In this article, Dudley Gould returns to explore how AI's ability to explain things and to receive clear explanations, is changing the way he learns, writes, and thinks about professional communication.
One of my most-used prompts at the moment is: "Explain this to me like I'm ten years old."
Followed by: "Now explain it like I'm fifteen."
When I'm getting to grips with a new concept – a new accounting standard, a regulatory development, a piece of technology I haven't encountered before – I'll often start there. AI is genuinely excellent at explanation. It can take a complex idea, strip it back to its essentials, and build it back up at whatever level you need. Better than most textbooks, and considerably more patient.
But there's a flip side to this that I've been thinking about more recently.
The bar is set by agents
Andrej Karpathy – co-founder of OpenAI – said something in a recent No Priors podcast episode with Sarah Guo that reframed how I think about all of this:
I'm not explaining things to people anymore. I'm explaining things to agents. If you can explain it to an agent, the agent is the distribution layer.
If AI is good at explaining things to you, the quality of that explanation depends on what you give it. Ask a vague question, get a vague answer. Ask a precise, well-framed question, and the output is precise and useful. The skill isn't in the AI — it's in the clarity of the explanation you give it.
The same applies when the direction is reversed: when you're using AI not to learn, but to produce.
The same principle, applied to writing
When writing my last article for this community – Is AI being asked to solve the wrong problems? – I used an open source AI agent called OpenClaw to help shape my ideas.
OpenClaw is an open-source AI agent. Unlike a chatbot – where you ask questions and get answers – OpenClaw runs persistently on your own machine, maintains memory, and can take actions on your behalf.
Here's how it worked.
- Brain dump
I use Open Wispr – a speech-to-text tool – to get my thinking out before I write a single word. I might talk for ten minutes: the argument I want to make, the examples I have in mind, the things I'm not sure about, along with any interesting links or articles I've saved. No structure, no editing. - Agent Q&A
I keep a repository of all my articles – my AccountingWEB columns going back to 2021 and my previous ICAEW pieces – and I've given OpenClaw access to it. The two audiences are quite different, so the styles are too, and I make sure OpenClaw takes that into account when it drafts.
My brief: read everything in the repo and use it to understand how I write. When I hand you raw material, ask me questions to sharpen the thinking, then draft in my voice.
OpenClaw came back with questions. Who's the main audience? What's the one thing you want them to take away? Is there a concrete example that makes the central point? What's the obvious counterargument? I answered each one directly. - Draft and edit
OpenClaw drafted the article. I read through it carefully and made edits to deliver a completed article.
The quality of the output depends entirely on the quality of the explanation going in, based on your idea, concepts, topics and themes. So, if the thinking of what you want to produce is unclear, the draft will be too. It’s also worth noting that I needed that repository of articles as a starting point – I still had to do the legwork to establish my own writing style. And ultimately, I'm responsible for what gets published.
The review remains crucial too. In fact, for this article, OpenClaw inserted the wrong link not just to the Karpathy podcast episode, but also to my last ICAEW article. Remember – AI is just a tool.
What this means for accountants
Vikram Pawar, who heads up the Claude Code Community in London, put it well on LinkedIn recently: "The question was never 'was this written with AI?' It's 'was this written with any respect for the reader's time?'"
The same principle that makes AI good at explaining things to you makes it good at helping you explain things to others. But only if you can explain clearly to the agent first.
A significant part of accounting and audit work involves explaining things: to clients, to junior colleagues, to regulators. Karpathy's insight is that the bar is now set by agents. Can you explain an accounting treatment clearly enough that an agent follows it and produces a useful output? Can you articulate a process precisely enough that an agent executes it reliably?
If you can, the agent handles the distribution. If you can't, it produces something generic that doesn't help anyone.
The constraint has moved. It's no longer about whether you can write quickly or present well. It's about whether your thinking is clear enough to instruct an agent well. That's where deep accounting expertise has a genuine advantage.
