If you're picking a spreadsheet AI for your firm in June 2026, the math is simpler than the marketing makes it sound.
Both Excel Copilot and Sheets Gemini can build a P&L dashboard from a plain-English prompt. Both run formulas, build charts, write summaries, flag variance. The feature gap that existed a year ago is almost closed.
So why does the choice still matter? Because spreadsheet AI is not really about the AI. It's about which ecosystem your firm already lives in. Switching, retraining, and re-integrating is where the actual money is.
What Excel Copilot does. Microsoft 365 Copilot in Excel handles plain-language analysis at a level that surprises long-time Excel users. Ask it: "Summarize the last three years of this client's revenue by category and flag any line where year-over-year variance exceeds 20%." It writes the formulas, runs the analysis, drops the summary into the sheet. As of early 2026, Microsoft added model choice inside Agent Mode — Anthropic or OpenAI per task.
The pricing is the gotcha. Copilot is an add-on, not a standalone. You must already be on Microsoft 365 Business Basic, Standard, Premium, E3, or E5. Promo rate is $18/user/month through June 30, 2026, then $21/user/month. For a 10-person firm already on Microsoft 365, that's the full cost. For a firm not on Microsoft 365, you're paying the base license too — closer to $33–$40 per user once you stack it.
What Sheets Gemini does. Google's recent updates turned Gemini in Sheets into a real end-to-end builder. Ask it to "build a P&L dashboard from my historic service incidents and rate cards" and Gemini will synthesize, propose a plan, and construct the formatted spreadsheet. Side-by-side editing — "add scorecards and bar charts above my sales data" — is native.
Gemini is included in Google Workspace Business and Enterprise plans at no add-on cost as of the 2025 packaging refresh. Business Standard is $14/user/month, Gemini included. Compare that to Microsoft's $21–$33 stack.
The deciding question. Where does your client data already live? If client deliverables are .xlsx and your tax software exports to Excel, Copilot wins by default — switching costs swamp the AI feature gap. If you're a Google-native CAS firm (rare in tax, more common in advisory and bookkeeping), Gemini wins on price alone.
The one place to actually compare: pull your three most complex client workbooks, run the same five questions through each tool for a week in a sandbox account. The AI that does the harder things on your real data wins. Not the AI that did better in the demo.
What neither does well yet: auditable trace of how a calculation was derived. Both rewrite formulas without leaving a clear log. For tax workpapers where a reviewer needs to see exactly what changed, both still require manual documentation.
Microsoft quietly added model choice in Agent Mode earlier this year — Anthropic or OpenAI per task. Practical use: Anthropic for narrative summaries, OpenAI for raw number work. The setting is per-workbook, so train staff on which to pick before turning it loose on client files.
The new optimization problem solver inside Gemini handles budget allocation, staff scheduling, and resource constraints — the kind of advisory work that used to take an FP&A consultant. Useful for CAS firms running budget conversations with clients. Nothing equivalent in Excel Copilot yet.
35% off Business Standard + Copilot and 25% off Business Premium + Copilot bundles end this month. If your firm has been on the fence, the 12 months of savings are worth signing for before July 1.
Select a folder of tax-related files in Drive and ask cross-document questions. For firms that store client paperwork in Drive, this is the closest thing to a real RAG system without buying a separate tool. Works on PDFs and Sheets together.
That's the fully-loaded monthly cost of Excel Copilot for a firm not already on Microsoft 365, versus Google Workspace Business Standard with Gemini included. A $228 per-user-per-year gap. Across a 15-person firm: $3,420 a year. The AI feature gap is not $3,420 a year of better.
What it is: An AI add-on to Microsoft 365 that runs inside Excel (and the other Office apps) using a mix of Anthropic and OpenAI models. Plain-language prompts return formulas, analysis, charts, and written summaries directly in the sheet.
What it does well: Variance analysis, trend summaries, client-specific narrative ("write a 4-sentence note on this client's revenue trajectory"), conditional formatting at speed. The Agent Mode model picker is a quiet power-user feature.
What it doesn't do well: Audit trails. Source attribution on formulas it rewrites. Cross-workbook reasoning. Anything that requires the AI to know what your firm's prior workpapers looked like — it doesn't.
Pricing: $18/user/month through June 30, 2026, then $21/user/month — add-on only, requires a qualifying Microsoft 365 plan ($6–$22/user). Bundle discounts expire June 30.
This week: a Claude Project that catches the variances Excel doesn't. I built this for a CAS client's month-end close, where the controller was spending 90 minutes a month chasing line items and writing the executive summary by hand.
Step 1: Create the Project. In Claude Team, new Project. Custom instructions: "You're reviewing a monthly P&L for a [industry] business. Compare actual to budget and prior month. Flag any GL line with variance over 10% in dollars or 15% in percent. For each flag, write a one-sentence hypothesis based on the trend, not the number."
Step 2: Upload context. A redacted 6-month rolling P&L, the budget, and the chart of accounts mapping. Claude needs to know what "office supplies" really means at this company.
Step 3: The monthly prompt. Drop the current P&L in and ask: "Run the variance review. Then draft the three-paragraph executive summary using the firm template I uploaded."
What this does that Excel Copilot still can't: cross-month memory. Excel sees one workbook at a time. The Claude Project keeps the prior six months in scope, so variance commentary references trend, not just delta. The controller's 90 minutes is now 12.
What this doesn't replace: the actual review. The Project will sound confident about accruals you haven't booked yet. The closing call is still the controller's.
Open your firm's last three client workbooks. Write down the five most repetitive analysis tasks you did in each. Run those five through whatever spreadsheet AI you're already paying for (or trial either tool free for 30 days). If three or more come back clean on the first try, you have your answer. If two or fewer, the firm isn't ready and the problem isn't the AI — it's the data hygiene. Fix that first.
P.S. CAS firms: reply with the variance prompt that drives you nuts and I'll publish a fix in a future Workflow Lab.
