Twelve Conversations That Do the Work of an Auditor: AI Prompts for Reconciling Acts, Catching Discrepancies, and Surviving a Tax Audit

Every accountant knows the ritual: a folder of counterparty acts arrives by email, each PDF a slightly different layout, and someone has to match every line against the ledger before the month closes. Miss one discrepancy and it surfaces months later — usually during a tax review, when nobody has time to reconstruct the paper trail.

The problem is not that accountants are careless. It is that reconciliation is repetitive pattern-matching, and human attention degrades after the fortieth invoice. LLMs are unusually good at exactly this: reading messy text, comparing two lists, flagging mismatches, and drafting the follow-up letter. They are not a replacement for your judgment — they are a tireless first pass.

Below are twelve complete prompts organized by difficulty. Each one is a working dialogue template: copy it, fill in the brackets, and iterate. Treat the output as a draft to review, never as a filed document.

Before You Start: The Ground Rules

Three rules keep AI-assisted bookkeeping defensible:

1. Never upload raw personal data or full bank statements to a public model without a data-processing agreement. Mask names and account numbers (e.g., LLC ***, acct ****1234).
2. Always demand a source citation — the model must point to the exact line it compared, not just say "mismatch."
3. Keep the prompt and the output in the working file. If a tax inspector asks how a figure was derived, you have the audit trail.

These map to the principle of documentary evidence in tax accounting: every entry must be traceable to a primary document. AI is a tool inside that chain, not outside it.

Basic Level: Getting Clean Comparisons

1. The Act-to-Ledger Line Matcher

Task: compare a counterparty act against your internal ledger and produce a line-by-line match table.

You are an accounting reconciliation assistant. I will give you two lists:
(A) lines from a counterparty act, (B) lines from our internal ledger.
For each line in A, find the best match in B by: document number, date, amount, and description.
Output a Markdown table with columns: Act line

| Ledger line | Match status | Difference | Note.
Match status must be exactly one of: MATCH, AMOUNT MISMATCH, MISSING IN LEDGER, MISSING IN ACT, AMBIGUOUS.
Do not guess. If two ledger lines could match, mark AMBIGUOUS and list both.
Act:
[PASTE]
Ledger:
[PASTE]

Example result: a table where line 4 reads `AMOUNT MISMATCH

| 12,400.00 vs 12,040.00 | transposition suspected` — the kind of digit-swap a human eye skips at 7 p.m.

2. The Digit-Transposition Hunter

Task: catch the classic accounting error where two digits are swapped.

Compare these two amounts and tell me if the difference is consistent with a digit transposition (e.g. 1240 vs 1204), a decimal shift, or an unrelated discrepancy. Explain your reasoning in one sentence.
Amount A: [X]
Amount B: [Y]

Example result: "12,400 vs 12,040 — the difference is 360, consistent with swapping the tens and units digits. Likely a data-entry transposition, not a pricing dispute."

3. The Missing-Act Detector

Task: given a list of expected acts for the period, identify which never arrived.

Here is the list of counterparties and expected act numbers for [month]. Here is the list of acts actually received. Return a bulleted list of acts that are missing, sorted by counterparty name. Do not invent act numbers.
Expected: [PASTE]
Received: [PASTE]

Example result: three counterparties flagged as "act not received," which is exactly the list you email before closing the period.

Advanced Level: Reasoning Across Documents

4. The Contract-Terms Cross-Checker

Task: verify that an act's amounts comply with the contract's pricing terms.

I will paste (1) key pricing clauses from a contract and (2) the line items of an act. Check each act line against the contract terms: unit price, quantity limits, VAT rate.
For every violation, quote the exact contract clause and the exact act line. If no violation, say COMPLIANT.
Contract clauses:
[PASTE]
Act lines:
[PASTE]

Example result: "Act line 2 charges 20% VAT; clause 4.2 specifies 10% for this category. Violation." This is the check that saves you from overpaying VAT you cannot recover.

5. The Discrepancy Narrative Builder

Task: turn a list of mismatches into a formal letter to the counterparty.

Here are the discrepancies found during reconciliation: [LIST]. Write a formal business letter to the counterparty requesting corrected acts. Tone: polite, factual, no accusations. Include: reference to the reconciliation period, a numbered list of discrepancies with amounts, and a requested response deadline of [N] business days. Keep it under 250 words.

Example result: a ready-to-send letter with a clean numbered list — the model does the tedious drafting, you do the sign-off.

6. The VAT Consistency Auditor

Task: check that VAT amounts are internally consistent across a batch of documents.

For each row below, verify: net amount + VAT = gross amount, and VAT rate matches the stated category. Flag any row where arithmetic does not close. Output only the flagged rows with the arithmetic shown.
Data:
[PASTE TABLE: net, vat, gross, rate]

Example result: flagged rows where net + VAT differs from gross by a few kopecks — trivial alone, suspicious in aggregate.

Expert Level: Preparing for a Tax Review

7. The Document-Trail Reconstructor

Task: given a transaction, assemble the chain of supporting documents.

For the transaction described below, list every primary document that should exist to support it under standard tax documentation practice: contract, act, invoice, payment confirmation, and any sector-specific document. For each, state what it must contain. Then tell me which of the documents I list as "on hand" are missing.
Transaction: [DESCRIBE]
On hand: [LIST]

Example result: a checklist revealing that the payment confirmation is missing — the gap an inspector finds first.

8. The Inspector-Question Simulator

Task: anticipate the questions a tax inspector would ask about a specific transaction.

Act as an experienced tax inspector reviewing the transaction below. Ask the five most likely questions you would pose, ordered from most to least probing. For each question, note what document would answer it. Do not give legal advice; frame as questions only.
Transaction summary: [PASTE]

Example result: five pointed questions, each paired with the document that resolves it — a rehearsal before the real meeting.

9. The Period-Close Consistency Report

Task: summarize reconciliation health for a month in one page.

Based on the reconciliation data below, write a one-page period-close report: total acts received, total matched, total mismatched, total missing, and the three largest discrepancies by amount. Use a Markdown table for the summary and a short bulleted "actions required" list. Do not invent figures not present in the data.
Data: [PASTE]

Example result: a management-ready summary that takes the model seconds and you would otherwise spend an hour assembling.

What These Prompts Will Not Do

Be honest about the boundary. An LLM cannot verify a signature, confirm that a counterparty is registered, or replace a statutory audit. It reads text and reasons over it — nothing more.

Use the table as a division of labor, not a limitation. The repetitive layer is where AI earns its keep; the judgment layer stays with you.

The Result: Hours Back, Fewer Surprises

The pattern across all twelve prompts is the same: structure the input, demand citations, review the flags. Do that and reconciliation stops being a late-night ritual and becomes a routine review of a short exception list.

Start with the three basic prompts this week. If they save you an evening, the advanced ones will save you a month-end. And when the tax review letter arrives, you will already have the document trail assembled — because you built it while closing the period, not after the knock on the door.