Before you prompt anything, get a clear mental model of what today's AI is actually good at — and where it will quietly let you down.
When people say "AI" in 2026, they almost always mean a large language model (LLM) — the technology behind ChatGPT, Claude, Gemini, and Microsoft Copilot. It's worth understanding what one actually does, because the mental model determines whether you use it well.
An LLM is a very sophisticated pattern-completion engine. It has read an enormous amount of text and learned the statistical shape of language — which words, ideas, and structures tend to follow which others. When you give it a prompt, it predicts a plausible continuation, one piece at a time. That's it. It is not looking anything up in a database, it is not doing arithmetic the way a calculator does, and it has no built-in sense of whether what it's saying is true.
This sounds like a weakness, and sometimes it is. But "predict the most plausible professional-sounding response" turns out to be shockingly useful for a huge slice of knowledge work — including a lot of what happens in an accounting practice.
The sweet spot is work that is language-heavy, judgment-light, and easy to verify. In those conditions AI acts like a fast, tireless junior who drafts something 80% of the way there in seconds, leaving you to review and finish.
Concretely, accountants get real value from AI on:
Notice the pattern: in every case, AI produces a draft or a hypothesis that a professional then checks. It compresses the blank-page time, not the judgment.
Because an LLM generates plausible text rather than retrieving facts, it will sometimes invent things — a court case that doesn't exist, a Code section with the wrong number, a deduction limit that was right three years ago. The dangerous part is the tone: fabrications arrive with exactly the same confidence as correct answers. Never treat a bare AI claim about a rule, rate, threshold, or citation as authoritative. Verify against the actual source.
An LLM is a language model, not a calculator. It can reason through the steps of a calculation well, but it can still drop a digit or round loosely in the final number. For anything that has to foot, use a real tool — a spreadsheet, a calculator, or the AI's ability to write and run code — rather than trusting a number it simply typed out.
A model's training has a cutoff date, and tax law moves. Unless a tool is explicitly connected to live, current sources, assume its knowledge of this year's brackets, mileage rates, and filing thresholds may be out of date or averaged across years. Always anchor time-sensitive work to the current-year authority.
Before you lean on any AI output, ask three questions:
The safest place to start is a language task where you can see the whole draft. Try one of these next:
Plain-English client explainer → Next: Keeping client data safe →Free weekly read on AI for accountants — and a heads-up when new Academy lessons go live.