← All posts

AI in Excel in 2026: Copilot, ChatGPT, Claude and agents, and what actually works.

THE GRIDPATH TEAM · OCTOBER 11, 2026

Every spreadsheet now ships with an AI button, and every chatbot now accepts an .xlsx. Most guides to this read like press releases. We build a spreadsheet agent for a living, so this is the version we'd want to read ourselves — what each approach actually does, what it quietly can't, and how to pick the right one for the job in front of you. It's cross-checked against the practitioner record — the long threads where accountants and CFOs compare notes after the demo glow wears off. Where their experience contradicts the marketing, we sided with them.

The four ways to put AI in a spreadsheet

"AI in Excel" is really four different product categories that happen to share a marketing phrase. They're good at different things and — more importantly — they fail in different places:

• Built-in copilots — Copilot in Excel, Gemini in Google Sheets. A sidebar inside the app you already use.
• Upload-and-chat — ChatGPT, Claude. You hand a chatbot the file and talk about it.
• AI as a formula — =COPILOT() in Excel, =AI() in Sheets, and a pile of add-ins. The model becomes a function in the grid.
• Agents on the file — tools that open your workbook and edit it through tool calls, multi-turn, with diffs. This is the category GridPath is in, so read that section knowing who's talking.

The short version, before the detail:

Built-in copilots: convenient, shallow

Copilot in Excel is genuinely useful for a specific shape of work: your data is a clean table, and you want one thing done to it. Suggest a formula and explain it, throw together a pivot or a chart, apply conditional formatting you describe in words, summarize the trend. For that, it's right there in the ribbon, and it's fine.

The limits show up fast. The good parts want your file in OneDrive or SharePoint with AutoSave on — awkward if your models live on disk. Licensing is its own maze — metered credits on consumer plans, a separate per-seat license (~$30/user/month) for the business version. And it's fundamentally single-shot: one instruction, one result. It doesn't carry a plan across steps, doesn't notice what step three broke, doesn't loop until done. Point it at a real workbook — merged headers, cross-sheet links, fifty tabs of inherited structure — and it mostly retreats to summarizing.

The practitioner record is consistent and rough: Copilot is better at troubleshooting formulas than generating them. Veterans report learning genuinely new functions from it, and that building from scratch falls apart as complexity rises — until fixing its hallucinated suggestions takes longer than doing the job yourself. Microsoft's own launch guidance said not to use it for tasks requiring accuracy or reproducibility, which an entire subreddit of accountants found extremely funny. One nuance the pile-on misses: the dedicated Excel agent under the Tools button is meaningfully better than the chat sidebar most people judge Copilot by.

Gemini in Google Sheets is the same shape: good at organizing data, generating tables, quick charts; same single-step ceiling, plus everything Sheets itself inherits (the 10-million-cell limit, pivots that crawl at scale).

Upload-and-chat: great analysis, dead output

For understanding a file, the chatbots are hard to beat. Upload a workbook to ChatGPT or Claude and it will run real analysis over it — group, filter, regress, chart, describe — and, crucially, it's conversational: you can interrogate the result, push back, ask the follow-up. For "someone sent me this file, what does it say?" this is the fastest tool that exists.

Power users have converged on a telling workflow here: paste the data as a markdown table and ask how to solve the problem — many report better results from that than from anything running inside the spreadsheet. Typing the prompt forces you to state the problem precisely, which was half the work all along.

The catch is what comes back. The analysis runs as Python over a copy of your data, so the output is a new file — and the round trip is lossy. Formatting, charts, and formulas frequently don't survive; what you download is computed values pasted where your live model used to be. (Why that matters more than it sounds is its own post.) And on big workbooks, context limits mean the model may be reasoning over a truncated slice of your file without telling you. The rule of thumb: upload-and-chat produces reports, not workbooks.

AI as a formula: clever, and a little cursed

=COPILOT("categorize this expense", A2) in Excel, =AI() in Sheets: the model becomes a function that recalculates like any other. For text work over rows — classify, extract, summarize, translate — this is genuinely great, and it covers exactly the ground regex-and-MID() formulas were always terrible at.

The cursed part: a calculation engine is built on the assumption that the same inputs produce the same outputs, and a language model doesn't promise that. A recalc can re-answer differently than it answered last week. Every recalc burns metered credits and adds latency. And there's no audit trail for what the model saw when a cell last calculated. Classification is also where the funniest failures live — one accounting team watched an AI file Uber Eats receipts under "Travel." Use these functions for text transforms you'll paste-freeze once you're happy — not for anything that feeds a number someone will rely on.

Agents on the file: the newest category

The fourth category is what happens when you take the loop coding agents proved out — read, plan, edit, check, repeat — and point it at a workbook. The defining features: it's multi-turn, it works through tools that read and write actual ranges, and it verifies its own output. The agent builds the block, reads it back, sees the #REF! it just created, and fixes it — without you re-prompting.

Full disclosure: GridPath is one of these — a desktop agent working directly on your local .xlsx, running on your own Claude or ChatGPT subscription. Others are approaching the same idea from different directions: task-pane add-ins that put an agent inside Excel itself (Claude for Excel is the one finance forums are excited about — users report it tracing dependency chains through multi-sheet models and collapsing week-long builds into hours), and AI-native spreadsheets that skip Excel entirely (Quadratic, with Python and SQL living in the grid). The category is young and moving fast.

Its honest costs: rendering fidelity is hard outside real Excel (our own in-app grid doesn't paint charts or pivots live while you edit — we've written about why); agents are confident, so you still review the diffs; and for a one-cell task, a full agent loop is overkill — the ribbon button is faster. Users of the add-in agents also report the current ceilings: context runs out once work spans multiple files, and nothing persists between sessions — people end up maintaining a dedicated "instructions sheet" just to re-teach the agent every morning.

The blocker the demos skip: your IT department

Read any long thread of professionals discussing these tools and one theme dominates, and it isn't capability — it's whether the data is allowed to leave. Workbooks are where the sensitive stuff lives: payroll, pricing, PII, the acquisition model. Companies block cloud AI tools outright; some have gone as far as moving to Office LTSC — the edition without the AI — to keep the buttons away from their data. Meanwhile Copilot's genuine advantage is the inverse of its reviews: the weakest tool in the lineup has the strongest compliance story, because everything stays inside the Microsoft tenancy your company already trusts.

So evaluate on two axes: what the tool can do, and where your cells go while it does it. Defensible answers: local-first tools (GridPath keeps files on disk and sends only structure plus the cells the agent reads or writes), enterprise plans with no-training guarantees, or staying inside your Microsoft or Google tenancy. "We pasted it into a chatbot" is not one, and your IT department knows it.

AI amplifies Excel skill — it doesn't substitute for it

The single most-repeated insight across every practitioner thread, nearly verbatim each time: if you're good at Excel, AI speeds you up; if you're not, it trips you up. These tools produce answers at a quality bar of roughly "plausible" — one FP&A shop described the output as five out of ten answers making sense, with the real work being the filtering. An expert recognizes the five instantly and banks the time saved. A novice can't tell which five, and now has confident wrong numbers with professional formatting. Whoever you are, the review pass isn't a chore on top of these tools — it is the job.

And the edge compounds: if AI writes all your formulas, you stop practicing the exact skill you need to check AI's formulas. Use these tools to go faster at work you understand — not to skip the understanding.

What none of them solve

Two problems survive every category, and any guide that skips them is selling something:

• The monster workbook. Fifty sheets and tens of thousands of formulas strain every approach. Agents that read on demand degrade the most gracefully, but nothing ingests a monster whole — anyone claiming otherwise is truncating quietly.
• Values where formulas should be. The industry default is still to paste computed values into your model. If the file has to recalculate next quarter, that's disqualifying — here's the full argument.

So which should you use?

• Quick formula help, a chart from a clean table — Copilot in Excel or Gemini in Sheets, if you're already paying for them. Right there, good enough.
• Understanding a file someone sent you — upload it to ChatGPT or Claude. Treat the output as a report, not a workbook.
• Bulk text work over rows — =AI() / =COPILOT(), then paste-freeze the results when you're done.
• Building or maintaining a real model — an agent on the file, with diffs you review. This is the only category doing multi-step work with a check pass.

One piece of borrowed wisdom from a CFO forum matches everything we've seen: ninety days with one tool beats ninety days dabbling across three. Pick one; go deep.

The honest summary: copilots are convenient, chatbots are insightful, formula-AI is handy, agents are powerful — and every one of them will cheerfully be wrong at scale if nobody looks. The skill in 2026 isn't picking the one true tool. It's knowing which job you're holding.