Adding Zeroes

By · · 5 min read

How do you bring AI into finance without exposing confidential data?

Give the AI files, not logins. Save the finance reports you need as files, put them in one folder, and let the AI read that folder.

Give the AI files, not logins.

Save the finance reports you need as files, put them in one folder, and let the AI read that folder. A file is a copy, so the AI can read it and cannot change your books through it. People call that "read-only," and it is the whole safety idea. Your bank login, card numbers, payroll and tax IDs stay with you. The AI writes the weekly summary from the copies.

That is the whole idea. The rest of this page shows it with a made-up brand, so every figure below is illustrative.

Which finance data should never reach an AI tool?

Four things stay out, whatever the tool promises: logins and passwords, bank and card account numbers, payroll and personal data about your staff, and tax IDs.

A weekly summary never needs them. If the AI cannot see them, it cannot leak them.

What a summary does need is totals. Sales by channel, spend by category, cash in, cash out, what customers still owe you. Totals are enough.

What does a read-only export setup look like?

Take a made-up snack brand, Fjord Snacks, at about $4M a year.

Before: the founder shares her accounting login with an AI app "so it can see everything." The app now holds a key to the whole building. If the app is hacked, or its settings are wrong, everything behind that key is open.

After: three files land in one folder every Monday at 6:00 AM. A sales file from the online store, which lists what sold and where. A profit-and-loss report from the accounting software, which shows money in against money out. A bank summary with totals by category and the account numbers removed. Each one is a scheduled report, which means the software builds it and files it on its own. Whoever keeps your books sets that up once. If that is nobody, it is the first job of a Reporting Sprint. No password changes hands.

Someone who got into that folder would find last week's totals. They would not find a way into the bank. That is the difference a read-only copy makes.

I wrote about the files themselves in the Monday revenue question, answered from two exports. The setup above is how you feed them to an AI safely.

How do you check where a vendor stores and uses your data?

In the older days you would read the vendor's privacy policy and terms line by line. You would mark where your data is stored, who else can see it, and whether the vendor keeps your uploads and uses them to improve its own AI product. Then you would email their sales rep about every vague clause and wait a week for an answer. That is an afternoon of reading plus a week of waiting.

Today, give this to Claude, an AI assistant you type instructions to, and you have your answer in thirty seconds. (Those typed instructions are called a prompt; you can paste this one as it is.) Paste the vendor's privacy policy and terms underneath it.

Below is the privacy policy and terms of an AI tool I am thinking of using with my company's finance reports. Answer these five questions in plain English. For each one, quote the exact sentence from the text that supports your answer. If the text does not say, write "not stated" and do not guess.

1. Where is my data stored, and in which country?
2. Does the vendor keep what I upload and use it to improve its own AI?
3. Who else can see my data, including other companies the vendor shares it with?
4. How long is my data kept, and can I have it deleted?
5. Can I switch off storage of my files?

Then write the email I should send the vendor to get every "not stated" answer in writing, ready to send.

Whatever the AI says, the vendor's written reply to that email is the answer that counts.

What can the AI do with the data once it has it?

It can read, add up, compare and write. That covers most of a weekly finance summary.

It should write back to one place only: a folder called Summaries. It does not post entries to your books, pay a bill, or change a number. If it cannot change anything, a wrong answer costs you a bad paragraph, not a bad payment.

In the older days you would open the sales file, the profit-and-loss report and the bank statement side by side. You would copy totals into a sheet, check them against each other, and write three paragraphs for yourself or your partner. That is a Monday morning, every week.

Today the same job takes the AI half a minute. Attach the three files from the folder, paste this, and read the result with your coffee.

Three files are attached for last week: a sales file, a profit-and-loss report and a bank summary. Write a one-page summary for me in plain English, under 250 words.

Start with one sentence: did we make or lose money last week, and by how much. Then give sales by channel, the three biggest costs, and cash in versus cash out.

Flag anything that moved more than 15 percent against the prior week, if the files show it. Flag anything that looks like a mistake, such as the same number appearing twice. If two files disagree, say which two and by how much. Do not pick a winner.

If any file is missing or looks empty, say so first and stop. Use only the attached files. If a number is missing, write "not in the files." Do not ask me for logins or account numbers.

If the two files disagree, the AI reports the gap and stops. Which number is right is a different problem, and your weekly meeting should not start with it. When the summary matters, let a second AI read it against the three files and say what it cannot support. I covered that in the AI that audits the AI.

CPG AI Assessment

Fix the CPG workflow that keeps landing back on your desk.

I map the recurring reporting or operating workflow first, then build only when the first useful fix is clear.