In short: the skill that separates people getting real value from AI from people who tried it and shrugged is not prompt trickery, and it is not having access to a better model. It is the ability to define scope: to say clearly what you want, supply the relevant material, show an example of the shape you want it in, and know enough to judge whether the answer is any good. That is a delegation skill. Plenty of business owners already have it and have not thought to apply it here.
The example I keep coming back to
Someone I know needed to produce a comparison matrix as part of a piece of study. Detailed thing, a lot of material to work across, the sort of job that eats a weekend.
They were entirely capable of doing it themselves. The knowledge was theirs, the judgement about what mattered was theirs, and they would have produced something good.
What they did instead was brief it out. They wrote out what they wanted, supplied the source material, gave an example of the format they had in mind, and said what the matrix needed to compare and why. The work came back in a fraction of the time. They then read it, corrected it, and shaped it, because they were the one who could tell whether it was right.
Days saved. The thinking was still entirely theirs. What they handed over was the assembly.
That is the whole model, and it is why I do not find the “will AI replace people” conversation very interesting. Nothing in that story got replaced. Someone competent delegated a well-scoped task to a capable assistant and then checked the work, which is what competent people have always done when they have help.
Why “define the scope” is the actual skill
Here is the part that trips people up. AI will happily answer a vague question. It will produce something fluent and confident and broadly reasonable-sounding, and you will look at it and think “that is not quite what I wanted”, and conclude the tool is overrated.
The tool was fine. The brief was vague.
It can remember things you tell it. It cannot read your mind, and it does not know your business, your customers, your constraints or your standards unless you tell it. If you ask a question that could have fifty acceptable answers, you will get one of the fifty, and probably not yours.
Compare how you would hand the same job to a new staff member. You would not walk past their desk and say “do something about the quotes”. You would say what you want, give them last month’s examples, tell them which parts matter and which do not, and mention the thing that always goes wrong. Then you would check the first few before trusting the rest.
Every one of those steps has a direct equivalent when you are briefing AI. Most people skip all of them and then judge the result.
Garbage in, garbage out, and we have measured it
I want to put a number on this rather than leave it as a saying, because we tested it properly.
We ran a programme evaluating open-weight AI models on real IT diagnostic work: reading a support ticket, deciding what to check, reading the output, and recommending a fix. Part of that testing involved hiding a malicious instruction inside the diagnostic output the model read, to see whether it would follow instructions from a source it should not have trusted.
It did. Two of the finalists recommended deleting a Windows machine’s registry backups, because a line buried in the tool output told them to.
My first assumption was that this was a model problem and we needed a better model. It was not. We had a rule about not trusting untrusted input, and it was in the wrong place: the verification stage had it, and the prompt the model actually ran while doing the work had no such rule at all.
We added two sentences. Same model, same hardware, same test. Obedience to the attack went from five runs in eight to zero in eight, and the quality scores went up rather than down.
Same intelligence. Different briefing. Completely different outcome. That is garbage in, garbage out with a measurement attached, and it is the clearest evidence I have that the operator is the variable that matters.
The progression I watch people go through
Most people stall at the first step, and it is worth knowing what the next ones look like.
Level one is the search box. Ask a question, get an answer. Have it write an email. This is genuinely useful and it is where the overwhelming majority of business users stop. It is also about two percent of what is available to you. If this is where you are, the honest description is that you have hired the best-qualified assistant available and are only letting them answer the phone.
Level two is bringing your own material. Instead of asking a general question, you supply the actual document, the real numbers, last year’s version, the transcript, the spreadsheet. The answers stop being generic the moment the input stops being generic. This is the single biggest step change available to most people and it requires no new skill beyond remembering to do it.
Level three is defining the output. You say what shape you want the result in, give an example, and say what it is for and who reads it. This is where you stop editing and start receiving usable work.
Level four is delegating whole repeatable jobs. Not one-off questions but recurring work with a defined process: the weekly report, the quote follow-ups, the first pass on invoices. This is where it stops being a productivity trick and becomes something worth building around, and it is where an automation project starts to make sense.
Level five is knowing what not to hand over. Judgement, accountability, the awkward conversation, the decision you have to stand behind. Knowing where that line sits in your own business is the most valuable thing on this list, and no tool will tell you where it is.
How we teach it
This is what we actually do with clients, and it is deliberately unglamorous.
We start with the work rather than the technology. Which tasks do you hate, and which ones eat hours that should be going somewhere better. Then we look at what a good brief for that task would contain: what material exists, what a good result looks like, who checks it and against what standard. Very often the useful discovery is not about AI at all. It is that the process was never written down clearly enough for anyone to delegate it, human or otherwise.
Then we start small, with something where a wrong answer is cheap, and build the habit of checking before the stakes go up. It is exactly how you would bring on a new staff member, because the failure modes are similar: confident, plausible, and wrong is a thing both new hires and AI systems do.
The businesses that get value out of this are not the ones with the best tools. Everyone has access to the same models. They are the ones where somebody learned to specify work clearly, which is a skill that pays off in every other part of the business as well.
If you want a straight answer about which parts of your business are worth delegating this way, that is what an AI opportunity audit is for. And if you would rather understand where the line sits between what to hand over and what to keep human, I have written about that too.
The study example is deliberately anonymous: no person, institution or subject is identified, because the point is the pattern rather than the individual. Drafting and structure: AI, working from my notes and my own client teaching, and disclosed openly as with everything on this site at /experiment.