AI tools are giving tax advice. Accountants should be behind them.

An AI tool was quietly telling tech founders at Network School they did not have to pay tax. It ran on ChatGPT, wired through a Tally form into a Notion database. 

The founder who built it was good at tech and not qualified in tax. There was no accountant behind it. There was no one accountable for anything it said. 

That was the moment I decided the profession's role in AI is bigger than most of us have been treating it. I gave my keynote at The Firm's AI in Practice Summit on this, and this is the version of the argument I want the profession to sit with.

Let me deal with the technical piece first. Language models are probabilistic, not deterministic. The answers they give are not based on rules. They are hallucinated. They are what you want to hear. That works for a lot of things. Tax advice sits outside what a probabilistic system can safely handle, and providing an AI tool that gives tax advice can itself constitute providing a regulated tax service, in Australia and elsewhere. The deeper issue is accountability. When the translation is wrong, no one carries the consequence.

The intermediary layer no one has built

I spent part of this year getting AI safety qualified. I did the frontier governance course. I read the Claude Opus system cards. What kept coming back at me was that every AI governance proposal assumes there will be non-conflicted intermediaries somewhere in the system. Someone qualified. Someone accountable. Someone with a public interest mandate. No proposal explains where those people actually come from.

That is the layer accountants are already built to fill. The International Federation of Accountants binds the profession to a public interest code of ethics that travels across borders. There are millions of us. We are used to being the accountable person when the numbers go wrong. That is the shape of the role AI safety keeps assuming will exist, and no one else has stepped into it.

The profession will not win a translation-work fight against language models. That fight is already lost. What we can do is verify, sign, and carry the accountability that AI cannot carry.

Build the proof now

The problem is that AI agents cannot find us. Our credentials sit in walled gardens. Our CPD sits in Excel spreadsheets on our own laptops. Our track record sits in client engagements no one else can see. If an AI agent needs a qualified human to verify a tax opinion at three in the morning, it has no way of knowing who is qualified, what they know, and how to pay them.

This is what I have been building at Network School. I put my professional graph on GitHub. Every commit is public. Fifteen years of work turned into a machine-readable record with nodes of thought leadership, credentials, and current builds. It is public CPD, and it is designed to be read by AI agents as much as by humans. I built an ask bot on top of it. AI can now query what I know and get an answer that traces back to what I have actually written and studied.

Once you have that, the rest of the stack makes sense. Credentials anchored on Bitcoin via SHA-256 hashes. If a certificate is fraudulent, the fraud is there forever, so I only put the real ones on. Payment via L402, an open protocol that combines digital authentication with Bitcoin Lightning Network payments. It activates a dormant web status code, HTTP 402 Payment Required, so an AI agent can pay for a verified opinion without credit cards, user accounts or passwords.

The end state is a proof of control standard. Cryptographic evidence of qualified human judgement at every AI decision checkpoint in a regulated workflow. The licensed human reviews the AI output, applies judgement, signs, and there is a record of that signature. The AI agent can only trade on the advice if it has a qualified human behind it.

Why accountants should study Bitcoin

Forget Bitcoin the asset for a moment. Bitcoin as a network is what matters here. A highly scalable network run by individuals without a central authority. That is the model the profession should learn from, because it is the model AI agents are going to operate on. Ask AI what kind of money it prefers, and it prefers money that can be controlled, verified and settled entirely through software. Bitcoin does that. The rails AI will run on are already being built.

Cryptography goes well beyond internet native money. It is what verifiability looks like as a system. Vendors are selling verifiability under a lot of different names, and some are delivering it without knowing what they have. If the profession does not have the fluency to evaluate those claims, someone else will define what qualified means, and we will live with the definition.

Raise the fluency bar, then build

Real AI fluency starts where the basic hygiene sessions end. Sitting through an hour on not pasting client data into ChatGPT is hygiene, and hygiene is a starting line. Fluency comes from building your own tools, coming face to face with the failure modes, and understanding what can actually go wrong. In my submission to the Tax Practitioners Board on AI in the code of professional conduct, I argued the profession should hold a much higher technical bar. I still believe that.

Get on GitHub. Build in public. Publish your work in a form AI can read. Anchor your credentials somewhere verifiable. Make yourself procurable. The play is to become the qualified human of choice AI has to introduce to its client. Discoverable. Verifiable. Procurable.

You can watch the full session here.