Can ChatGPT do Excel models?

The Excel Modeling Bottleneck: Why Teams Are Asking This Question

Every finance team knows the pain: a three-statement model that eats a week, a forecasting workbook that breaks the moment someone inserts a row, a DCF that gets rebuilt from scratch every quarter. Building financial, statistical, and forecasting models in Excel is slow, error-prone, and expensive — and it's exactly why teams are now asking whether ChatGPT can finally take over the heavy lifting.

The stakes are real. Practitioners report financial analysis that once took 3–4 days of manual work being completed in roughly 5 minutes with AI assistance, and a revenue analysis that would normally consume 6 hours finished in minutes after just three prompt iterations (one practitioner newsletter documents these cases firsthand). A finance AI workflow report found a technical accounting memo that took 4 hours was cut to about 30 minutes including human review. Those numbers explain the urgency behind the question.

Early AI workarounds only partially helped. Teams would:

• Copy-paste formulas ChatGPT generated into their own spreadsheets — workable, but ChatGPT couldn't produce a live, downloadable Excel file (as finance trainers noted)
• Export CSVs through Advanced Data Analysis, uploading files so ChatGPT could run Python and hand back downloadable results (per MIT Sloan EdTech's guidance)
• Wire ChatGPT into Excel via Office Scripts and an API key — a DIY approach that, as one CPA wrote, still required checking every response because "sometimes I received an inaccurate response" (Journal of Accountancy)

These approaches made ChatGPT a formula helper, not a model builder. The bottleneck remained: a human still had to assemble the structure, wire the sheets together, and audit every assumption. Worse, AI could hallucinate plausible-but-flawed formulas that looked correct but were mathematically wrong, so the "time saved" often came back as review time.

That's the gap this question is really about. Teams don't need help typing a SUMIFS faster — they need a tool that builds the whole model: revenue build, scenario toggles, sensitivity tables, all traceable and auditable. OpenAI's ChatGPT for Excel add-in now claims to do exactly that, creating live models directly inside workbooks from plain-language descriptions. Whether it actually delivers — and what still requires human judgment — is what the rest of this article tests. If you extend these workflows with AI agents, SkillzBench's free directory lets you browse and test security-scanned, benchmarked agent skills before trusting them with your spreadsheets.

What ChatGPT Can Now Build: From Formula Helper to Live Model Builder

For years, ChatGPT could only hand you a formula through the window while you did the driving. That changed when OpenAI shipped an add-in that builds the whole car inside your garage.

The ChatGPT for Excel add-in (beta March 5, 2026, generally available May 5, 2026) creates and updates live Excel models directly in the workbook from plain-language descriptions, keeping structure, formulas, and assumptions Excel-native so calculations run in Excel and you can trace and audit every input. It reasons across sheets, explains why outputs changed, and asks permission before touching your file.

Beyond the add-in, three tools round out the landscape:

• Advanced Data Analysis — uploads datasets, runs regressions in Python, and exports downloadable files; MIT Sloan notes tasks that "formerly took several hours can be done in minutes."
• ChatGPT for Financial Services (launched September 10, 2026) — pairs GPT-6 Astra reasoning with premium data from Daloopa, PitchBook, LSEG, and Crunchbase, and lets admins publish firm-formatted Excel templates for valuation models and pitchbooks.
• Financial data integrations — FactSet, S&P Global, Moody's, MSCI, and others feed the add-in, with proprietary data connectable via MCP.

Model coverage is genuinely broad: DCF components, three-statement models, scenario toggles, sensitivity tables, and WACC formulas. OpenAI's internal investment-banking benchmark for three-statement models jumped from 43.7% (GPT-5) to 87.3% (GPT-5.4 Thinking) — a vendor-internal figure, but consistent with practitioner reports of coherent first-draft DCFs.

The honest limits matter just as much. OpenAI itself flags beta constraints: slow responses, outputs needing formatting cleanup, and complex formulas or edge cases requiring manual refinement. The Wall Street School warns of the "black box" problem — AI can hallucinate plausible-but-flawed formulas — while Coursiv's verdict is blunt: ChatGPT is "accurate enough to accelerate finance work, but not accurate enough to run finance work unsupervised."

The consensus across every source is co-pilot, not captain. Excel still does the math; ChatGPT writes the formula you would have typed. The judgment layer — which assumptions hold, where the model breaks — stays with you. If you're extending this workflow with AI agents, SkillzBench is building a free directory of security-scanned, benchmarked agent skills to help you pick ones that actually deliver lift.

The Verification Gap: Why 'Accurate Enough to Accelerate' Isn't 'Accurate Enough to Run'

Every source on this topic lands on the same uncomfortable truth: ChatGPT is accurate enough to accelerate finance work, but not accurate enough to run it unsupervised. That distinction is the whole ballgame when you're building models that inform real decisions.

The core problem is what practitioners call the black-box issue. ChatGPT can hallucinate numbers or formulas that look correct but are mathematically flawed — plausible syntax, confident delivery, wrong logic. Even OpenAI's own guidance on the Excel add-in acknowledges that outputs may need cleanup and that complex formulas or edge cases may require manual refinement.

The vendor's internal benchmark tells a revealing story: three-statement model accuracy improved from 43.7% with GPT-5 to 87.3% with GPT-5.4 Thinking — impressive progress, but still a vendor-internal number with no independent verification, and it still leaves roughly one in eight attempts falling short. As one CPA put it after testing ChatGPT in Excel: "Sometimes I received an inaccurate response... Be sure you check the accuracy."

Practical self-checking tactics close some of that gap:

• Ask ChatGPT to explain its work step by step — MIT Sloan EdTech recommends this to force the model to check its own logic.
• Inspect the underlying Python code via the "view analysis" icon when using Advanced Data Analysis, rather than trusting the output at face value.
• Build models in pieces (revenue build, then depreciation, then working capital) so errors surface early instead of compounding.
• Test every generated formula directly in Excel before wiring it into your model.

The verification burden extends beyond accuracy to confidentiality. On Free, Plus, and Pro plans, OpenAI trains on user content by default, with opt-out available — MIT's own policy prohibits sharing non-public data with public AI tools. Business and Enterprise workspaces don't train on inputs, but a financial model full of client figures pasted into a personal ChatGPT account is a genuine compliance risk.

That verification mindset applies to the AI tooling ecosystem too. SkillzBench is being built on exactly this principle — that agent skills should be security-scanned, version-pinned, and benchmarked for actual lift before you trust them in a workflow, because published research found roughly 26% of skills carry at least one vulnerability. Treat ChatGPT's Excel output the same way: useful draft, unverified until you've checked the math yourself.

Your Implementation Playbook: Prompts, Workflow, and Going Further with Agent Skills

Knowing ChatGPT can build Excel models is one thing; getting reliable output week after week is another. The difference, according to practitioners, isn't model capability — it's workflow.

Finance trainers recommend a simple three-step build workflow: define your assumptions up front, request the model logic in pieces (revenue build first, then depreciation, then working capital), and audit every formula at the end (The Wall Street School — see can ChatGPT do financial modeling). Breaking the build into chunks keeps each output verifiable, and it forces you to catch the plausible-but-flawed formulas AI can hallucinate.

Prompting matters just as much. One practitioner found that generic asks like "generate a SUMIFS formula" produced nothing he could use — what actually worked was deliverable-focused prompting: "a board-ready revenue model with variance analysis, currency conversions, and scenario toggles." His 18 production-ready prompts are organized around deliverables, not AI features, and he reports building in roughly 5 minutes analysis that previously took 3–4 days (Nate's Newsletter). Those speed claims are anecdotal, but the pattern — iterate in smaller steps, verify every number — recurs across sources (Coursiv's 2026 finance review).

Privacy is the other non-negotiable. Free, Plus, and Pro plans train on user content by default; Business and Enterprise plans do not (Coursiv). For confidential financials, use an enterprise plan or anonymize data first.

Your working checklist:

• Define assumptions before prompting — output accuracy depends entirely on the quality of assumptions you supply.
• Request logic in pieces and audit each formula before moving on.
• Ask ChatGPT to explain its work to force self-checking (MIT Sloan EdTech).
• Keep non-public data out of plans that train on user content.

For team adoption, dedicated calendar time is the real unlock — junior staff adopt faster than seniors, and protected time is what embeds the workflow (Coursiv). One accounting team cut a 4-hour technical memo to about 30 minutes including review, after roughly 60 hours of setup.

Going further, agent skills can automate the surrounding workflow — data prep, formatting, validation — beyond what a single chat session covers. SkillzBench is building a free directory of security-scanned, version-pinned, benchmarked skills for exactly this; since published research found roughly 26% of skills carry at least one vulnerability, verification before adoption is worth the few extra minutes.