Want to Summarize Your Company's Financial Data in ChatGPT? Is It Dangerous?

More and more people are pasting financial data into ChatGPT to get an instant summary: sales reports, bank statements, receivables lists, sometimes even payslips. It's fast and convenient. But the same question is always in the back of your mind: is this actually safe?

The honest answer depends on three things: which ChatGPT plan you use, what kind of data it is, and your team's habits. I'll go through them one by one, and at the end there's an approach that makes sense to me for getting started.

First, what happens to the data you paste

Personal plans (Free, Plus, Pro)

In OpenAI's Terms of Use, effective January 1, 2026, OpenAI says it may use your content to provide, maintain, develop, and improve its services, and that you can choose to keep your content from being used to train models. The setting is in Data controls, specifically Improve the model for everyone.

A few details from that help page that are easy to miss:

• Turning that setting off does not delete chats that are already saved. They stay in your history.
• Temporary Chat doesn't appear in your history and isn't used to train models, but it can be kept for up to 30 days for safety purposes.
• If you give a thumbs up or thumbs down on an answer, the whole conversation tied to that feedback can be used to train models, even if you've opted out.

Business plans (ChatGPT Business, Enterprise, and the API)

OpenAI's Enterprise privacy page (updated January 8, 2026) says business data isn't used to train models by default, unless you opt in through the feedback mechanism. Data is encrypted (AES-256 at rest, TLS 1.2 or higher in transit), the service has passed a SOC 2 Type 2 audit, and companies can sign a Data Processing Addendum. Workspace admins can control how long data is retained.

That doesn't mean nobody else can ever see it. For ChatGPT Business, OpenAI says access to conversations is limited to authorized employees and third-party contractors bound by confidentiality, solely to investigate abuse. For the API, inputs and outputs can be kept for up to 30 days for abuse monitoring, with a zero data retention option for certain endpoints if you qualify.

The biggest difference isn't "ChatGPT or not". It's a personal account versus a workspace managed by the company. And this is only OpenAI's story. Other AI services have their own rules, so check per service and per plan.

The danger isn't only about model training

1. Personal accounts outside company control

LayerX, a browser security company, reported from the browser telemetry of its enterprise customers that 77% of employees paste data into GenAI prompts, and 82% of those pastes come from personal accounts. Treat this as a picture, not a general figure: the source is a security vendor, and its blog doesn't explain the sample size.

Employees who paste data into GenAI prompts, according to LayerX

77%

Of those pastes come through personal accounts

82%

Average pastes into non-corporate accounts, at least 3 of them containing sensitive data

14/day

A personal account also means the data leaves with the employee when they resign, which is exactly the problem I cover in When the Company's AI Resigns Along With the Employee.

2. Data that shouldn't leave

In April 2023, Samsung engineers were reported to have accidentally leaked internal source code by uploading it to ChatGPT. In a memo dated May 1, 2023, Samsung banned ChatGPT, Google Bard, and Bing on company devices and networks. The reason, according to Bloomberg's report, was that data sent to external AI platforms "is stored on external servers, making it difficult to retrieve and delete, and could end up being disclosed to other users." That isn't financial data, but the pattern is the same: once data leaves, someone else holds the controls.

3. Personal data and Indonesia's PDP Law

A company financial report with revenue, profit, and costs isn't personal data. But once it includes people, such as payslips, employee bank account numbers, or customer invoices with names and national ID numbers (NIK), Indonesia's Personal Data Protection Law (the PDP Law) applies. Law No. 27 of 2022 classifies "personal financial data" as specific personal data (Article 4, paragraph 2, letter f). Processing specific data is on the list of high-risk processing that requires a personal data protection impact assessment (Article 34, paragraph 2, letter b).

If data is sent to a party outside Indonesia, Article 56 requires the receiving country to have an equal or higher level of protection. If that isn't met, there must be adequate and binding protection, and if that's missing too, the data subject's consent is needed. This is an orientation summary, not legal advice.

4. The numbers can be wrong

Safety aside, AI can add things up wrong or pull the wrong figure from a large table. Don't make AI the calculator: compute totals in a spreadsheet, then ask the AI to explain or summarize numbers that are already settled. That habit of reconciling results matters, and I explain why in Your Team Stays Up Every Month-End Compiling Reports: When Is It Worth Automating?.

So, is it allowed? Split your data into three tiers

How to use it more safely

1. Use a company-managed account, meaning a business plan or the API, not an employee's personal account. The simple rules are in A Simple SOP to Keep Company AI a Team Asset, Not Personal Property.
2. Process it in a spreadsheet first, and give the AI only the summary. Calculate totals and subtotals with formulas, then paste a compact table, not thousands of raw rows.
3. Anonymize. Replace customer and vendor names with codes (for example P-001), and strip out national ID numbers, account numbers, and addresses. Keep the code table somewhere separate.
4. If you have to use a personal account for non-sensitive data, turn off Improve the model for everyone and use Temporary Chat. That reduces the risk, it doesn't remove it.
5. Reconcile the numbers. The AI's total must match the spreadsheet. Ask the AI to show its steps or formula, and read them yourself before using the result.
6. Write the rules down. What data is allowed, who may use it, and through which account. Rules that only exist in someone's head won't be followed.
7. For the most sensitive data that still needs AI help, consider a model run on your own server. It has its own costs and limits, which I cover in What Is a VPS for AI, and What Can It Actually Do?.

If you've already pasted something

Stay calm, but don't leave it. Check the plan and account settings, turn off training if it's a personal account, delete the chat, change any credentials that were pasted along, and note down what was pasted for review. Deleting a chat doesn't always mean every trace disappears, because retention and legal obligations can keep data longer. If what was pasted is a large amount of other people's personal data, talk to a legal consultant about your obligations.

Wrapping up

So, is it dangerous? It is if you're careless: personal account, raw data, no rules. It's far safer when the account is company-managed, the data is summarized and anonymized, and there are written rules. And if the data you want to summarize can't leave at all, that's a sign that what you need may not be ChatGPT, but a system that runs on your own infrastructure. For AI risks in companies more broadly, see The Dangers of Using AI in Companies.

If you're thinking about how to use AI safely on financial or other internal data, tell me about your situation first, and I'll help you think it through from the risk and cost side.