Accounting Prompts

Accounting Prompts That Actually Work (And The Ones That Destroy Your Ledger)

I've been running month-end closes for about a decade, and somewhere around 2023, everything shifted toward AI-assisted workflows. I was skeptical at first, honestly. I had spent years relying on macros, VBA scripts, and manual cross-checks that were slow but predictable. Then a colleague shared some prompts they'd been using for reconciliation and variance analysis, and I decided to test them. Some of them worked great. Others made things worse in ways that took me three weeks to catch. At their core, Accounting Prompts are structured instructions fed into AI systems to handle accounting tasks — journal entries, reconciliation, variance explanations, financial statement summaries, general ledger reviews, and so on. They're not magic. They're basically templated questions designed to extract useful, accurate output from a language model. The key difference between a good prompt and a bad one usually comes down to specificity, context, and formatting constraints you build into the instruction. I keep a folder of prompts I've refined over the last two years. Some are one-liners. Others are multi-step instructions that tell the model exactly what format to use, what accounts to reference, and what kind of errors to flag. Here's how I structure them.

The Prompt Format I Actually Use

The prompts that survive real-world use tend to follow a pattern I've developed through trial and error. Here's a working example I use for bank reconciliation: "You are reviewing a bank reconciliation for Q3 2025. Compare the attached bank statement line items against the general ledger accounts. Identify any transactions that appear on the bank statement but not in the GL. For each discrepancy, classify it as: timing difference, duplicate entry, missing entry, or error. Output results in a table with columns for Transaction Date, Description, Bank Amount, GL Amount, Variance, and Classification. Flag any variance exceeding $500 as 'requires review'." This prompt works because it gives the model a role, a clear task, a classification framework, and a formatting requirement. Without the formatting instruction, the output tends to be messy prose that you have to restructure anyway. The $500 threshold keeps noise out of the important stuff.

For journal entry generation, I use something different. This is my go-to: "Based on the following transaction description, generate the appropriate journal entry with debit and credit accounts, amounts, and a brief narrative. Use standard US GAAP chart of accounts. The transaction: [insert description]. Include any required accruals or adjustments." This has saved me maybe two hours per close cycle. The output isn't perfect — I still review every entry — but the draft quality is usually high enough that I'm editing rather than writing from scratch.

Where These Prompts Fail Hard

I need to be blunt about this because nobody talking about Accounting Prompts seems willing to admit it. They fail in specific, painful ways. First, they hallucinate account numbers. I learned this the hard way in November 2024. I ran a prompt to generate adjusting entries for prepaid insurance across fourteen cost centers. The model produced entries that looked completely normal — correct account structures, reasonable amounts, proper debit-credit balance. But two of the cost centers used account 2180 for prepaid expenses while the other twelve used account 2175. The model defaulted to one account number and applied it everywhere. I caught it during my own variance check because the prepaid balance jumped unexpectedly on one cost center. If I hadn't been doing a line-by-line review, that would have slipped into the financials. The workaround I use now is to include the exact chart of accounts in the prompt context. I paste the relevant account list directly into the instruction, not as an attachment. The model references it much more accurately when it's in the text. It also helps to ask for a validation step where the model confirms each account number against the provided chart before outputting the entry.

Second, they struggle with multi-currency consolidation. I run prompts for intercompany elimination entries across three subsidiaries in different currencies. The math is usually right, but the FX translation logic is occasionally wrong, and the model doesn't always apply the correct spot rate to the right date. I cross-check every single elimination entry now. There's no shortcut around that. Third, and this is the big one — they don't understand your internal policies. If your company has a policy that travel expenses above $2,000 require a specific cost code, the model won't know unless you tell it. I've seen prompts generate entries that technically comply with GAAP but violate internal controls. That's a compliance risk you need to manage yourself.

My Actual Workflow

Here's how I actually use Accounting Prompts in my close process. I don't delegate anything to them blindly. I use them as a first pass, then I review everything. For reconciliation, I run the prompt, get the output, and then I match every flagged item against supporting documentation. The prompt identifies candidates. I confirm them. For journal entries, I feed the transaction description and get a draft entry. I then verify the account codes, the amounts, and the narrative against our policy manual. If something looks off, I rewrite the prompt with more context and run it again. Usually the second pass is better.

For variance analysis, I use prompts to explain deviations. I'll paste the budget versus actuals and ask the model to identify the top three variances and suggest possible causes. The suggestions are often reasonable starting points for my own investigation. I've had prompts correctly flag that a software expense spike was due to a license renewal I'd forgotten about. That one was genuinely useful. I spend roughly 15 minutes per close cycle running and reviewing prompts, compared to maybe 45 minutes doing the same work manually. That's a meaningful difference when you're closing multiple entities.

Prompts I Keep in My Library

Here are the ones I come back to most often. I've included the ones that actually work, not the theoretical ones that sound good in a blog post. Accruals calculation prompt: "Calculate the monthly accrual for [expense category] based on the following data: [insert data]. Use the straight-line method unless otherwise specified. Show the calculation steps and output the journal entry in standard format with account codes, debit/credit amounts, and a narrative that includes the accrual period and basis."

Depreciation schedule review prompt: "Review the attached fixed asset depreciation schedule for errors. Check that: (1) accumulated depreciation matches the sum of monthly depreciation amounts, (2) useful lives are consistent with our policy table, (3) no assets are depreciated beyond their salvage value, and (4) disposals are properly reflected. Flag any issues and suggest corrections." Revenue recognition check prompt:

"Review the following revenue transactions against ASC 606 criteria. For each transaction, determine whether revenue should be recognized at a point in time or over time, and identify any potential performance obligations that may not have been properly allocated. Output a summary table with transaction ID, customer, amount, recognition method, and any flags." GL health check prompt: "Scan the general ledger for the following period and identify: (1) accounts with zero activity that should have activity, (2) accounts with unusually large balances compared to the prior period, (3) any negative balances in expense accounts, (4) transactions posted without a cost center, and (5) any journal entries dated outside the current period that haven't been reversed. Present findings in a prioritized list."

Bottom Line

Accounting Prompts are a tool, not a replacement for judgment. The ones that work are the ones you refine through repeated use and failure. You'll make mistakes. I made a bunch of them in the first six months. The prompts that survive are the ones you've tested against real data and corrected when they produced wrong output. If you're just starting with this, I'd recommend beginning with reconciliation and variance analysis. Those tasks have clear success criteria — the output is either right or it isn't. Journal entry generation is trickier because the stakes are higher and the errors are harder to spot. Spend more time reviewing those outputs before they hit the ledger. The biggest lesson I've learned is that the prompt is only as good as the context you give it. The more specific information you include — chart of accounts, policy rules, thresholds, formatting requirements — the more reliable the output. Generic prompts produce generic, often wrong, results. Specific prompts require more setup upfront but save you time in review.

I still do the work myself. The prompts just get me to the answer faster, and they catch patterns I might have missed after staring at a spreadsheet for six hours. That's about all you can reasonably ask for.