Is It Safe to Put Borrower Financials Into ChatGPT or Claude?

1. The Short Answer

It can be safe to put borrower financials into ChatGPT or Claude, on a business plan, once you have read the confidentiality clause in each credit agreement and written your own rules. On a personal account, the answer is no.

For a lender of 10 to 40 people, three things decide it:

• The plan. Business plans, the versions sold to companies, do not use your data to train the vendor's models by default. Personal plans do until each user switches a setting off.
• The confidentiality clause. Each credit agreement lists who may receive the borrower's information, and your counsel decides whether an AI vendor is on that list.
• Your own rules. One page on who may upload which documents.

This guide is operational guidance, not legal advice.

2. What Each Plan Does With Your Data

The plan is the part you decide alone.

Anthropic's training page says that by default it "will not use your inputs or outputs from our commercial products" to train its models, and OpenAI's data controls page says the same of ChatGPT Business and Enterprise. One exception on Claude: a thumbs-up or thumbs-down rating sends that chat to Anthropic, which can keep it for up to five years and train on it. An owner can switch ratings off.

A promise not to train still leaves your file on the vendor's servers. When you delete a chat, Anthropic and OpenAI each say they remove it within 30 days.

The business contracts add a confidentiality promise. Anthropic's commercial terms say, "Customer Content is Customer's Confidential Information," and OpenAI's services agreement has confidentiality terms of its own. Have counsel confirm which contract covers the plan you buy.

Cost is no reason to stay on personal accounts. Claude Team and ChatGPT Business each cost $20 a seat a month billed annually: $2,400 a year for 10 people and $9,600 for 40.

3. What Your Credit Agreements Say

The plan settles what the vendor does with a file. The credit agreement settles whether you may hand it over, in its confidentiality clause.

Three credit agreements on file with the SEC show how such a clause reads: Loar Holdings' (a private credit firm, First Eagle Alternative Credit, is the agent), Blue Owl Capital Corporation's revolving facility (a credit fund is the borrower) and Luxfer Holdings' English-law facility. Yours will have their own wording.

In all three, the lender may pass the borrower's information to its own directors, officers and employees, and to its agents or advisors, provided they are told it is confidential. In the Loar agreement, each recipient "will be informed of the confidential nature of the Information and instructed to keep such information confidential."

The two US agreements name service providers only for "the existence of this Agreement" and information about it, with no mention of the borrower's financial statements.

None of the three mentions AI. Three full-text searches of SEC filings from 2024 to November 2026 found no credit agreement whose confidentiality clause does, which is no proof that none exists.

The closest published analysis comes from law firms writing about confidentiality agreements (NDAs) in company sales; none found for this guide addresses lenders.

• Baker Donelson (August 2026): "Uploading a target's data room contents into a third-party AI tool may constitute disclosure to a third party in breach of the NDA, particularly if the vendor's terms permit customer inputs to be used to train or improve its models."
• KJK (March 2026): "An AI platform provider may not qualify as a permitted representative under the NDA."
• Pillsbury (November 2026) advises treating an upload to a third party's AI system as a disclosure.

So for each agreement, and for the NDA on any deal you are still looking at, ask counsel whether a business AI vendor, under contract and bound to confidentiality, counts as a permitted recipient. Where the answer is no or unclear, ask the borrower: the Loar clause also permits disclosure to anyone the borrower consents to.

On new loans, counsel can settle it in the drafting: Baker Donelson's advice for NDAs is to "address permitted AI use expressly."

4. Information That Needs Extra Care

Material non-public information (MNPI). This is information the market does not have that would matter to someone trading a company's securities, and a lender can come to hold it. The SEC checks the written policies that guard it: in 2024, two much larger credit managers settled such cases, Sound Point Capital Management for $1.8 million and Marathon Asset Management for $1.5 million.

Neither case involved AI, and no SEC statement found for this guide mentions MNPI and AI tools together. The working rule: nothing about a company on your restricted list (the companies whose securities your firm may not trade) goes into an AI tool.

Personal data. Regulation S-P, the SEC's rule on protecting customer information, covers personal information about individual customers, so a borrower company's financial statements sit mostly outside it. The part to watch is personal data inside the file: a guarantor's personal financial statement or tax return, a payroll list.

SEC-registered advisers with $1.5 billion or more in assets had to comply with the rule's 2024 amendments by December 3, 2025, and smaller ones by June 3, 2026. Your counsel can confirm how it applies to you.

5. Five Rules for a Small Lender

Put these five rules in place while counsel reads the clauses.

1. Business plan only, with the terms on file. Borrower packages go into the firm's business-plan account and nowhere else. Save the vendor's contract and training page with the date you read them, and switch off ratings on Claude.
2. Named people only. List who may upload borrower documents; the plan's owner adds and removes them.
3. Names off for exploratory work. Anyone trying the tool out or learning it uses a closed deal with the borrower's name removed.
4. Nothing from the restricted list, and no personal data. Guarantor statements, tax returns and payroll lists stay out until counsel says otherwise.
5. Write it down. One page: the tool, the people, the documents, who checks the clause on each new loan and who to tell when something goes in by mistake.

Accuracy is a separate question: see Can ChatGPT or Claude Spread Borrower Financials?

6. Where to Start

1. Ask who uses a personal account for work today, without blame. Shadow AI at Investment Firms covers the fix.
2. Buy one business plan for everyone who handles borrower documents. Who sets it up is in Can a Private Credit Firm Use AI Without an IT Team?
3. Send counsel the confidentiality clause from each credit agreement, with the vendor's contract.
4. Write the one-page rule and have each named person sign it.
5. Keep the rule, the vendor's terms and counsel's answer in one folder, ready for an LP's questionnaire or an SEC exam.