ChatGPT in an accounting firm: the spreadsheet goes to AI, client data stays in the office

What happens in the films

Both films show a day without the extension, then "take two": the same move with Do Not Feed AI.

On Friday a file with the national ID numbers and bank accounts of 25 employees lands on an AI vendor's server. The firm is the data controller here, so a procedure begins: an entry in the breach register, a risk assessment with the data protection officer, a decision on whether to report, and explanations for the staff. The weekend is gone. In take two the extension stops the file, and the program on Alex's computer lists the detected data. The chat gets a copy with realistic replacements. The answer to who is missing a bank account number comes back with the real names.

On Thursday the data doesn't belong to the firm but to the client. The firm is a processor, so it has to notify Mr. Henry without undue delay, and it is he who assesses the situation and decides on reporting. In take two amounts and dates stay, while detected names, national ID numbers, accounts and tax IDs reach the chat as replacements. The list of overdue invoices comes back with the real names of the counterparties, and the call to the client ends with thanks.

What the program detects in an accounting firm

The detection rules are tuned for Polish data: PESEL (the Polish national ID number), NIP (the Polish tax ID), REGON (the Polish business registry number), Polish IBAN account numbers and addresses. Names, e-mails and phone numbers are also found in English text.

• first names and surnames (inflected Polish forms too), national ID numbers, tax IDs, business registry numbers, bank account numbers, ID card numbers, e-mails, phone numbers, addresses and postal codes with the city;
• company names with a legal form (for example "sp. z o.o.", the Polish limited liability company);
• in xlsx spreadsheets: the text of cells in all sheets, with columns also recognized by their headers (first name, surname, national ID, account, tax ID);
• xlsx, docx, csv and txt files dragged in, pasted or added with the attachment button. The chat gets a clean copy with the content only: cell values (formula results as numbers), without formatting, formulas, charts or macros.

What the program doesn't change, or may miss:

• amounts, dates, deadlines and invoice numbers: they stay on purpose, because the AI works on them;
• a tax ID with a wrong checksum, and company names written without a legal form: add such client names as custom phrases to mask;
• PDF files and photos of scans: they are stopped by default, not masked. Old .xls and .doc files and password-protected files are stopped too.

What the rules say

When the firm sends its own employees' data to a chat, it is the data controller. It records every breach in a register (GDPR Art. 33(5)), assesses the risk and, unless the breach is unlikely to result in a risk to people, reports it to the data protection authority without undue delay and, where feasible, within 72 hours (Art. 33(1)). If the risk is high, it also notifies the employees (Art. 34).

The data of a client's employees and counterparties the firm usually processes as a processor (Art. 28). Then it notifies the client without undue delay (Art. 33(2)), and the client makes the assessment and any report. The deadline and procedure are usually set in the data processing agreement.

GDPR names pseudonymisation as an example of a technical measure (Art. 25 and 32). Data with replacements are still personal data, because the firm can attribute them to people (recital 26). The program reduces the risk but doesn't release you from your obligations. This is a general description of the EU rules, not legal advice, and national laws differ.

Questions and answers