Accounting has always been a profession where the numbers are the easy part. Anyone with sufficient training can produce a balance sheet. The actual work, the part that justifies the billing rate and differentiates one firm from another, is the interpretation, the communication, and the documentation that explains what the numbers mean and what a client should do about them.
That distinction matters for understanding where AI fits. AI cannot check arithmetic with the reliability that accounting requires. It does not know your client's chart of accounts, their prior-year audited financials, or the specific tax jurisdiction rules that govern their situation. What it can do is write. And a significant portion of what accountants spend their time on, the narrative sections, the client emails, the audit documentation language, the management letters, is writing that follows predictable structures and eats time that could be spent on analysis.
Narrative Report Writing for QuickBooks and Xero Clients
Many accounting firms send monthly or quarterly reports to bookkeeping and write-up clients that include a narrative section explaining what happened in the financials. These narratives follow a consistent structure: revenue versus prior period or budget, expense categories that moved significantly, cash position, anything notable. The content varies by client; the structure almost never does.
Here is a workflow that works. Export your QuickBooks or Xero comparison report to PDF or copy the key figures into a text summary. Then use a prompt like this:
"Write a two-paragraph narrative for a monthly accounting report for a [type of business] client. Here are the key figures: Revenue was $[X] versus $[Y] prior month and $[Z] same month last year. Gross margin was [X]%. Notable expense changes: [list the categories and amounts that moved meaningfully]. Cash at month end was $[X]. Owner draws were $[X]. Write this in plain language suitable for a business owner who reads the report but doesn't dig into the underlying detail. Tone should be professional but not stiff. Flag anything that looks like it needs attention without being alarmist."
The output will need light editing. AI sometimes overexplains normal fluctuations or underweights genuinely significant movements because it lacks the client context you carry. The edit takes five minutes. Writing from scratch takes twenty. For a practice with 30 write-up clients, this workflow compresses the most tedious part of monthly reporting by several hours.
Variance Explanation Emails to Clients
When a client's financials show something unexpected, the explanation email is one of the highest-stakes communication tasks in accounting. Clients often conflate "the numbers look bad" with "something is wrong with my accountant's work," and the email has to simultaneously explain the accounting, reassure where appropriate, and flag genuine concerns without creating panic.
AI is good at this structure because the framework is consistent even when the specifics vary. Try this approach:
"Write a client email explaining a variance in their financial results. The client is a [type of business] owner. Here is the situation: [describe the variance in plain terms, what moved, by how much, compared to what]. Here is the actual explanation: [the real reason, whether timing, category reclassification, one-time event, or underlying business trend]. The tone should be calm, clear, and direct. Don't be condescending. End with what, if anything, they should do next or what we're watching."
Read the output before sending. Check that the actual explanation is accurate, that the numbers you gave it are correctly represented, and that the tone matches your relationship with this particular client. Adjust accordingly. The value is that you're editing a complete draft rather than staring at a blank email at 4 PM on a Friday.
Excel and Google Sheets Formula Generation
This is one of the highest-ROI uses of AI for accountants and it requires almost no workflow change. When you need a formula you don't have memorized, describe what you want in plain English and ask Claude or ChatGPT to write it.
Examples that come up regularly: a SUMIFS formula pulling from a multi-column transaction log where you need to filter by date range, account category, and department. A nested IF statement for a commission calculation with multiple tiers and a cap. An XLOOKUP across sheets with a fallback value. A formula that calculates days outstanding for an accounts receivable aging report.
Describe the spreadsheet structure, the columns and what they contain, and exactly what you want the formula to return. You'll typically get a working formula on the first try or a formula that's close enough that you can spot the adjustment needed. This is faster than searching Stack Overflow and significantly faster than building complex formulas by trial and error.
One important check: test AI-generated formulas on a small data set before applying them to production data. Formulas that look correct can have edge-case errors, particularly with date math, blank cell handling, and text matching. Verify the output matches what you expect before relying on it for client-facing numbers.
Audit Documentation Language
Audit workpapers require a specific register of language. The documentation has to be precise, describe what was done without overstating scope, note exceptions clearly, and support conclusions in a way that a reviewer (or an inspecting regulator) can follow. Writing workpaper language well is a skill that takes years to develop, and junior staff often produce documentation that senior reviewers have to significantly rework.
AI can help at both ends. For senior accountants, it can generate a first draft of workpaper narrative that you edit for accuracy and precision, saving 15 to 20 minutes per section. For developing junior staff, it can serve as a model: give it your bullet notes from a procedure and ask it to write the workpaper documentation, then show the junior how you've edited it and why. That's a faster training loop than waiting for review comments.
Prompt structure for workpaper documentation:
"Write audit workpaper documentation for the following procedure. Use professional audit documentation language. Structure: what was done, what was tested, what was found, conclusion. Here are my notes from the procedure: [paste your notes]. Don't add conclusions I haven't given you. Flag any places where I need to add specific evidence references."
Management Letter Writing
Management letters are one of the most time-consuming deliverables in audit and review engagements. They have to describe observed deficiencies without overstating their significance, recommend corrective actions that are practical for the client's size and resources, and be written in language that management can act on rather than just file away. Firms that issue thoughtful management letters get more traction on recommendations. Firms that issue boilerplate get letters that go in a drawer.
AI produces good management letter drafts when you give it the right inputs. For each finding, provide: what the deficiency is, what the risk is, what you observed specifically, and what you're recommending. Ask for professional language appropriate for management letter format. Ask it to vary the language across findings so the letter doesn't read as a template with the names changed.
Review every word before it goes out under your signature. AI occasionally softens findings to the point where the significance is unclear, or recommends solutions that are impractical for a five-person operation. Your judgment on severity and practicality is what makes the letter worth something.
What AI Is Unreliable For in Accounting Work
Number checking is the obvious one. AI is not a calculator, and it is not reliable for verifying that a set of financial statements ties or that a calculation is arithmetically correct. It can make arithmetic errors, and it has no mechanism to catch them. Never use AI to verify numbers. Use a spreadsheet.
Tax jurisdiction specificity is the other major reliability gap. Federal tax questions at a general level, AI can orient you reasonably well. State and local tax is where things break down. Nexus rules, apportionment formulas, credit availability, entity-specific elections, these vary enough by state that AI answers in this domain require verification against primary sources before you act on them. The same applies to international tax. The framework AI describes may be directionally correct; the specific treatment may not be.
AI also does not know about recent regulatory changes unless it has been specifically updated. Tax law changes constantly. If you're relying on AI for tax research, treat its output the same way you'd treat a law review article from an uncertain date: useful for understanding the framework, requiring verification before application.
How to Quality-Check AI Output Before It Goes to Clients
The standard that matters is simple: everything that goes to a client under your name is your work product, and you are responsible for its accuracy. AI makes that standard easier to meet by giving you more time to review because you spent less time drafting. It does not change the standard.
For narrative content: read for accuracy (do the numbers match what you gave it), tone (does it match your firm's register and your relationship with this client), and completeness (did it miss anything material you mentioned). This takes two to three minutes for a well-crafted prompt.
For technical content like workpapers or management letters: verify that conclusions stated are conclusions you reached and can support, that evidence references are accurate, and that the language accurately represents scope. AI occasionally expands scope in language ("we verified all transactions" when you sampled 25) or understates what you actually found.
Build a review step into your workflow explicitly. The risk is not that AI produces bad output. The risk is that you're in a hurry and you send AI output with a skim instead of a real read. The hurry is exactly when errors get through. Give yourself the time the review requires.