Remove the client's name, paste the general ledger, ask for anomalies. Except the narratives give everything away. Answers to the questions a firm asks before handing an accounting export to AI.
It is a familiar scene at any firm during year-end close. A team member exports a small business's general ledger, deletes the column with the company name, and pastes the four thousand lines into a consumer AI tool with a single instruction: "find the anomalies". They feel they have anonymised it.
Then you read the export again. In the narrative column: "Salary transfer Martin S.", "Expense refund Dupont", "Advance on maintenance — agreed by MD", the name of every supplier, a cheque reference. The customer and supplier accounts carry the names of the business's own clients. The director's loan account points to a person and their assets. None of it was removed, and all of it went out.
Because an accounting export says far more than its header. Entry narratives are typed by hand, often with a proper name; subsidiary ledger accounts carry the names of customers and suppliers; payroll and shareholder accounts refer to individuals. Even without a single name, a turnover, a sector and two large customers are enough to recognise a small business in its town.
Removing a column is at best pseudonymisation, not anonymisation: our guide on anonymisation and pseudonymisation explains the difference. For an export, the practical rule is simple: if you would not send it as it is to a colleague at another firm, do not send it to a consumer AI tool.
The same as about a balance sheet, in more detail. In many countries, accountants are bound by professional secrecy, and breaching it can be a criminal offence; in Switzerland, auditors are covered by Article 321 of the Swiss Criminal Code, and the fiduciary who keeps the books owes the client discretion and compliance with the FADP. On top of that comes the personal data a general ledger always contains: salaries, expense claims, sometimes a wage garnishment. In the European Union, the GDPR applies to all of it.
A ledger export has intended recipients: the client, the auditor, the tax authority when it asks. A consumer AI tool is not one of them: it is a third party, which may keep what it is given and fall under foreign law that gives its authorities access. Your engagement letter and your firm's policy may also set their own rules: check them.
No, and that is the second mistake, less visible than the first. A language model reads and writes; it does not calculate reliably over thousands of lines. It can state a total of expenses that is wrong to the cent, with full confidence, or "find" a difference that does not exist. Totals, balances and reconciliations remain the job of your accounting software or your spreadsheet.
What AI does well is help you build those calculations: write the spreadsheet formula that isolates duplicates, suggest the pivot table that breaks expenses down by month, explain why a column of amounts is being read as text. You run the calculation; it helps you design it and read the result.
Plausibility checks, the ones you do by eye while leafing through and never have time to do on the whole file. Spotting entries with no narrative or a vague one, repeated round amounts, entries posted on a Sunday or on the last day of the financial year, a supplier with a debit balance, a suspense account that never clears, two invoices from the same party for the same amount a few days apart. It flags, you investigate.
Here is an extract of a general ledger in CSV format (date, journal, account, narrative, debit, credit), closed financial year, third-party names replaced with codes. Do not calculate any totals. List the entries that deserve a check: likely duplicates, empty or vague narratives, repeated round amounts, entries dated on a weekend or on the last day of the financial year, customer or supplier accounts with an unusual balance direction. For each entry you flag, give its exact date, account and amount, and say in one sentence why you are flagging it.
Asking for the exact references of every flagged line is not a detail: it lets you find the entry in your software in ten seconds, and see straight away if the AI has invented a line.
Rarely. A CSV export is a text file: it can be attached to a message like a document. A spreadsheet workbook, on the other hand, is not read as it is; save it as CSV first. With a very large file, the amount of text the assistant can take into account depends on your plan: beyond that, it will only see part of the export, and its conclusions will only hold for that part.
So it is better to split. One journal rather than all of them, one quarter rather than the whole year, one class of accounts at a time. And for overall questions, the trial balance is often enough where you would have sent the general ledger: fewer lines, fewer narratives, fewer names.
In IA Confidential, questions are handled by default by confidential models: open models installed on servers in Switzerland and Europe, which keep nothing of your conversations. For an export, stay there: the file has no reason to go anywhere else, and these models are all you need for plausibility checks.
If you choose an anonymised model, the Confidentiality filter steps in before anything is sent. It detects sensitive data and offers to answer with a confidential model, message unchanged, to send an anonymised version that you review, or to send it as it is if you accept that. For an attachment, only pseudonymised extracts are sent; if the document is judged too identifying, the assistant tells you and offers to have it handled by a confidential model, or to send only the question. In the Privacy panel, specify what must always be masked: your clients' names, and those of their main customers and suppliers.
Whatever adds nothing to the question. For a duplicate check, detailed narratives and document numbers are enough; employees' names, bank details and free-text comments serve no purpose. Replacing subsidiary accounts with neutral codes (CUS-01, SUP-07) takes two minutes in a spreadsheet and keeps the export usable. Even on a confidential model, sending less remains the best habit.
With one space per client or per engagement. In the space's instructions, describe your framework once: the chart of accounts used, the currency, the CSV separator, the year-end date, and the rule "do not calculate any totals, flag". These instructions take precedence over your general preferences, and every conversation in the space picks them up without you having to copy them again.
On a computer, with Chrome, Edge or Opera, you can also add to the space the folder where you keep this client's exports. Tick the "Confidential models" option: the folder is then never sent to an anonymised model. Without it, the relevant passages would be sent as they are if you chose an anonymised model. The assistant then searches the folder for relevant passages on its own; the index stays on your computer.
Yes, and that is often where it saves the most time. Once your checks are done, it turns a list of doubtful entries into a clear letter that the director can understand without being an accountant. The assistant writes it as a document: you edit it, print it as a PDF or export it to Word or OpenDocument, with a fresh layout to adapt to your firm's template.
Using the entries flagged above and my notes, write a request for supporting documents for the director: one introductory sentence with no reproach, then a table (date, amount, what we do not understand, document expected), sorted by amount in descending order. End with the date by which we need the answers to meet the year-end close. Friendly tone, no technical term without an explanation.
That will happen often, and it is normal: it knows nothing of the client's business, their posting habits or the arrangement agreed with a supplier. Conversely, it will let through errors that a trained eye would have caught. Treat its list like that of an attentive trainee: a lead to investigate, never a conclusion. What goes to the client and what appears in the accounts remains your responsibility.
With a confidential model, the export, your questions and the answers are processed on servers in Switzerland and Europe, and nothing is kept on our side. Attachments are read in your browser, no file is stored on our servers, and your conversation history stays in your browser.
With an anonymised model, what is sent depends on your decision, once the filter has run: an anonymised version, pseudonymised extracts, or the message as it is if you choose so. A folder added with the "Confidential models" option is never sent to an anonymised model.
On a shared computer at the firm, the history can be read by anyone using the same browser session: keep a personal session and close it when you leave. And the original export stays where you filed it: it does not need to live in the chat once the check is done.
For reporting back, our guide on explaining a balance sheet to a client takes over. Before attaching your next file, reread the list of data you should never paste into an AI.
To try it on a fictitious extract of a few dozen lines: open the chat.
