Give a frontier model a clean set of filings and ask it to build a three-statement model, and it will get roughly 89% of the answers right. Move that same task onto a real desk, with messy inputs and a retrieval step in front of it, and accuracy collapses: enterprise systems tested against the FinanceBench dataset fail more than 80% of the time. That 70-point gap between the demo and the day job is the whole story of AI in spreadsheets right now, and almost nobody selling you a tool will say it out loud.
I have spent 20 years in IT operations and the last few doing fractional COO work, which means I have built and, more often, torn apart other people’s financial models and operating dashboards. The pattern with AI-assisted spreadsheets is not that the tools are useless. It is that they are useful in exactly the way that gets you in trouble: fast, confident, and wrong in places you were not trained to look. This is a guide to using them anyway, without signing your name under a number the model invented.
What the tools genuinely earn their keep on
Start with the good news, because it is real and it is where almost all of the measured return on investment actually lands.
AI is strong at structured extraction and at mechanical, checkable transformations. Pulling a specific line item out of a 10-K, rolling a prior-period schedule forward into the current period, reconciling two versions of the same tab, converting a wall of pasted text into a clean table: these tasks have what one 2026 review of the evidence called “clear mechanical correctness criteria.” Either the filing contains the number or it does not. Retrieval-augmented systems beat general-purpose chat on this kind of work by a wide margin, and it shows up in adoption data. A KPMG survey of roughly 2,900 executives found 52% already using AI somewhere in financial reporting, with 92% saying the return met expectations, but the return concentrated in document processing, not in analytical generation. The machine is a very good clerk. It is a dangerous analyst.
The second thing it does well is get you off a blank sheet. In a 2026 WallStreetPrep evaluation that pitted the Excel add-in Shortcut, Claude, Microsoft Copilot, and ChatGPT against a real modeling task, the reviewers were blunt about where the value sits: these tools are good for “kickstarting models from scratch,” roughly the zero-to-60% stretch. Claude and Shortcut both scored a perfect 10 on understanding intent, because they asked thoughtful clarifying questions before touching a cell, about forecast assumptions, revenue segmentation, share repurchases, and how the schedules should be laid out. ChatGPT, by contrast, asked nothing and scored a zero on that dimension. If you have ever handed a model to a junior analyst who started typing before they understood the assignment, you know which behavior you want.
The errors that do not look like errors
Here is the part the marketing skips. When these tools fail, they rarely fail loudly. They fail in a way engineered, almost, to pass your review.
The failure mode has a name now. Researchers call them “ecological errors”: precise mechanical mistakes such as selecting an adjacent temporal column in a financial table, or applying perfectly correct logic to a variable that was extracted incorrectly. The reasoning looks sound. The formula looks sound. The source number is off by one column, so the answer is confidently, quietly wrong. In the WallStreetPrep test, both Shortcut and Claude “hallucinated significant portions of historical data” on their first pass, and the reviewers flagged the real danger directly: the errors were subtle enough that the wrong line items still rolled up into correct-looking subtotals. Your eye checks the total, the total ties, and you move on. That is how a bad number ships.
This is worse in finance than in most domains because the tell that usually saves you, obvious nonsense, is absent. General summarization hallucination rates are low for the best models: on Vectara’s short-document leaderboard, GPT-4o sits around 1.5% and Claude 3.5 Sonnet around 4.6%. But those numbers describe “did it make something up,” not “did it grab the right cell.” Point the same models at genuinely hard, multi-step reasoning and the range blows out. Stanford researchers documented hallucination rates of 58% to 88% on legal questions, and financial modeling carries the same structural risk: many steps, each one a chance to shift a period, misread a sign, or reconcile the wrong two documents.
The WallStreetPrep reviewers found every tool they tested failed at circularity, the interest-income-and-expense loop that runs off cash balances, and not one of them modeled shares outstanding correctly. Those are not exotic edge cases. They are Tuesday. Their summary is the line I would attach to every AI spreadsheet pilot: these tools “are very good at tempting you to give them more responsibility than they can handle,” and analysts who over-trust them “quietly create problems that take longer to fix than it takes to do the job without them.”
Match the tool to the mistake it is least likely to make
There is no single best AI for spreadsheet work. There is a best tool for a specific failure you are trying to avoid, and the practical differences between the platforms are mostly about architecture, not intelligence.
ChatGPT runs a real Python sandbox with pandas and matplotlib. When the task is data manipulation, cleaning a messy export, deduping, pivoting a hundred thousand rows, building a chart, this is the one that actually executes code against your file instead of reasoning about it in its head. It takes CSV and XLSX up to 512 MB and 20 files per chat. The catch, per the modeling test, is that its output was “easiest to audit” partly because it did the least thinking about your intent up front.
Claude is LLM-only in its chat interface, no code execution, but it is the strongest at multi-tab reasoning and at explaining where a number came from, which is exactly what you need when you are auditing rather than generating. Its chat upload cap is smaller (around 30 MB, or 500 MB through the API), and it reaches into Google Sheets through a =CLAUDE() add-on function. If the hard part is business logic and long, messy context, this is the tool. Anthropic has also pushed Claude directly into Office; I wrote about what Claude in Chrome, Excel, and PowerPoint actually changes when that shipped.
Gemini is the native choice if your work lives in Google Sheets. It sits in a sidebar and exposes an =AI() cell function, with a million-token context window on the paid tiers, though it does not execute code the way ChatGPT does. If you want the fuller picture of Google’s stack, the Gemini platform guide covers it.
Microsoft Copilot, embedded in Excel, scored highest on raw data accuracy in the WallStreetPrep test (8 out of 10) precisely because it is closest to the grid, and Shortcut, a purpose-built Excel add-in, won overall by treating modeling as its only job. The takeaway is not “buy the winner.” It is that a general chatbot, a code interpreter, and a native add-in fail differently, so pick the one whose failure mode you can most easily catch.
One rule cuts across all of them, and it is the single most important operational habit: do not let the agent go find the data. The evidence is consistent that retrieval is where these systems break. Upload the PDF, upload the spreadsheet, hand it the source, and make it work from what you gave it. The moment it is browsing for figures on its own, you have reintroduced the 80% failure rate.
The workflow that survives an audit
None of this is an argument against using AI in spreadsheets. It is an argument for a workflow that assumes the tool is a fast, overconfident clerk and builds the guardrails around that assumption. Here is the one I use and recommend.
Feed it the source, never the search. Attach the actual filing or export. If you find yourself pasting a prompt like “look up Acme’s 2025 revenue,” stop; that is the exact request that gets a plausible, wrong number back.
Make it show its extraction before its analysis. Ask for the raw pulled figures with cell or page references first, in a separate step, and check those against the source. Extraction is the cheap thing to verify and the expensive thing to get wrong. This is the same discipline as defining the outcome and the constraints instead of the steps: you are specifying what “correct” looks like so the model has a target you can score.
Audit the line items, not the totals. This is the one that matters most. Because subtle errors roll up into clean subtotals, checking the bottom line proves nothing. Spot-check individual cells, especially anything involving a time period, a sign convention, or a variable pulled from a table. Temporal shifts and extraction errors are where the money leaks.
Keep the model on the zero-to-60% work and do the last 40% yourself. Let it scaffold, structure, and draft. Do the forecasting logic, the circular references, and the judgment calls by hand, or at least rebuild them until you understand every dependency. The reviewers were right that finishing is where trust becomes a liability.
Measure it on your own data, not a leaderboard. A model’s benchmark score tells you almost nothing about how it handles your chart of accounts and your naming conventions. Build a small set of tasks you already know the right answer to, run each candidate tool against them, and grade the results. That is the same logic I laid out for building an eval that predicts what actually ships, applied to spreadsheets: the eval is the moat, not the model. And once you know which tool to trust for which job, route deliberately by cost and capability instead of defaulting to the most expensive option for everything.
The uncomfortable truth is that the time AI saves you on the front half of a model can be entirely eaten by the audit on the back half if you skip the guardrails. Used as a clerk with a supervisor, it is a genuine speed-up. Used as an analyst you trust, it is a misstated forecast waiting for your name on it. Treat every number it hands you as a claim to be checked, not a result to be filed, and you get most of the upside with almost none of the risk. If you want the broader mental model for what these systems can and cannot be trusted to do on their own, start with what an AI agent actually is and work back to your spreadsheet from there.