You cannot buy a lower touch rate. You train for it.

Every organisation I work with has funded AI training. Almost all of it was prompt training.

In the 4D framework for AI fluency, developed by Professors Rick Dakan of Ringling College and Joseph Feller of University College Cork in partnership with Anthropic, and released under a Creative Commons licence, prompting sits inside one competency called Description. Which makes prompt engineering roughly a quarter of one of four things, or about six per cent of AI fluency.

Almost every rand your organisation has spent on AI training bought that six per cent.

The four competencies, and which one you are missing

Delegation is deciding what to hand over and what to keep. This is a value screen, but at the level of one person's inbox rather than one board paper. Most people are poor at it because nobody has ever asked them to be explicit about which parts of their job are judgement and which are throughput.

Description is telling the system clearly what you want and how to work. This is prompting, plus the surrounding discipline of specifying context, format and constraints. It is the competency that is easiest to teach, which is precisely why it is the one everyone teaches.

Discernment is catching it when it is confidently wrong. This is the one that matters, and I will come back to it.

Diligence is owning the result you put your name to. Verification, attribution, and accepting that the output is now your work product regardless of how it was produced.

The framework's own insight is that Description and Discernment form a loop. People do not get better at asking by being taught to ask. They get better at asking by getting better at judging what came back, which then informs the next request. Train one without the other and you get fluent requests for work nobody can evaluate.

Why discernment is the one that moves the number

In an earlier post in this series I argued for human-touch rate as the metric that matters: the share of what your agents produce that a person still has to check before the work counts as done.

Here is why that metric and this competency are the same subject.

Touch rate falls when checking becomes sampling rather than gating. Checking can only become sampling when the people doing it are good enough at judgement to know which outputs need a second look and which do not. That is discernment, and there is no other route to it. You cannot buy a lower touch rate. You cannot procure it, licence it, or negotiate it into a contract. You train for it.

And the failure mode is worse than neutral. Weak discernment plus strong description ships polished mistakes faster. An operator who instructs the model well but judges its output poorly is not merely unproductive. They are actively harmful, because they increase the throughput of confidently wrong work while every fluency metric you are tracking goes up.

The Aristotle distinction

Aristotle separates techne, craft knowledge, from phronesis, practical wisdom. Techne is the knowledge of how to make a thing. Phronesis is the judgement to know which action is right in this particular situation, with these particular circumstances, at this particular moment.

You cannot get phronesis from a manual. It comes from doing the thing repeatedly, under supervision, with someone competent telling you when you were wrong.

Prompt training is techne. Discernment is phronesis. Organisations have been buying the first and hoping the second arrives free, which is not how it has ever worked in any professional discipline, from medicine to auditing to engineering.

What a discernment programme actually looks like

This is the part where most training strategies stall, because discernment does not decompose into modules.

It does decompose into practice, though, and the structure is well understood because every apprenticeship in history has used it.

A corpus of wrong answers. Collect real agent outputs from your own workflows, including the ones that were subtly incorrect. Not the catastrophic failures, which everyone spots, but the plausible ones. This corpus is the single most valuable training asset your organisation can build, and it costs nothing but the discipline to keep the failures instead of deleting them.

Calibration sessions. People score outputs independently, then compare and argue about disagreements. The argument is the learning. Inter-rater agreement is a measurable proxy for organisational discernment and almost nobody tracks it.

Feedback with a delay. People need to find out whether they were right, and they need to find out often enough for the loop to close. If nobody ever learns which of their approvals turned into a complaint six weeks later, no learning is occurring at all.

Supervision by someone better. The uncomfortable part. Discernment transfers from people who have it, which means identifying who those people are and giving them time they currently do not have.

Train the room that signs the cheque

One more thing, and it is the point I would make to any CHRO or learning lead.

Your AI fluency is not the percentage of staff who have used a copilot. It is the percentage of your decision-makers who can interrogate an AI claim without help.

Four questions every executive should be able to ask unprompted, and none of them are technical:

• Is this evidence or is it a coincidence? What was it compared against?
• What would a person cost to do this same unit of work?
• If it is wrong, can we undo it, and how long does that take?
• What condition, not what date, unlocks the next stage?

An organisation where the approvers can ask those four questions will make better AI decisions than one where every practitioner has a prompt certificate and the exco cannot tell a rung-one claim from a rung-three one.

Train the room that signs the cheque, not only the room that writes the prompt.