once with the job, the material, the standard and the limits. Put the two answers side by side. Nobody needs convincing after that.
3. Test briefs on a person first. In pairs, each person writes the brief for their job, then hands it to their partner, who plays the new hire and says out loud what they would need to ask. Every question is a missing piece. This exercise needs no AI at all, and it is the fastest way I know to make the skill visible.
4. Keep the house briefs. Collect the briefs that produced good work in one shared document, each with the job it belongs to, the person who owns it and the date it last worked. This is not a folder of magic prompts. It is a record of how your business explains its own work, and it needs a quick review every month or two, because models change and the study is clear that people's prompts change with them.
5. Protect twenty minutes a week. Jahani's advice to companies, reported by MIT Sloan, is that "companies need to continually invest in their human resources." In a small business that means something modest and specific: a standing slot where people try their brief on the week's real work and share one that worked.
What should you do this week?
You can start this without a budget, a consultant or a new tool.
- Pick three recurring jobs in the business that involve writing: a customer reply, a follow-up, a weekly summary.
- Write the New Hire Test on one card and put it where people use AI: "Could a capable new hire, on their first morning, do this job well from these words alone?"
- Run the side-by-side at your next team meeting: one task, a one-line brief and a full brief, both answers on screen.
- Start the house briefs document with the full brief from that meeting as entry number one.
- Ask one question on Friday: "Which brief saved you the most time this week?" Add the answer to the document.
Where to go next
If your team has the tools but nobody has shown them how to brief AI on their own work, that is what my AI training for teams is built to do: practical sessions on your real jobs, not a tour of features. And if you want to know where your business stands first, the free Business AI Readiness Scorecard takes about five minutes: 12 questions across six areas, with an instant report and a 90-day plan.
Sources and further reading
- Eaman Jahani, Benjamin S. Manning, Joe Zhang, Hong-Yi TuYe, Mohammed Alsobay, Christos Nicolaides, Siddharth Suri and David Holtz, Prompt Adaptation as a Dynamic Complement in Generative AI Systems, arXiv, first posted July 19, 2024, revised January 7, 2026.
- Seb Murray, Study: Generative AI results depend on user prompts as much as models, MIT Sloan Ideas Made to Matter, August 4, 2025.
- NatWest Group, AI growth gap emerging across UK businesses, as smaller firms risk missing out on AI opportunity, September 25, 2026.
- Goldman Sachs, Survey: Small Businesses Embrace AI, But Need Training and Support to Fully Harness It, March 17, 2026.
- Workera via PR Newswire, AI Training More Than Doubled This Year, but 56% of Employees Report No Time at Work to Build the Skills, September 23, 2026.
Dan Fitzpatrick is The AI Educator: a former teacher, assistant headteacher and Director of Digital Strategy who now helps businesses train their people to use AI well, and hosts a daily podcast and a newsletter with more than 44,000 subscribers.
