OpenAI puts ChatGPT inside the Excel workbook, not next to it

The add-in lives in the grid, not a chat sidebar

The framing that matters here is where the interface sits. OpenAI describes ChatGPT for Excel as "a version of ChatGPT embedded directly in spreadsheets" that builds, analyzes, and updates models "using the same formulas and structures teams already rely on."

That is a deliberate frontend decision. Instead of asking an analyst to copy cells into a chat window and paste results back, the model operates on the live workbook. The announcement is explicit that calculations "run directly in Excel," not in a separate reasoning environment that returns a rendered answer.

The practical consequence: outputs stay Excel-native. OpenAI says teams can run analysis, reporting, budgeting, and scenario work "while preserving structure, formulas, and assumptions in a formatted, Excel-native workbook." The deliverable is a spreadsheet, not a transcript describing one.

Permission gates and cell-level citations are the trust surface

Editing someone's financial model is high-stakes, and the interface reflects that. ChatGPT "asks for permission" before making changes, so users can "review each step and undo edits if needed." This is a consent gate built into the frontend, not an afterthought.

It also grounds its claims. OpenAI says ChatGPT "links answers to the exact cells it references and updates," so analysts can "trace assumptions, audit formulas, and verify how results were produced." Citations point at cells rather than sources—a form of provenance suited to the medium.

The company is candid that this is beta. Responses "may take longer," generated output "may occasionally require cleanup," and "complex formulas or edge cases may still require manual refinement." The permission-and-undo pattern exists precisely because the model will sometimes get formulas wrong.

The benchmark that anchors the finance pitch

OpenAI ties the launch to GPT-5.4 (as GPT-5.4 Thinking) and a single hard number. On its internal investment-banking benchmark—which evaluates tasks like "building a three-statement model with proper formatting and citations"—performance rose from 43.7% with GPT-5 to 87.3% with GPT-5.4 Thinking.

Two caveats are worth stating plainly. The benchmark is OpenAI's own and internal, so it is not independently verifiable from the announcement. And it measures real workflow output—formatting and citations included—rather than an abstract reasoning score, which makes the near-doubling meaningful but also self-defined.

The customer quote from Hg's head of AI, Amr Ellabban, reinforces the workflow framing rather than the number:

ChatGPT has materially accelerated our research and due diligence workflows—from financial analysis and market research to legal review and writing internal memos—while improving consistency across teams.Montana Labs

Data integrations and MCP move the workflow upstream

Alongside the Excel add-in, OpenAI added financial data integrations inside ChatGPT—Moody's, Dow Jones Factiva, MSCI, Third Bridge, and MT Newswires at launch, with FactSet "coming soon." The stated goal is to bring "market, company, and internal data into a single workflow."

Firms can also wire in proprietary data by building their own apps via the Model Context Protocol. That extends the reach beyond vendor feeds into internal sources, and the research features can export cited outputs to PDF or Microsoft Word.

Access is gated for organizations. In Enterprise, Edu, and Teacher workspaces the feature is "off by default," and admins enable it per user through custom roles and group permissions—consistent with the RBAC, SSO, and data-residency controls OpenAI lists.

What building inside the cell means for how finance teams adopt AI

The specific implication of this launch is that OpenAI is meeting analysts inside the tool they already fight with, rather than asking them to migrate to a new one. The model edits the same cells, obeys the same formulas, and returns the same file type.

That lowers the switching cost dramatically, but it also raises the bar for correctness. When the output is a live three-statement model an analyst will present, a wrong formula is not a bad paragraph—it is a wrong number in a deliverable. The permission gates, undo, and cell-level citations are the mechanisms OpenAI is betting can carry that weight.

The general-availability note—GPT-5.5 powering both Excel and Google Sheets across all plans as of May—signals OpenAI moved quickly from beta. For teams evaluating it, the honest test is not the 87.3% benchmark but whether the audit trail holds up when a model breaks in production.