Best AI for Excel in 2026: Copilot vs Claude vs ChatGPT for Spreadsheet Work

Written by Mike Pearlstein, CISSP, CEO of Fusion Computing Limited. Helping Canadian businesses build and manage secure IT infrastructure since 2012 across Toronto, Hamilton, and Metro Vancouver.

Excel is where most Canadian SMB analysts, finance leads and operations teams spend half the workday. The question keeps showing up in our intake calls: which AI assistant actually helps with a real spreadsheet. The 2026 answer is different from the 2025 one. All three vendors now ship an Excel add-in that edits cells directly, so the deciding factor moved from reach to governance.

Why this comparison needs its own post

Our parent platform comparison answers which assistant gets the primary seat for the whole company. This piece answers a narrower question. When an analyst opens a workbook with 40,000 rows and three linked tabs, which AI saves the most time without sending that file somewhere the privacy officer has not approved? The answer still splits by job, and it now splits by audit trail too.

Building a SUMIFS across three sheets is a different problem from auditing a 200-row LAMBDA. Extracting a PivotTable summary from raw CRM data is different again. Each tool wins one of those jobs.

What is an Excel AI add-in, and what does each vendor actually ship?

According to Anthropic’s Claude for Excel documentation (2026), an Excel AI add-in is a task-pane app that reads the open workbook, answers with cell-level citations and writes changes back while preserving formula relationships. All three vendors now ship one. The differences sit in the trust boundary, the model bench and the operations each add-in is allowed to perform. Source: claude.com, 2026.

Through 2025 this comparison had an easy shape. Copilot lived in the ribbon and the other two lived in a browser tab. Three releases across ten weeks ended that.

• Copilot Agent Mode, April 22, 2026. Plans, executes and validates multi-step work in the grid rather than answering in chat.
• ChatGPT for Excel, May 5, 2026. Generally available on every ChatGPT plan, powered by GPT-5.5, with market-data connectors from Moody’s, FactSet and MSCI.
• Claude for Excel, May 7, 2026. Out of preview and generally available on every paid Claude plan, alongside Word and PowerPoint.

Snapshot: how the three tools approach Excel work

Six fields decide most analyst shortlists. Microsoft publishes Canadian dollar list prices directly, so Copilot figures below are Microsoft Canada list. OpenAI and Anthropic bill in US dollars, so those rows convert at the August 4, 2026 rate of roughly 1.41 CAD per USD. That single difference explains most of the confusion in vendor quotes.

The seat prices now land within about CA$1 of each other, so price no longer decides this. Copilot carries a base-licence requirement the other two do not: a qualifying Microsoft 365 plan, per Microsoft Learn (2026). Most Canadian SMBs already pay for that plan, which makes Copilot’s marginal cost the honest number to compare.

Curious which combination fits a 50-seat team? Book a consultation → and we will map the assistants against your actual workbook portfolio. The review is CISSP-led at a Microsoft Solutions Partner.

Book a Consultation

What Copilot in Excel is actually built for

Copilot is the only one of the three whose Excel work stays within the Microsoft 365 tenant. You highlight a range, prompt, and it operates on those cells. Per Microsoft Support (2026), Copilot in Excel generates and explains formulas, builds PivotTables, creates charts, highlights and filters data, and summarises numerical or text ranges under tenant permissions.

Agent Mode raised the ceiling on April 22, 2026. Rather than answering in a chat pane, it plans a multi-step job, writes the formulas, cleans the data and validates the result in place. A companion Plan mode shows the intended steps before anything changes, which matters when the workbook is a client deliverable.

The governance advantage is mechanical. Because Copilot reads tenant files directly, the workbook never leaves the Microsoft 365 boundary. If the tenant sits in Canada Central or Canada East, prompts and responses stay inside that geography under Microsoft 365 enterprise data protection. Purview sensitivity labels carry through every prompt and response.

Runtime note: Python in Excel runs on Microsoft Cloud, not on the local machine. External libraries such as pandas.read_csv are unsupported, so data must come from the worksheet or through Power Query. Copilot can now invoke Python itself for advanced analysis inside Edit with Copilot. Source: Microsoft Support, Get started with Python in Excel (2026).

“Across our 40 Canadian SMB client fleets running Copilot, finance and operations adopt fastest. They open a workbook three or four times a day. The ergonomic cost of switching to a browser was always the real barrier, and Agent Mode removed the last of it. Adoption now stalls on data hygiene instead, which is a fixable problem.”

Mike Pearlstein, CISSP, CEO, Fusion Computing.

Copilot is weakest at two things. Long Power Query M pipelines still sit outside the Copilot surface in the query editor. Dense reasoning across a whole workbook is also awkward, because the in-grid pane rewards short instructions rather than long analytical back-and-forth.

What ChatGPT does well for spreadsheets

ChatGPT now has two spreadsheet surfaces. The older one is Advanced Data Analysis, a sandboxed Python environment where you upload a .xlsx file and ask questions in natural language. It returns a cleaned workbook, a .csv, a chart image or a derived file. That path is still available on every paid tier and remains the fastest way to reshape a messy export.

The newer surface is the ChatGPT for Excel add-in. It entered beta on March 5, 2026 and reached general availability on May 5, 2026, powered by GPT-5.5. It builds and updates live models in the open workbook, handles scenario analysis and explains formulas, and OpenAI pairs it with market-data connectors including Moody’s, FactSet and MSCI.

The structural cost is still the trust boundary. Workbook content reaches OpenAI infrastructure rather than staying in your tenant. On the no-cost tier, conversations may be used for training, which rules that tier out for business data under the Office of the Privacy Commissioner of Canada principles (2023). Business and Enterprise exclude inputs from training by default under SOC 2 Type II controls.

Seat price: ChatGPT Business is US$20 per seat on annual billing or US$25 monthly, with a two-seat minimum. OpenAI cut that seat price by US$5 on April 2, 2026, so any quote written before that date overstates today’s cost. Enterprise is quote-only and 2026 procurement reporting clusters it between US$45 and US$75 per seat with a 150-seat minimum. Source: OpenAI published plan pricing, 2026.

Chart generation deserves its own mention. Asking for a bar chart with monthly grouping, a log y-axis and a regression line produces a usable image in one turn. Rebuilding that in native Excel takes several manual steps plus a calculated column, so the export speed is the win.

What Claude does well for spreadsheets

Claude for Excel closed the reach gap on May 7, 2026. Per Anthropic’s documentation (2026), the add-in is generally available on Pro, Max, Team and Enterprise. It answers questions about the open workbook with cell-level citations, adjusts assumptions while keeping formula relationships intact, traces the root cause of errors and populates or builds models across multiple tabs.

The cell-level citation behaviour is what analysts notice first. Ask how the revenue figure in C42 was derived and the answer links back to each referenced cell. Ask it to flex a growth rate and it shows which downstream cells recompute before it writes anything.

Two limits matter for Canadian SMBs. Anthropic lists data tables, macros and VBA as unsupported capabilities, so legacy automation still needs the copy-and-paste review workflow. Model choice is also a curated subset rather than the full bench: Claude Opus 4.7, Opus 4.6 and Sonnet 4.6 inside the add-in.

Plan detail: Claude Team is US$20 per seat on annual billing or US$25 monthly, sold from two seats up to 150. A Premium seat adds Claude Code and Cowork at US$100 annual. Enterprise is published at US$20 per seat plus metered usage at API rates. Per Anthropic’s pricing page (2026), Claude for Microsoft 365 is included from Pro upward at no extra charge. Source: claude.com/pricing.

Context depth is worth stating precisely, because vendor marketing blurs it. Anthropic’s model documentation (2026) gives Claude Fable 5, Opus 5 and Sonnet 5 a 1M-token window on the API. The claude.ai subscription plans that drive the Excel add-in are documented at 200k. For a 4,000-line macro workbook, that difference decides whether the review fits in one pass.

Curious how this looks against a real client workload? Talk to Fusion → and we will map it against your team’s workbook portfolio, with PIPEDA-aware governance handled first.

Use case by use case: who wins what

Eight jobs cover most of what a Canadian finance or operations analyst does in a week. We benchmarked all three tools against the same workbook portfolio during Q2 2026, and the ranking below reflects an FC internal benchmark from Q2 2026 rather than vendor claims. Two of the eight rows changed hands during 2026, which is why a 2025 shortlist is now misleading.

Formula generation (SUMIFS, INDEX/MATCH, XLOOKUP)

Winner: Copilot. The in-grid prompt reads neighbouring cells and existing formulas, so the suggestion lands in context. The other two now do this from their own panes, though they still ask the analyst to confirm ranges more often. Time saved per formula is roughly 60 to 90 seconds.

PivotTable building

Winner: Copilot. Copilot builds the PivotTable in place, refreshes it against the source range and rearranges fields by prompt. Claude for Excel can now edit pivot tables directly too, which is new in 2026, but Copilot still handles refresh and source-range changes more cleanly.

Data cleanup at scale (10,000+ rows)

Winner: ChatGPT. Advanced Data Analysis with pandas handles raw exports quickly. A prompt to drop null rows, deduplicate on a key column and normalise postal codes to A1A 1A1 finishes in three turns. The .xlsx download lands ready for Excel.

Chart creation

Winner: ChatGPT. Python matplotlib generates a chart in one turn with full control over axis labels, legend placement and styling. Copilot produces a chart inside the workbook, which is often what the analyst actually wants, though iteration lags. In our practice a monthly board pack of 12 charts is where that gap becomes visible.

Formula auditing and debugging

Winner: Claude. Drop a tangled INDIRECT and SUMPRODUCT stack in, ask it to trace the calculation for cell K47, and Claude walks the chain with clickable cell citations. That citation behaviour is the reason auditors reach for it first.

Custom function design (LAMBDA, LET)

Winner: Claude. Best at recursive design and edge-case enumeration. Copilot generates LAMBDA syntax but treats it like a regular formula, so base cases and termination conditions get less scrutiny. On a 60-line recursive LAMBDA our engineers found Claude the only one of the three to flag the missing termination branch.

Power Query M code

Winner: Claude. Long M-scripts reward careful reasoning. ChatGPT places second. Copilot places third because the Power Query editor still does not carry the Copilot surface that the main grid does.

VBA and Office Scripts

Winner: Copilot. This row flipped in 2026. Claude for Excel excludes macros and VBA from the add-in, so legacy refactoring means pasting code into the chat product instead. Copilot generates Office Scripts natively in Excel for the web.

Why this matters for Canadian SMBs: Analysts in regulated firms often have one workbook that sits at the centre of their job. A wrong tool choice costs hours every week and can move regulated data outside the tenant without an audit record. Sources: microsoft.com/en-ca, openai.com, claude.com, priv.gc.ca, learn.microsoft.com.

Pricing reality for a 25-seat finance team

Run the arithmetic on a 25-seat Canadian finance team. Microsoft publishes Canadian list prices, so the Copilot line needs no conversion. OpenAI and Anthropic bill in US dollars, and the figures below convert at the August 4, 2026 rate of roughly 1.41. Every line is the AI seat only, because all three assume the team already owns Excel.

• Copilot path. 25 seats at CA$28.50 is CA$712.50 a month, about CA$8,550 a year, on top of a qualifying Microsoft 365 plan. A promotional annual rate of CA$24.43 for tenants under 300 seats brings that to CA$610.75 a month while it lasts. Our Microsoft 365 Copilot pricing in Canada 2026 breakdown tracks the current window.
• ChatGPT Business path. 25 seats at US$20 annual is US$500 a month, roughly CA$705. Standalone, no underlying suite required, and the add-in works on any current Excel build.
• Claude Team path. 25 seats at US$20 annual is US$500 a month, roughly CA$705. Same shape as ChatGPT, and the same absence of tenant-side audit logging.

The three paths now land within about CA$10 a month of each other at 25 seats. Price stopped being the deciding variable in 2026. What separates them is whether the workbook stays inside a boundary your privacy officer already signed off.

Version and licence requirements before you install any of the three

Half the failed pilots we see are a licensing or build problem rather than a model problem. Each add-in has a hard floor, and two of the three quietly exclude older Excel installs entirely. Run this checklist before a single seat is purchased, because a blocked install after the purchase order is a painful conversation to have with a finance director.

• Copilot Agent Mode. Requires a Microsoft 365 Copilot licence on top of a qualifying base plan, and the workbook must be saved to OneDrive, OneDrive for Business or SharePoint. Agent Mode does not run on Office 2019, Office 2021 or Microsoft 365 Basic.
• Claude for Excel. Runs on Excel for the web, Excel on Windows with a Microsoft 365 subscription at build 16.0.13127.20296 or later, and Excel on Mac at version 16.46 or later. Excel 2016 and 2019 perpetual or volume licences are unsupported, as are iPad and Android.
• ChatGPT for Excel. Available across every ChatGPT plan since May 5, 2026, including Business and Enterprise. Confirm regional availability with OpenAI before promising it to a team.
• Add-in deployment control. All three are installable by an end user unless an admin restricts the Office Store. Set that policy in the Microsoft 365 admin centre first, then deploy the approved add-in to a named security group.
• Audit and retention. Confirm which surfaces write to Purview. Anthropic documents that Claude for Excel activity is excluded from Enterprise audit logs and the Compliance API, and that it does not inherit an organisation’s custom retention settings.

Canadian residency, PIPEDA, and Law 25 implications

Three rules govern this for Canadian SMBs, and the 2026 add-in wave sharpened all three. An assistant that lives in the ribbon lowers the effort required to move regulated data outside the tenant, so the policy work now has to run ahead of the install rather than beside it.

• One. Personal information sent to a free-tier consumer AI is a privacy risk. Per the Office of the Privacy Commissioner of Canada (2023), organizations must document a legal basis for AI data collection and prefer anonymized or synthetic data.
• Two. Residency configuration differs by vendor. Copilot inherits the Microsoft 365 tenant region, so a Canada Central or Canada East tenant keeps prompts and grounded data inside that geography. ChatGPT Enterprise offers Canadian residency via a signed DPA. Anthropic offers zero data retention on Enterprise rather than a Canada-specific region.
• Three. Quebec Law 25 raises the bar for cross-border transfers. The privacy impact assessment requirement applies whether the server sits in Quebec or in Texas. Copilot’s inheritance of Microsoft’s existing Canadian commitments simplifies that materially.

Security note: Anthropic publishes an explicit prompt-injection warning for Claude for Excel. Downloaded templates, vendor files and data imports can carry hidden instructions that manipulate the add-in into extracting data or performing destructive actions. Anthropic also advises against using it for audit-critical calculations without verification. Source: Anthropic, Use Claude for Excel (2026).

That warning generalises beyond Anthropic. Any assistant with write access to a workbook inherits the trust level of every file the analyst opens. A supplier pricing template downloaded from a portal becomes an input to an agent that can edit records, which is a PIPEDA exposure a 40-seat firm rarely has a control for.

If acceptable-use policy is next on the queue, our piece on what should be in an AI acceptable-use policy sets out the Canadian clauses worth including before licensing any of these three. For deeper Microsoft 365 governance work, the Copilot SharePoint oversharing audit is the one most pilots skip.

Need help running the residency review across your client portfolio? Get in touch → and we will structure the conversation around PIPEDA exposure first and the tool stack second. Reviews for Toronto, Hamilton and Metro Vancouver SMBs are CISSP-led at a Microsoft Solutions Partner.

The recommendation we give 25-100 seat Canadian SMBs

The default stack is Microsoft 365 Copilot for everyone who opens a workbook daily, plus 3 to 8 Claude Team seats for dense audit work. That split is an FC internal benchmark from Q2 2026, built from anonymized client data across 25-to-100-seat finance-team engagements. In our practice the split held even after both challengers shipped add-ins. Source: Fusion Computing engagements, 2026.

“People expected the May releases to change our recommendation. They did not. Copilot still takes the company-wide slot, because it is the only one whose Excel activity lands in Purview. Claude earns a small analyst pool for formula tracing. The change is that the governance work now has to happen before the add-in appears, not after the licence renews.”

Mike Pearlstein, CISSP, CEO, Fusion Computing.

The decision flow we walk clients through:

1. Microsoft 365-centric team with Excel as the daily surface: start with Copilot. The in-grid surface plus tenant residency make it the safest default.
2. A subset doing dense formula audit or model review: add 3 to 8 Claude Team seats for that subset, with the add-in scoped to a security group.
3. Marketing or research needing ad-hoc cleanup and chart export: add 2 to 5 ChatGPT Business seats for that subset.
4. Team not on Microsoft 365 at all: ChatGPT Business becomes primary, Claude Team second, Copilot deferred until the suite question is settled.

This echoes what we see across the broader assistant decision. The parent platform comparison handles the company-wide choice; this post stays on the Excel surface.

Get in Touch

Common selection mistakes

Four errors show up repeatedly in Canadian SMB shortlists, and none of them are about model quality. Three of the four became more expensive in 2026, because an add-in that installs from AppSource in 30 seconds removes the friction that used to keep unapproved tools out of a regulated workbook.

• “Copilot covers Excel, so we are done.” Copilot wins inside the grid and owns the audit trail. It still trails on long Power Query pipelines and on multi-tab model tracing. Across our client base, most teams that start Copilot-only add a small Claude pool within six months for the analyst subset.
• “An add-in means the data is safe.” The opposite is closer to true. A ribbon button removes the pause that an upload used to force. Confirm where the content goes, whether it lands in Purview, and whether the vendor retains it.
• “Free tiers are fine for non-sensitive data.” The Office of the Privacy Commissioner of Canada principles apply to any personal information, including the names, emails and salary figures common in HR workbooks. A free tier processing that data can put the organisation offside under PIPEDA.
• “Standardise on one assistant, no exceptions.” Single-tool standardisation on Excel work is the most common mistake we see. The workbook portfolio is heterogeneous, so one tool rarely covers formula generation, audit, cleanup, charting and macro review equally well.

FAQ

Related Resources

• Agents in ChatGPT for Work
• Custom business AI platform
• RAG for business explainer
• AI knowledge management playbook

Contact Us

Last updated: August 2026 · Reviewed by Mike Pearlstein, CISSP · Fusion Computing