In the last few months, the models your team types into stopped behaving like the ones they learned on. Opus 4.8 and GPT-5.5 shipped, and the agents built on them now run for hours on a single instruction — reading files, drafting, and handing back something close to finished. Your firm's AI training hasn't moved with it.
Most owners will close that gap the way the profession closes every skills gap. Book a course, log the CPD hours, tick the box. And the AI training on offer to accounting firms today is almost always a prompt-engineering course built for 2024 — back when we told everyone to talk to the model like a junior.
Here's the uncomfortable part. The skills that actually build an AI-native team are mostly learned through reps in your own work, not in a lecture — and the most valuable of them may not qualify as CPD at all. That shouldn't be the reason you skip them.
Prompt engineering was a 2024 skill
In 2024, the model needed hand-holding. You specified the format and made every structural decision. That advice was right for the model we had then.
The model we have now works like a senior you delegate to. You give it the goal, the context, the sources, the constraints, the standard, and where to stop — then you get out of the way. Nate Jones calls the shift "from prompting to briefing."
The magic words and role-play tricks are dead. What survived got renamed — and your profession already has a name for it. Making your intent legible is a planning memo. It's a review note. It's the workpaper standard: could someone else pick this up and act on it?
So a course that drills prompt syntax teaches a skill the models already outgrew. I can say that plainly: I used to teach one — nine hours of it. The tool literacy that still matters — what the model can and can't see, that it will invent a confident answer, how a file differs from a chat — now fits in a half-day.
Why your CPD hours quietly work against you
CPD measures hours. It rewards structured, assessed, instructor-led content — a precise description of the format that doesn't transfer. The thing that does transfer — a person running reps on their own client files, getting it wrong, and learning to see why — is messy, self-directed, and often unclaimable.
Read that again, because it's the trap. An owner optimising for claimable hours is selecting against capability. The format that ticks the box is the one your team forgets by Friday. CPD exists to keep professionals current, and the way it's counted now fights the one technology moving fastest.
This isn't an argument against structured learning. It's an argument against training whose only output is a completion certificate. Spend the hours you're required to spend — just don't let "claimable" decide what your people learn.
What an AI-native team actually practises
There are two different relationships with AI, and most training teaches only one. You can delegate to it — brief it, let it produce, judge the result. And you can think with it — let it test your assumptions, argue the other side, surface what you missed.
Delegation builds throughput. Thinking with it builds judgment. Train only the first and you've taught your team to commission output faster while the judgment muscle wastes.
The skill underneath both is recognition: looking at confident output and knowing what's wrong and why. That used to be built for free. Juniors developed it by grinding through the production work — the recs, the returns, the workpapers — that AI now does.
OpenAI's GDPval benchmark put frontier models at professional quality on 70 to 83 percent of tasks; the 17 to 30 percent they get wrong is exactly where judgment lives. A practitioner who can't spot that slice is dangerous with an agent, not productive with one. Recognition is the apprenticeship you now have to rebuild on purpose — because the old path to it is being automated away.
Training installs habits. The firm builds the conditions.
Even the best course isn't enough on its own. Reps only compound inside conditions: a private space to test a hunch at zero social cost, quality systems that surface the exceptions, a culture where "I checked it with AI" counts as real work, and a scoreboard that tracks capability moved rather than hours logged. That's not a training decision. That's your job as the owner. The course installs the habit; the firm keeps it alive.
The six questions to ask before you book any AI training
Before you spend a dollar or an hour on AI training — yours or a vendor's — grade it.
Does it run on your actual work, or on generic examples? Does each person walk out with something the firm keeps — an encoded process, a correction log — or a certificate? Does it build recognition, or just production?
Will it change how they work on Monday? Does it match the person — a senior needs delegation reps, a junior needs recognition reps? And is success measured in hours, or in capability?
If a program only scores well on "can we claim it," you already know what it is.
I built the redesigned version of this for accountants inside AI Essentials — reps on your own files, not a prompt lecture. Want the one-page scorecard to grade what you're already running? It's yours when you subscribe below. So here's the real question: are you about to train your team, or just file the hours?