Can ChatGPT or Claude Spread Borrower Financials?

1. The Short Answer

Yes. ChatGPT or Claude can spread borrower financials when the statements arrive as clean, text-based PDFs (files where you can select the numbers with your cursor), as long as an analyst checks every figure against its source page.

Scanned pages, restated periods and EBITDA add-backs need the hardest checking, and they are where a second reviewer or a dedicated spreading tool earns its place.

No published, independent test measures either assistant on spreading itself. So a firm of 10 to 40 people has to run its own, on borrower packages its analysts have already spread by hand.

The makers ask for the same care. OpenAI's ChatGPT for Excel help page says: "AI can also make mistakes, so verify the outputs." Anthropic's credit underwriting example, written for commercial bankers, says: "Claude pulls spreads, runs ratios, and structures the writeup while you own the credit decision."

2. What Works and What Breaks

The closest published evidence is a 2024 study in the journal Computers, which ran pages from the annual reports of 46 large listed companies through two older assistants, ChatGPT-4 and Google's Bard. ChatGPT-4 reached "an accuracy of 89.5% when it came to balance sheet questions" but "struggled to indicate EBIT and gross profit with a precision of only 44.2% and 62.8% of correct answers."

For today's versions, the nearest outside measure is the Vals Finance Index, a third-party test of broad finance work that leaves spreading out. Its top-scoring AI model reached 67.05% as of November 9, 2026.

Anthropic's help page says Claude for Excel is "not recommended for" work that includes "Final client deliverables without human review." It adds: "Only use Claude for Excel with trusted spreadsheets." A borrower's workbook is an outside file, so work from the PDF or copy its numbers into your own template.

3. A Spreading Routine That Holds Up

Run it first on three to five borrower packages your analysts spread by hand last quarter, so you already hold the right answers.

1. Give the assistant your template and a mapping note: one page on where your firm puts the awkward lines. A research paper called FinTagging, which tests how AI models match figures to standard accounting labels, found they "generalize well in extraction" but "struggle significantly with fine grained concept linking." They read a number more reliably than they choose its line.
2. Ask for the page, the printed label and the units beside every figure. In the 2024 study, 55 of the 57 partly correct answers shared one fault: "the returned number was a thousand times too small." That is what happens when a figure reported in thousands loses its scale.
3. Tell it to leave a blank and list what it could not place. A guess looks like every other number in the spread.
4. Run the tie-outs before anyone reads a ratio. Assets equal liabilities plus equity. Subtotals add up. Closing cash matches the balance sheet. The prior period matches the spread you already hold, and any difference is a restatement to write down.
5. Let Excel do the arithmetic. The assistant fills the input cells, and your template's own formulas work out the ratios.
6. Have an analyst check every figure against its page and initial the spread. Count the errors by type: units, wrong line, misread digit, missed restatement.

No survey publishes hours per spread for direct lenders, so time each package both ways.

4. Use a Business Plan, Never a Personal Account

Borrower financials go into a business plan only, the version each vendor sells to companies: Claude Team or ChatGPT Business, each $20 a standard seat a month billed annually, or $4,800 a year for 20 people. By default neither Anthropic nor OpenAI trains its models on business-plan data. On personal ChatGPT plans, OpenAI says, "data sharing is enabled by default."

Both Excel add-ins also run on personal plans, so put the rule in writing. The confidentiality clause in each credit agreement comes before any upload: see Is It Safe to Put Borrower Financials Into ChatGPT or Claude?

5. When a Dedicated Spreading Tool Is Worth It

Claude and ChatGPT work one package at a time. A dedicated tool earns its price when:

• most packages arrive as scans, or in a different format every quarter;
• several people and the auditors need one record, with each figure tied to its page;
• the spread has to feed covenant tests and portfolio reviews across the whole book.

Vendors that sell spreading or borrower data intake to private credit funds include, in their own words: Chronograph, which "automates borrower data collection, analytics, valuations, reporting, and information warehousing"; Lumonic, which "reads borrower financial statements, board decks, and compliance certificates in whatever format they arrive"; and Oxane Partners, which lists "Financial Extraction & Spreading with AI."

No vendor we checked publishes a price, so ask each for a quote and a trial on your own borrower packages, scans included. The scorecard is in How Should a Private Credit Firm Choose AI Software?

6. Where to Start

1. Pick three to five borrower packages already spread by hand.
2. Read the confidentiality clause in those credit agreements, then buy business seats.
3. Run the routine in section 3 and compare each AI spread with the hand spread.
4. Count the errors by type and time both versions.
5. Decide after one quarter: keep the assistant with its checks, or ask the spreading vendors for quotes.

The step after a checked spread is the memo draft, covered in Claude for Credit Memos.