The best AI agent for Excel Automation in 2026

What to look for

• Works on the actual .xlsx file on your disk — opens it, edits it, saves it back — without requiring you to upload the workbook to OneDrive, SharePoint, or a third-party cloud. The buyer test: does the AI feature stay enabled when AutoSave is off and the file lives in ~/Downloads?
• Drives Excel as an application, not just a single cell — opens workbooks, switches sheets, creates pivots, applies formats, runs Power Query refreshes, exports to PDF — so a one-prompt request like 'clean column C, build a pivot by region, save as Q2-clean.xlsx' executes end to end without leaving the agent
• Permission-gated for any destructive action — overwriting a file, deleting rows, running a VBA macro, sending the file by email — with a visible preview of the change before it applies and an audit trail of every cell that was rewritten
• Handles real Excel messiness — merged headers, multi-tab workbooks, sheet protections, named ranges, formulas you do not want to silently overwrite with values, and number formats that differ across regions — rather than flattening everything to plain text on read
• Cross-platform — runs against Excel on both macOS and Windows with the same prompts and the same workflows, so a Mac analyst and a Windows ops lead can share the same skill library
• Writes the cleanup or analysis as an inspectable plan first (formulas it will add, columns it will delete, sheets it will create), not as opaque magic — so the user can correct the plan before any cell is touched, and can re-run the same plan on next month's export

Top tools compared

1. 1. Lapu AI

   High fit

   Built for desktop-native Excel automation on the file where it actually lives. Open an .xlsx in ~/Downloads, describe the job in plain English — 'dedupe by email, normalize the Date column to ISO, build a pivot by region, save as Q2-clean.xlsx' — and the agent reads the workbook, drafts the plan (which columns it would change, which formulas it would write, which sheets it would create), waits for approval, then executes against the real file. Works the same on macOS and Windows, against the Excel you already have installed, with no requirement that the file be saved to OneDrive or that AutoSave be on. Every destructive action — overwriting a workbook, deleting rows, running a macro, sending the result by email — is gated by an [explicit permission prompt](/blog/is-desktop-ai-safe-permission-models-explained) the first time the workflow runs. The [audit trail](/blog/ai-agent-audit-trail-explained) records every formula written, every range rewritten, and every file saved so a job can be replayed or rolled back. Bridges well to other formats: see [the PDF-to-Excel guide](/blog/pdf-to-excel-ai-on-desktop) for the common job of turning a folder of PDF invoices into a clean .xlsx without uploading anything. Where it shines: messy one-off cleanup; recurring monthly reports where the data shape drifts; jobs that span Excel plus the rest of your desktop (the agent can also read the source PDF, send the result on Slack, or commit the file to git). Where it is weaker: it is not a centrally-administered RPA platform with 24/7 unattended Excel bots — for that scale of orchestration, Power Automate or UiPath are the right shape.

   Learn more →

2. 2. Microsoft Copilot in Excel

   Medium fit

   Microsoft's first-party AI in Excel, included in the Microsoft 365 Copilot license. Strong at the in-cell jobs people most often ask Excel to do: explain a formula in plain English, generate a formula from a description, suggest a chart, summarize a table, highlight or sort data, and propose pivots. Where it shines: any team already paying for Microsoft 365 Copilot whose Excel work lives in OneDrive or SharePoint with AutoSave on — the integration is deep, the latency is low, and it knows the workbook context natively. Where it falls short for this task: it requires AutoSave with the file saved in OneDrive or SharePoint, and the Copilot button is greyed out on a local-disk file. That is a hard wall for the very common case of a one-off .xlsx in Downloads, a workbook with confidential data the team has not approved to sync to OneDrive, or a Mac analyst who does not want every spreadsheet round-tripping through Microsoft's cloud. It also does not drive Excel as an application across multiple files — it is scoped to assisting inside the currently open workbook.

   Learn more →

3. 3. Microsoft Office Scripts + Power Automate

   Medium fit

   Microsoft's modern replacement for Excel VBA: write or record TypeScript that runs against the Excel JavaScript API, then trigger it from a button, a Power Automate flow, or (until script scheduling returns) a Power Automate schedule. Cross-platform in the sense that scripts run inside Excel for the web on any device. Where it shines: stable, predictable repeating Excel jobs — turning a daily CSV into a formatted, pivot-bearing report; running the same cleanup on every weekly export; chaining Excel with the rest of the Power Platform (Outlook, SharePoint, Dataverse). Where it falls short for this task: Office Scripts targets Excel for the web specifically (scripts are stored in OneDrive and run in the cloud), so a workbook that sits on the local disk and is never uploaded is out of scope. It is also a coding tool, not an agent — the script does not adapt when the source file's columns drift; you go back to the editor. Power Automate Desktop adds Windows-only RPA on top of this stack, but the unattended desktop tier is per-bot priced ($150/bot/month historically), and the macOS audience is unsupported on the desktop-RPA side.

   Learn more →

4. 4. Bricks

   Medium fit

   Browser-based AI spreadsheet that takes uploaded data and generates dashboards, charts, and reports from a natural-language prompt — a Notion-style document around a Google-Sheets-like grid. Where it shines: turning a clean CSV or .xlsx into a polished visual deliverable (dashboards, charts, slide-shaped reports) without designing them by hand; teams whose Excel work ends in a presentation rather than a refreshed model. Where it falls short for this task: Bricks is a cloud product — you upload the file to clean it or visualize it, which is a non-starter for sensitive data with no third-party processor agreement, and the output is a Bricks document rather than a refreshed copy of the original .xlsx your downstream tools depend on. It is also visualization-first; deep Excel-specific automation (Power Query refreshes, formula rewrites, macro execution) is not the job it is built for.

   Learn more →

5. 5. Numerous.ai

   Medium fit

   Lightweight Excel and Google Sheets add-in that exposes ChatGPT as a spreadsheet function: write =NUM.AI("summarize this review", A2) in a cell, drag it down, and the formula runs for every row. Pricing starts around $10/month. Where it shines: bulk per-row AI work inside an existing Excel workflow — categorize 5,000 support tickets, extract sentiment from product reviews, write product descriptions from a SKU table, generate a formula from a plain-English description. Where it falls short for this task: it is a cell-level tool, not an agent — it does not open files, build pivots across sheets, run Power Query, fix merged headers, or save the workbook back under a different name; everything happens inside a single =AI call. For multi-step Excel automation across an entire workbook, you still need a person (or a real agent) driving Excel. Excellent complement to Lapu or to a built-in Copilot; not a replacement.

   Learn more →

Why Lapu AI is built for Excel Automation

Lapu AI is the rare Excel-automation tool that meets the buyer where their .xlsx actually lives — on the disk, not in OneDrive. Microsoft Copilot in Excel is the obvious first answer for in-cell help, but the moment the file is local (no AutoSave, no OneDrive sync) the Copilot button is greyed out; the very files most people work on every day fail the precondition. Office Scripts and Power Automate fix the recurring-job problem but live in the cloud and target Excel for the web. Bricks and Numerous.ai each cover one slice — visualization, per-cell prompts — but neither drives Excel as an application across a multi-step workflow. Lapu AI is a desktop agent: it opens the workbook where it sits, reads the actual sheet structure, drafts a plan that lists the formulas it would write and the columns it would change, waits for approval, then runs against the real file. The same agent that cleans the spreadsheet can read the upstream PDF, send the result on Slack, and commit the cleaned file to git — see [the PDF-to-Excel walkthrough](/blog/pdf-to-excel-ai-on-desktop) for the canonical cross-app version of that job, and [the file-organization step-by-step](/blog/automate-file-organization-with-ai-step-by-step) for the broader pattern of cross-app desktop automation. A practical decision framework: if your team already pays for Microsoft 365 Copilot and every workbook lives in OneDrive with AutoSave on, Copilot in Excel is excellent and you should use it. If you need a deterministic, scheduled Excel job that always runs against a known table shape, write an Office Script and trigger it with Power Automate. If your problem is a one-off messy .xlsx on your Mac or Windows machine, a confidential workbook your team has not approved to sync to the cloud, or a multi-step job that spans Excel plus a PDF plus a Slack post — that is what Lapu AI is built for.

FAQ

Why can't Microsoft Copilot in Excel automate my local .xlsx file?

Copilot in Excel requires the workbook to be saved in OneDrive or SharePoint with AutoSave enabled. Microsoft confirms this in their Q&A: on a file sitting on a local drive, the Copilot button is greyed out regardless of the license. For workbooks that live in ~/Downloads, on an external drive, or in a synced folder for a different cloud (Dropbox, Google Drive), Copilot in Excel does not engage. Lapu AI was built for that case — it drives the Excel you already have installed, against the file where it actually lives, without forcing it through Microsoft's cloud.

Does Lapu AI need Microsoft 365 Copilot or Office Scripts to automate Excel?

No. Lapu AI drives the Excel application directly through the OS — on macOS via Excel's AppleScript and accessibility surface, on Windows via UI Automation and the COM bridge — so a standard Excel install is enough. You do not need a Microsoft 365 Copilot license, you do not need OneDrive, and you do not need Office Scripts enabled in your tenant. Lapu calls its own AI model layer to plan the job; the Excel side is a normal automation surface. If you have Microsoft 365 Copilot too, the two complement each other — Copilot for in-cell formula help, Lapu for cross-file, cross-app workflows on the desktop.

Can Lapu AI handle multi-tab workbooks with merged headers and protected sheets?

Yes. Before any change the agent inspects each tab — detects the data region, identifies whether the top rows are merged headers or a header band, notes which sheets are protected, and reads which named ranges and formulas exist. The plan it presents calls those out explicitly: which sheets it would touch, which merged-header cells it would flatten (and to what convention), which protected sheets it would skip or ask you to unlock, which formulas it would preserve and which it would rewrite. Nothing is silently flattened, and a workbook that fails the check (a sheet protected with a password Lapu does not have) is reported back as a blocker rather than partially executed.

Will Lapu AI break formulas, named ranges, or number formats when it saves the file?

By default it preserves them. When the agent writes back to an .xlsx file it keeps existing formulas, named ranges, conditional formatting, and number formats intact. If a transformation would change a value that a formula depends on, the preview flags the formula so you can decide whether to recalculate, leave it in place, or convert the cell to a static value. If you explicitly ask the agent to rewrite a formula (for example, replace a 200-character nested IF with a LET) it shows you the rewritten formula before applying it, and the audit trail records both the original and the rewrite.

Can the agent automate Excel jobs that span more than one file or more than one app?

Yes — this is the case the desktop-agent shape exists for. A single Lapu workflow can open a folder of PDF invoices, extract the line items into a new .xlsx, run a Power Query refresh against an existing model, build a pivot, export a chart to PNG, paste it into a Slack message, and commit the cleaned file to a git repository — all on your machine, with permission prompts at each destructive step. For the specific PDF-to-Excel job, see the [PDF to Excel on the desktop walkthrough](/blog/pdf-to-excel-ai-on-desktop). Copilot in Excel and Office Scripts both stop at the workbook boundary; Lapu does not.

Does Excel automation with Lapu AI work on macOS, or only Windows?

Both. Lapu AI runs on macOS 12+ and Windows 10+ with the same prompts, the same workflow library, and the same permission model. The platform-specific layers differ — AppleScript and accessibility APIs on macOS, UI Automation and the COM bridge on Windows — but the agent presents one interface and most Excel workflows port across platforms unchanged. The genuine exceptions are platform-specific Excel features (running a VBA macro through COM on Windows, AppleScript-driven Numbers conversion on macOS); the agent flags those when a workflow includes them so you know up front.

How does this compare to Numerous.ai and other =AI-cell add-ins?

Numerous.ai and similar add-ins (Excel Formula Bot, Sheet+, AI Office Bot) live inside a single cell — write =NUM.AI("categorize this review", A2), drag it down, and the prompt runs per row. That is great for bulk per-row AI work inside a sheet you already have open. Lapu AI works at the workflow level: it opens the file, plans the multi-step job (clean, dedupe, pivot, format, save), executes against the workbook, and can chain to other apps. The two are complementary — many users keep a per-cell add-in for one-off categorization or summarization inside a sheet, and use Lapu for the cross-file, multi-step automation around it.

Is it safe to let an AI agent edit my Excel files?

Lapu AI is built around three answers to that question. First, every destructive action — overwriting a workbook, deleting rows, running a macro, sending the file by email — is gated by an explicit permission prompt the first time a workflow runs, and you can promote a step to auto-approve once you trust it. Second, the agent shows the plan before any cell is touched, so you correct it ahead of the change rather than after. Third, every cell that gets rewritten is recorded in an audit trail you can replay or roll back. See [the permission model post](/blog/is-desktop-ai-safe-permission-models-explained) for the full explanation. For the same question framed against cloud chatbots that upload your file, the answer is different: Lapu does not upload the workbook for storage — only the minimal context the model needs to reason about a step leaves the machine.

Related

• Best AI agent for Code Review →
• Best AI agent for Codebase Q&A →
• Best AI agent for Competitor Monitoring →
• Best AI agent for Data Cleanup →
• Best AI agent for Desktop Automation →
• Best AI agent for File Organization →
• More best-for guides →
• See Lapu AI use cases →
• Compare with alternatives →
• Pricing →