Financial data that stays out of the model

Finance work is full of data that is valuable to the wrong people. One bank statement holds account numbers, the names of payers and payees, and your organisation's payment patterns. One payroll list holds each person's pay. One unreleased profit and loss statement can move decisions by people outside the organisation. So before you use AI for finance work, first set what never goes into a general AI service, meaning a service you open with a personal or free account that is not covered by your organisation's contract.

The basic list is short and easy to remember:

• Bank account and card numbers, belonging to the organisation, staff, customers or vendors.
• Tax ids (NPWP) and national ids (NIK), including those printed on invoices, withholding slips or vendor registration forms.
• Payroll data: pay per person, deductions, allowances and each employee's income tax detail.
• Customer and vendor financial data: receivable balances per customer, payment history, contract prices, credit limits.
• Unreleased results: monthly reports, projections, budgets and performance figures before they are approved and announced.
• Passwords, internet banking access codes and one-time codes from SMS, an app or a bank security device, in any form.

Part of this list is personal data, such as pay, national ids and accounts held by people. Its protection is governed by the PDP Law (UU PDP), explained in the PDP Law basics lesson. Part of it is not personal data but an organisational secret: a vendor's contract price or this month's profit. Both follow the same rule here: they stay out of general services.

The good news is that almost all useful AI help for finance work needs no real data. The model does not need an account number to suggest how to match bank lines with the cash book. It only needs the structure: column names, date format, the kinds of transactions that usually appear, and two or three made-up sample rows. Nor does it need anyone's pay to write an overtime recap formula or a draft notice about a change to the payroll schedule.

So make three moves a habit before pasting anything:

• Replace real data with structure. Write "column A date, column B description, column C debit, column D credit" instead of pasting the statement.
• Make made-up samples. An invented vendor such as "PT Contoh Jaya", codes such as INV-0142, round figures that are clearly invented. Made-up samples also make the result easier to check, because you already know the answer.
• Run it yourself on the real data. The formula, matching steps or template the model gives you, you apply in your organisation's spreadsheet or accounting system, on your computer.

Masking data with initials or removing names is often not enough. Unique amounts, dates and transfer descriptions can identify the person or vendor. How to mask properly, including the difference between pseudonymisation and anonymisation, is in the masking data lesson.

Some work does need real data: summarising a hundred vendor invoices, matching thousands of statement lines, or reading a contract with prices. For such work the question is not "may I use AI or not" but "in which tool". An approved tool is one your organisation chose and governs: a contract covers how data is stored and used, settings prevent the data from training the model, and someone is responsible for it. If no such tool exists yet, work with real data is done without AI for now, or with AI only on structure and made-up samples.

Ask three things before using any tool for real data: who approved this tool for financial data, where the data is processed and how long it is kept, and who can see the chats or files you upload. If you cannot answer all three, treat the tool as a general service.