The most useful question an accountant can ask a client this year isn’t “what are your numbers?” It’s “what are your numbers compared to what they should be?” The first question takes a spreadsheet. The second one used to take a senior partner and half a day. With the right workflow, AI handles the bulk of the comparison and lets you spend your time on the conversation that actually moves the client’s profit.
The AICPA and CPA.com 2025 AI in Accounting Report found that 85% of accountants feel excited or intrigued by AI, with speed and error reduction as the top cited benefits. Wolters Kluwer’s Future Ready Accountant 2025 reported AI adoption inside firms jumping from 9% to 41% year over year. The market is moving fast. The advisory work is what’s getting harder to deliver, and that’s exactly where AI changes what’s possible for a small firm.
Pass one: turn the GL into a question list, not an answer
Before you ask AI for insights, ask it to be sceptical. The mistake most practitioners make on their first profitability prompt is asking the model to “find opportunities,” which produces a list of generic suggestions (“review subscriptions,” “renegotiate supplier terms”) that any reader could have written without the data.
Instead, run this against a 12 to 24 month general ledger export from Xero, QuickBooks, or Sage Intacct:
“You are a forensic management accountant reviewing the attached GL for [CLIENT NAME], a [INDUSTRY] business with annual revenue around [REVENUE]. Do not give recommendations yet. Instead, produce: (1) a table of every expense category that grew faster than revenue over the period, with the % growth difference, (2) any month where gross margin moved by more than 3 percentage points versus the prior 3 month average, (3) any vendor where total spend exceeds 5% of revenue, and (4) any recurring transaction that appeared mid-year and has continued. Output as four short tables. No commentary.”
You now have the four most likely places profit is leaking. The model hasn’t told you what to do, it’s told you where to look. That’s a much better foundation for the next step.
Pass two: ask the dumb questions a client would never volunteer
This is the pass that separates AI-assisted advisory from AI-flavoured admin. Take the four tables from pass one and run them through a second prompt that interrogates them.
“Here are the four anomaly tables from [CLIENT]‘s GL. For each row, write the single most useful diagnostic question I should ask the owner before drawing a conclusion. Then, for each row, list two business-side explanations that would be benign and two that would indicate a real margin problem. Be specific to a [INDUSTRY] business of this size.”
What comes back is a meeting agenda, not a report. You walk into the client conversation with a tight list of “I noticed advertising spend grew 47% while revenue grew 8%, can you tell me what changed in May?” rather than a vague “let’s look at your costs.” Clients can answer specific questions. They glaze over at general ones.
The failure mode: confident hallucinations on percentages
Large language models are bad at arithmetic when they’re working from raw text. They are reasonably reliable when they run actual code on the data. The single biggest mistake new users make is asking ChatGPT or Claude to “calculate the gross margin trend” without enabling code execution, and then trusting the percentages that come back.
In ChatGPT, that means making sure Advanced Data Analysis (the Python code interpreter) is engaged. In Claude, attach the file and explicitly say “use the analysis tool to compute these figures, do not estimate.” Either way, ask the model to show its working in a small printed table you can sanity check. If the numbers in the table don’t match what you’d get from a 60 second pivot, the model is hallucinating, throw the run away.
A worked sanity check: if a client did $1.2M revenue last year and you ask for marketing spend as a percentage, the answer should be roughly the line-item total divided by the revenue. If the model says “marketing was 4.2% of revenue” but the line-item total is $86,000, the percentage should be around 7.2%, not 4.2%. Numbers that don’t reconcile to the source are the cleanest tell that something has gone wrong, and they catch most hallucinations in under a minute.
Pass three: turn the meeting into a deliverable the client keeps
After the client conversation, you have answers to your diagnostic questions. Pass three converts that into a one-page profit-opportunity memo that’s specific enough to act on.
“I’m writing a one-page profit improvement memo for [CLIENT]. Based on these confirmed findings from our meeting [PASTE NOTES], structure the memo as: a single sentence top-line finding, three opportunities ranked by estimated annual margin impact (with the calculation shown), one risk we should watch, and a 30/60/90 day action list with owners assigned (client vs us). Plain English, no jargon, written for a business owner. Maximum 400 words.”
This is the deliverable clients pay advisory fees for. It used to take a senior accountant two to three hours to draft. With the three-pass workflow, the analytical heavy lifting is done and the writing is mostly editorial polish.
What’s quietly changing about advisory work
The accountants pulling ahead in 2026 aren’t the ones who automated bookkeeping fastest. They’re the ones who used the time AI gave them to do more advisory engagements per partner. Tools like Fathom and Spotlight Reporting have turned monthly management reporting into a near-zero-effort task once the data is mapped. The bottleneck is no longer the report. It’s the conversation around it, and that’s the part clients pay the most for.
The counterintuitive insight worth ending on is this. The accounting AI market is projected to grow from $6.68B in 2025 to $37.6B by 2030, but only 37% of firms invest in AI training (per the AICPA 2025 report). That gap is the opportunity. The firms that train their team on these workflows now don’t need to win a price war on compliance work. They win on the part of the job that was always supposed to be the most valuable, helping clients make more money.