Five accounting tasks AI does faster, and one you must never hand over

The prompt patterns below work in Claude, ChatGPT, Copilot and Gemini. Before using them on real documents: a business account, and personal and client data replaced with placeholders. How to do that in half a minute is shown in the opening module.

1. Reading data from a batch of invoices or a bank statement

Ten PDF invoices in ten layouts. Instead of retyping, you attach them and ask for a table.

From the attached invoices, produce a table with the columns: invoice number, issue date, supplier, supplier VAT number, net, VAT, gross, due date. If a value is not on the document, write “missing”; do not guess. At the end, give the totals of the net, VAT and gross columns.

What to check: the gross total in the table against one calculated on the side, two random invoices field by field, every “missing”. AI slips most often on credit notes and multiple VAT rates on one document.

2. Reconciling two lists that should match

A bank statement and the ledger, a sales register and a client’s file. Two lists, a few hundred rows, something does not add up.

Below are two payment lists: A (from the bank statement) and B (from the ledger). Match items by amount and counterparty, with a tolerance of 1. List separately: items only in A, items only in B, pairs with a difference in amount. Do not match items unless you are sure; mark those as “to check”.

What to check: the row count before and after (nothing lost), every “to check” pair. Above a few hundred rows, ask AI for a formula rather than a comparison “in its head”.

3. The spreadsheet formula nobody remembers

I have a sheet: column A is the date, B the counterparty, C the gross amount, D the status (“paid”/“open”). Write a Google Sheets formula that sums the open amounts for the counterparty in cell F1, only where the due date is more than 30 days before today. Explain the formula in two sentences.

What to check: the result on three rows calculated by hand. Whether the formula assumes your date format and works in your version (Excel and Sheets differ). Only then copy it down the column.

4. A draft letter to a client or the tax office

A reminder about documents, a reply to a client’s question, a letter of explanation. A draft in a minute, corrections in five.

Write a polite, short letter to a client of an accountancy practice reminding them that we need their purchase invoices for the previous month by the 10th, and that missing documents will delay the filing. Tone: businesslike, no threats. Leave placeholders in square brackets for the client name, month and deadline. At most 120 words.

What to check: every date, amount and legal reference, if AI added any. Models happily add section numbers that do not exist. If the letter needs a legal basis, you type it in yourself.

5. A summary of a regulatory change for clients

A long circular or ruling, and clients want to know what it changes for them.

Below is the text of a notice about a regulatory change. Summarise it in three points from the perspective of a small services company: what changes, from when, what needs doing. Quote the relevant passage of the notice under each point so I can verify it. Do not add information from outside the text.

What to check: every quote against the original. Asking for quotes is the simplest way to keep a model to the source. The summary goes to clients only after you have read the notice yourself.

The one task AI does not get

The decision. Which VAT rate, whether an expense is deductible, how to book an unusual transaction, what to tell the tax office in a disputed case. AI can prepare the material: gather the rules, show the options, point out what to check. The accountant makes and signs the decision: she is responsible for it and sees what is not in the documents.

A language model does not know whether it is right. It answers just as confidently when it is wrong. In accounting, the price of such a mistake is specific and has a due date.

The common thread

Task	AI does	The human does
Invoices into a table	Transcribes	Checks totals and random items
Two lists	Matches	Resolves uncertain pairs
Formula	Writes	Tests on three rows
Letter	Drafts	Checks dates and references, sends
Summary	Shortens, with quotes	Reads the original, approves
Tax decision	—	Makes it and signs

These five tasks are the skeleton of the accounting module: in the workshop we do them on your own anonymised documents and finish with a list of what else can be shortened the same way.