There is a curious divide in the profession. Maybe there always has been. It’s most apparent these days in the contrasting attitudes to Large Language Models (LLMs), often referenced as AI. Among the most well know are ChatGPT, Claude and Gemini.
On one side of the divide are those accountants who dismiss LLMs as unreliable toys, citing well-publicised hallucinations and early missteps as though nothing material has changed. Spoiler alert, today’s LLMs are vastly more reliable that those of 6 months ago – and this will continue to be true no matter how long after I write this that you are reading it!
On the other side of the divide are those accountants who are experimenting daily — drafting emails, refining processes, brainstorming ideas and aware they are only scratching the surface.
Most sit somewhere in the middle. Curious. Cautious. Trying things, but not yet systematic.
At the most recent meeting of my Sole Practice Club, the range of attitudes was striking. Some members had barely begun to experiment. Others were quietly integrating AI into drafting, research and internal documentation.
When I suggested a few additional practical applications, it became clear that the real divide is not between enthusiasts and sceptics. It is between those who are using AI casually and those beginning to use it deliberately.
The same pattern shows up in conversations I have on the Accountants’ Success Secrets podcast. Some guests are enthusiastic early adopters. Others are watchful and wary. Very few have developed a structured approach to getting reliable output.
And reliability is the issue.
The spreadsheet parallel we have forgotten
Much of the scepticism I hear rests on a simple assumption: “It makes mistakes.”
That sounds reasonable until you remember we said exactly the same thing back when spreadsheets were the exciting, scary new technology.
When spreadsheets first appeared, many accountants saw them as error-prone and dangerous. In reality, spreadsheets did not make mistakes. Only poorly constructed spreadsheets did.
Once we learned basic spreadsheet design discipline (such as separating inputs from outputs, avoiding hard-coded numbers and embedded formulas; and building in checks) reliability improved dramatically. And we realised that any mistakes were down to the human who built the spreadsheet model and hadn’t tested it sufficiently.
These days most accountants don’t risk being the cause of mistakes in the spreadsheets they use. Instead we rely on commercial software, which is often little more than a very well-engineered spreadsheet maintained by specialists.
The technology matured. But so did we.
Large language models are following a similar path. You can choose to rely only on the AI features embedded in commercial software, where much of the risk has been engineered out. That is perfectly sensible.
You can also, if you choose to do so, learn how to use LLMs so that you can rely on their output.
What does not work is expecting them to read your mind.
AI is not in The Magic Circle
I speak as a long-standing member of The Magic Circle, and I can assure you that no AI model is currently enrolled. It cannot read you mind or intuit your intention. It processes patterns.
If the output is vague, generic, wrong or even just slightly off target, it is most often because the brief was not sufficiently detailed and specific.
If, instead of asking an LLM, you were briefing an intern, you would explain what you are trying to achieve, describe the audience, set boundaries and expectations, and review their draft before it went anywhere near a client.
Reliable AI output is not accidental. It is engineered.
Introducing the CLEAR equation
That brings me to a simple framework I have found helpful.
Think CLEAR
C — Clarity
What do you actually want? Not “write something about tax planning” but “draft a 400-word explanation for small limited company clients about pension contributions before year end, highlighting cash flow considerations”.
L — Load AI with context
Background matters. Who is the audience? What do they already know? What assumptions are we working with? What is the commercial objective?
E — Establish constraints
Tone, format, length, exclusions and regulatory boundaries. Are you looking for bullet points, a client email or a blog draft? Should it avoid technical jargon?
These three elements combine to create a specific prompt.
Only then do we reach:
A — AI generates
The model produces a draft based on patterns. It does not “understand” intent in a human sense. It responds to what you have actually written.
R — Review and Refine
Your professional judgement turns a competent draft into reliable output. You adjust nuance, correct emphasis, and ensure compliance and commercial alignment.
Put simply: CLEAR → Specific prompt + AI + Review and Refine = Reliable output
Where most people go wrong
Notice what is absent from that equation. Blind trust.
The mistake many practitioners make is jumping straight in. They open the LLM tool and type a loose question. The result feels generic. Confidence drops. The tool is blamed.
Yet if the brief is woolly, the output will amplify that woolliness. AI amplifies clarity. It also amplifies confusion.
The other common mistake is failing to review the output before sharing it. If it is not quite right, go back and refine your prompt. Ask the LLM to adjust tone, expand a section, tighten an explanation or remove assumptions. Iteration is not a weakness; it is a key part of the process.
More intern than oracle
It is tempting to treat LLMs either as infallible or as fundamentally flawed. Both positions are too simplistic. They are better understood as capable but literal, more like an eager junior team member than an all-knowing oracle.
An overeager teenage intern is not permanently unreliable. They require structure, context and supervision. Given that, they often become remarkably useful. And often grow up to become reliable mature adults.
Assuming LLMs will never be reliable is even more shortsighted than assuming that spreadsheets would always be dangerous curiosities. In the future no LLM will still be making the mistakes we see today. They are growing up, maturing and becoming more reliable much faster than you might think.
The real question for sole practitioners
The real question for sole practitioners is not whether AI is perfect. It is whether you are developing the discipline to use LLMs properly to assist you in running and promoting your practice and in serving your clients.
Some in our community are already moving from ad hoc experimentation to structured integration. Others are waiting for certainty that may never arrive.
The accountants that understand this will quietly gain an advantage, not because the technology is magical, but because their thinking is clearer.
If you would like to explore how to integrate AI more systematically into your practice, that is precisely the sort of conversation we can have within the Sole Practice Club or in a mentoring session. The opportunity is not in chasing novelty. It is in applying professional discipline to new tools, just as we have done before.
