Automation

ChatGPT for FP&A: What It Can and Can't Do

Marc SeanJuly 8, 20266 min read

What ChatGPT Gets Right for Financial Work

Formula generation is the real win. Describe what you need in plain English and ChatGPT will produce a working formula faster than you can write it from scratch. "Give me a SUMIFS that pulls revenue from column C on the P&L tab where the date in column B is on or after the date in Assumptions B3" yields something like:

=SUMIFS('P&L'!C:C,'P&L'!B:B,">="&Assumptions!$B$3)

That's legitimately fast. For analysts who live in structured references and occasionally need to construct a gnarly nested IFERROR or an XIRR across a non-contiguous range, the time savings are real. OpenAI's GPT-4o handles this kind of formula construction with high accuracy - testing in mid-2026 puts simple-to-intermediate formula generation at roughly 85-90% accuracy on first attempt, with the failure rate climbing for complex array formulas or LAMBDA chains.

Debugging error messages is the other area where ChatGPT earns its keep. Paste a #VALUE! and the formula producing it and you'll get a clear diagnosis in seconds. Same for explaining why your IRR function is returning #NUM! when your cash flows look fine (usually a sign-convention problem, and ChatGPT will tell you that immediately).

Narrative drafting is underrated. If you need to turn a EBITDA bridge into 3 sentences for a board slide, or draft the assumptions section of a memo, ChatGPT does it well. That's not a spreadsheet task anyway.

The Copy-Paste Tax

Here's the problem: ChatGPT is blind to your model.

Every time you need it to do something context-specific, you're manually copying your column headers, pasting sample data, and describing your tab structure in prose. For a one-off formula question, that's 60 seconds. For an iterative modeling session - "now adjust that to exclude intercompany eliminations, and pull through to the FCFF tab" - you're doing that dance 8 times.

GPT-4o's context window is 128,000 tokens, which sounds enormous until you try to paste a fully-built 8-tab model with 3 years of monthly actuals. A single tab with 500 rows and 15 columns of financial data runs 20,000-40,000 tokens. You can't fit the whole model, so you're constantly deciding what context to include and what to leave out. That decision is cognitive overhead, and it's on you, not the AI.

There's also no write-back. ChatGPT produces text. You copy that formula, switch windows, find the right cell, paste it. Multiply that by every formula in a sensitivity table and the workflow starts to feel like using a calculator with a broken display - technically functional, deeply annoying.

Where It Falls Apart

No memory between sessions. ChatGPT doesn't know that your discount rate lives in Assumptions!$C$12 or that your fiscal year starts in April. You re-explain this every session. For analysts who return to the same model weekly, that's a real friction cost.

Hallucination on edge-case functions. ChatGPT occasionally writes formulas that look syntactically correct but produce wrong results. It has confidently produced XNPV formulas with the dates and values arguments reversed - the formula runs, returns a number, and the error only surfaces when you sanity-check the output against your DCF. According to Google's official Sheets function documentation, XNPV(discount, cashflows, dates) - but ChatGPT periodically swaps cashflows and dates. Trust but verify.

No awareness of your data types. If your revenue column has 3 cells formatted as text because someone pasted from a PDF, ChatGPT can't see it. It writes a formula that works in theory, fails in practice, and you spend 20 minutes diagnosing a problem it created by not knowing your actual data state.

Cross-tab formula writing requires manual scaffolding. To get a formula that correctly references 'Cash Flow'!D:D filtered by a date range on 'Assumptions'!$B$3, you have to manually describe the tab names, column letters, and structure. ChatGPT doesn't know any of it. In a real model - an LBO with a Sources & Uses tab, Debt Schedule, Returns Analysis, and three operating scenarios - you're managing that mental overhead yourself.

How ChatGPT Compares to Dedicated Spreadsheet AI

CapabilityChatGPTSpreadsheet-native AI (e.g. ModelMonkey)
Formula generationStrongStrong
Sees your actual dataNoYes
Writes to cells directlyNoYes
Cross-tab awarenessManualAutomatic
Memory of model structureNoneSession-level
Write-back without copy-pasteNoYes
Works inside Google SheetsNo (separate tab)Yes (sidebar)

The gap isn't formula quality. It's context. ChatGPT can write the right formula for an abstract problem. A spreadsheet-native AI can write the right formula for your model, with your actual column headers and your actual tab names, and drop it in the right cell.

ModelMonkey sits inside Google Sheets as a sidebar. You type "what's my trailing 6-month MRR trend?" and it reads your actual ranges, runs the calculation, and returns a table directly in your sheet - not in a separate browser tab you then copy from. For a quarterly board pack where you're iterating on 15 different number pulls, that workflow difference compounds fast. Try ModelMonkey free for 14 days - it works in both Google Sheets and Excel.

Use ChatGPT when you have a self-contained question: a formula to write, an error to debug, a paragraph to draft. Don't expect it to be a modeling partner for anything that requires seeing your actual sheet. The copy-paste loop is a real cost, and it scales badly with model complexity.

For FP&A work where the whole point is that numbers tie out across tabs, you need something that can actually see those tabs.


Frequently Asked Questions