Claude for finance and ChatGPT for finance: what they’re good at, and what changes when they can read your numbers

Search for Claude for finance and the first pages you'll find are the vendors' own. Anthropic's Claude for Financial Services is built for banks, insurers, asset managers and advisors, with connectors to FactSet, S&P Global, LSEG and Morningstar. OpenAI announced ChatGPT for Financial Services on September 10, 2026, starting with investment banking and equity research, with data from Daloopa and PitchBook built in.

Those are real products for people whose numbers live in market data. If you're a controller, an FP&A lead or a fractional CFO, yours don't. They're in QuickBooks, Xero or NetSuite, and in the workbooks built from exports of them. How much finance the assistant knows matters less than whether it can read those numbers.

This guide covers what Claude, ChatGPT, or any MCP client is good at before it sees your data, the three ways to put your numbers in front of one, and the review habit that goes with each.

What the vendors sell

As of September 2026, from each vendor's own page:

• Claude for Financial Services (Anthropic). Pre-built connectors to LSEG, FactSet, S&P Global, Morningstar and others, aimed at pitch materials, credit underwriting and portfolio reporting. In May 2026 Anthropic added agent templates, including a month-end closer and a general ledger reconciler, as plugins for Claude Cowork and Claude Code on paid plans. Anthropic's page, the May announcement.
• ChatGPT for Financial Services (OpenAI). Offered to eligible financial institutions, with Daloopa, PitchBook and LSEG News data built in, S&P Global and FactSet connectors tuned for it, and more than 50 connectors in all. OpenAI's announcement.
• QuickBooks in both (Intuit). A connector in Claude and a plugin in ChatGPT. Intuit's July 2026 update describes invoices, estimates, customers, payroll lookups and lending benchmarks. Intuit's post.

These change every few months, so check the page before you choose.

What it's good at before it sees your numbers

Explaining. How deferred revenue unwinds, what a covenant's EBITDA definition adds back, why a lease sits on the balance sheet. Ask it to name the standard or the clause it's relying on, so you can check it.

Reading a document you give it. Pull the covenants and their definitions out of a loan agreement. List the terms in a lease that matter for the schedule. Find the payment terms in a contract. It reads forty pages faster than you do, and it's reading the actual text.

Writing the words around numbers you supply. Variance commentary, a one-page summary for an owner, the memo that supports a journal entry. The variance commentary skill and the owner summary skill work in any assistant.

Structure. A close checklist, a 13-week cash layout, the outline of a board pack. You'll change it, but you won't start from a blank page.

Where it goes wrong without them

It answers anyway. Ask a model with no data why gross margin fell and you'll get the usual reasons: price, mix, freight, a supplier increase. They read well, and none of them is about your company.

Numbers in prose. A figure it writes into a sentence is a claim, even when it came from a table you pasted. Check it against the table.

Your chart of accounts isn't anyone else's. Freight sits in cost of goods sold at one company and in operating expenses at the next. The model can't know which without seeing the books.

Three ways to put your numbers in front of it

1. Paste or upload an export. This works in every assistant today. It's also a snapshot: the books as of whenever you pulled the report, at the level of detail you pulled. When the answer turns out to depend on which transactions make up a line, you go back and pull another report. You're the integration.

2. Work inside the spreadsheet. Claude, Copilot and ChatGPT all run inside Excel now. That's the right place for a model you're building, and the workbook is still an export. AI in your spreadsheet covers the three.

3. Connect the assistant to the source. Claude, ChatGPT, or any MCP client can take a connector that reads a system directly, so the assistant looks up what it needs instead of working from what you gave it. The QuickBooks connector covers the transactional side. A connector to the general ledger answers the questions a finance lead asks: what posted, why a balance moved, which period is still open.

Public companies are a source too, for when the board asks how your margins compare. RoboSystems SEC, which we build, is a read-only connector to SEC filings, listed in Claude's directory and ChatGPT's. The public peer skill does the comparison with xbrlkit instead, which runs on your own machine and reads only the filings you load.

Three questions, asked two ways

An illustration, on made-up numbers.

"Why did gross margin fall from 41% to 37% in August?" On a pasted P&L, the assistant splits the change into the revenue and cost lines on the report, then offers likely causes. Connected to the ledger, it opens cost of goods sold, finds $18,400 of freight for July shipments posted in August, and shows you the entries.

"Is anything unusual in this month's expenses?" On a pasted P&L, it compares against whichever months you pasted. Connected, it compares against the full history and flags a vendor paid twice in the same week.

"What does cash look like if we hire two people in November?" On a pasted P&L, it needs you to paste payroll, receivables and the bills due. Connected, it reads the payroll history, the receivables aging and the open bills itself.

The pasted answers are still useful. The connected ones are the ones you can act on without doing the digging yourself. The margin skill and the hiring cash skill are written for a connected assistant.

Claude or ChatGPT?

For finance work inside a company, two plain things decide it: which assistant your firm has approved for company or client data, and what you can connect it to. Both take MCP connectors, and QuickBooks is in both. Use the one you already pay for, and put the effort into the connection.

The review habit

1. Ask what it looked at. Before you use an answer, ask which data it used. An answer from an export is as of the export date.
2. Tie every number back. A figure headed for a board deck or a client gets traced to the report or the entry it came from.
3. Read before write. Some connectors can create invoices or entries. Start with read access, and allow writes only once you've watched how it reads your books.
4. A named person approves. The assistant drafts; the person who signs the numbers checks them.

Going further

The skills library has chat skills for any assistant and connected ones for when it can read the books. We connect our own books through RoboLedger, and Claude in Excel vs. Claude in the ledger makes the longer case for reading the source.