Finance teams spend a disproportionate share of their time rebuilding the same spreadsheet structures every month - monthly business reviews, budget variance reports, scenario models. OpenAI published a practical guide through its Academy platform showing how Codex, the company's AI coding agent that writes and runs code without constant user input, can handle that work from real financial data.
The guide covers five specific tasks: building monthly business reviews (MBRs), assembling reporting packs, constructing variance bridges (analyses that break down the gap between budgeted and actual numbers by line item), running model integrity checks, and generating planning scenarios from existing inputs. Crucially, it works from actual finance artifacts - spreadsheets and data exports - rather than clean demo datasets.
What Codex Actually Does Here
Codex operates as a coding agent, meaning it writes Python or SQL code, runs it in a sandboxed environment, and returns formatted results - without the user writing any code. For a variance bridge, that means feeding it your actuals and budget files and asking for a categorized breakdown of the differences. For an MBR, it means generating a formatted output from raw data instead of manually updating the same slide deck for the fourth month running.
The practical ceiling is data quality. If source files have inconsistent column headers or shifting naming conventions - both common in real finance environments - Codex will either ask for clarification or make wrong assumptions. Someone with domain knowledge still needs to review outputs before they go to a CFO.
This puts Codex in a different category from spreadsheet-native tools like Excel Copilot or Gemini in Google Sheets. Those work inside the file. Codex treats the spreadsheet as an input to a broader, multi-step workflow.
Who Benefits Most
The guide is most useful for small FP&A teams - one to three people covering planning and analysis for a mid-size business - where the same analyst builds the model, runs the report, and presents to leadership. There's no dedicated data engineering function to hand recurring work off to.
Larger finance organizations with established BI pipelines already automate most of what's covered here. For them, Codex is better suited to ad hoc analysis requests than replacing existing infrastructure.
OpenAI's Academy framing suggests a deliberate push to develop domain-specific onboarding content for Codex, extending its reach beyond the software developers who were its original core audience.