Excel AI Agents: Which Tools Run Multi-Step Workbook Work

The Excel AI tools that plan and execute multi-step workbook work instead of returning one formula are Viktor, Copilot in Excel (formerly Agent Mode), Claude for Excel, ChatGPT for Excel and add-ins such as Shortcut and ElyxAI. For a recurring report in November 2026, Viktor, an AI employee who lives in Slack and Microsoft Teams, handles the scheduled work by building Excel files with formulas on a schedule.

The target job is the weekly or monthly sheet somebody rebuilds by hand. The FP and A Trends Survey 2025 reports that teams relying on spreadsheets spend 50% of their time on manual data collection and validation and 28% on insight.

One question decides the choice. A tool has to run that sheet without anyone opening Excel, and Microsoft’s Copilot in Excel FAQ describes edits made inside an open workbook and documents no scheduled runs, so a recurring report needs a scheduler on top of any in-workbook tool.

What an Excel AI agent is

A multi-step Excel AI tool reads the workbook, edits cells and formulas, builds the report and checks its own output before handing it back. Microsoft’s Copilot in Excel FAQ says its edit mode lets Copilot “plan, run, and verify results directly in your workbook.”

A formula generator or chat tool returns one answer and leaves the person at the keyboard to apply it.

If a task needs more than one edit to finish, treat it as agent work. If it needs one formula, a chat answer is enough.

How an Excel AI agent works: plan, act, verify

Many Excel AI tools follow some or all of the same loop: read the workbook, plan, edit cells, check the result and report the changes. Vendors differ in which of these steps they document and which steps pause for a person’s approval.

The loop behind a task

The loop runs from a plain-language request to an edited workbook:

1. Read workbook context: The tool scans tabs, headers and formulas first; Claude for Excel reads complex multi-tab workbooks, and the SpreadsheetBench 2 authors found insufficient spreadsheet inspection was a dominant cause of failed tasks.
2. Plan: Copilot in Excel has a plan mode to “come up with a plan before editing your workbook,” and Anthropic says Claude in Excel can plan before acting.
3. Ask, then act on cells: ChatGPT for Excel asks for permission before changing a workbook, and Claude for Excel asks you to confirm before a risky operation runs.
4. Verify and retry: The AI Agent for Excel listing says the tool verifies AI results and retries before delivering them.
5. Report the changes: Claude for Excel tracks and explains its changes and links to the cells it references.

When comparing tools, ask which steps pause for approval and whether the change log shows every edited cell.

One prompt, five steps

The prompt below is an illustration for a weekly sell-through report, not a benchmark. It assumes a workbook with a Raw tab holding the week’s order export, a Settings tab holding the reorder threshold and a Summary tab.

Take the order export in the Raw tab. Remove duplicate order lines. Calculate sell-through by SKU as units sold divided by units received. Flag every SKU below the threshold in Settings. Chart the ten lowest SKUs on the Summary tab. Write a three-line summary at the top of Summary.

A tool that completes the loop makes these edits:

1. Dedupe: The tool removes duplicate order lines from the Raw tab and moves them to a Removed tab so a reviewer can check the deletions.
2. Calculate: The tool adds a Sell-through column with a formula in every row, not pasted values.
3. Flag: The tool applies conditional formatting to mark SKUs below the threshold, referencing the Settings cell instead of a typed number.
4. Chart: The tool adds a chart of the ten lowest SKUs to the Summary tab.
5. Summarize: The tool writes three lines at the top of Summary naming the weakest SKUs and the change from last week.

For Copilot in Excel, Claude for Excel and ChatGPT for Excel, test this illustrative prompt in an open workbook before relying on the edits. Confirm a scheduling route before expecting the report to run again next Monday.

Agent vs Copilot chat vs formula generators

Viktor, Copilot in Excel’s edit mode, the Claude and ChatGPT add-ins, Copilot’s chat mode and formula generators handle a workbook in different ways:

Microsoft retired its own COPILOT cell function on September 14, 2026, and points users to the Copilot pane. For a sheet that must refresh itself every week, choose a tool with a documented recurring run.

Tasks an Excel AI agent can execute beyond formulas

Recurring reports come first because they repay setup every week; the other jobs below run once per request:

• Recurring reports: Viktor runs the Monday report as a scheduled task set up once in plain English, and Claude Cowork scheduled tasks run in the cloud with weekly reports from spreadsheets or connected tools among the listed uses.
• Cleaning and reconciliation: In a Journal of Accountancy test on simulated data, an AI tool cleaned, validated and reconciled 59,157 rows of general ledger data to the trial balance in about 10 minutes on average.
• Budget vs actuals: Microsoft lists reshaping data, merging sheets and multi-element reports as Agent Mode work, and Claude for Excel adjusts assumptions while keeping formula relationships intact.
• Dashboards: Copilot in Excel creates PivotTables and charts from a description, and Viktor writes docs from live data that stay current at the same link.
• Model audit: Anthropic says Claude in Excel “audits formulas across linked workbooks, and runs sensitivity analyses”.
• Live market data: Copilot in Excel fetches information from the web, and Claude for Excel pulls context through MCP connectors such as S&P Global, LSEG and FactSet.

Hand over the report the team rebuilds most often. Give Viktor the Monday report: start free with up to $100 in credits, no credit card →

Which Excel AI agents are available

The options group into an AI employee, a native Excel mode and add-ins:

Viktor: an AI employee who owns the recurring sheet

Viktor is the AI employee who lives in Slack and Microsoft Teams, connects to 9,800+ tools, and does the work. Teams can also start in the web app at app.viktor.com, and Discord teams can run him in their server.

Illustrative only: for the recurring sheet, the setup is one plain-English request turned into a scheduled task, such as “Every Monday, pull last week’s charges from Stripe, expenses from QuickBooks and closed deals from HubSpot, build the weekly revenue workbook with formulas, and post it in #finance.” This is an example request, not a documented customer workflow.

He connects to those sources through the integrations directory, most through one-click connections, and reads spreadsheets, PDFs and decks people send him. He works on his own cloud computer and hands back the Excel file in the thread, which is the core of what an AI employee is: a hire who takes the whole job.

Viktor is an AI employee who waits for approval before he changes an account. Approval cards in Slack or Microsoft Teams carry Approve and Reject buttons, and a rejected action never runs; Viktor in Microsoft Teams supports scheduled tasks and approval cards.

Hampton connected 24 tools in 44 days, handed reporting to Viktor and gave him 26 recurring jobs a week, avoiding $440K in planned hires. Read Hampton’s story →

Viktor’s strength is the schedule: the report runs without anyone opening Excel. He returns a finished file rather than editing inside a workbook someone has open, while the add-ins below edit live workbooks in a session and their vendors document no schedule for the add-in itself. Choose Viktor when the sheet draws on tools outside Excel; choose an in-workbook add-in when the work lives inside one model a person is editing live.

Copilot and Agent Mode in Excel

Copilot in Excel is Microsoft’s native multi-step mode, renamed in Microsoft’s March 10, 2026 release notes as Edit with Copilot in Excel (formerly Agent Mode). Microsoft’s January 2026 Excel blog declared it generally available on web, Windows and Mac, and an April 2026 Microsoft post made the agentic features the default for Microsoft 365 Copilot and Premium subscribers.

The Copilot in Excel FAQ lists four routes to access: Microsoft 365 Personal or Family with an AI credits plan, Microsoft 365 Premium, a commercial Microsoft Copilot subscription, or a Copilot Chat-eligible business or enterprise subscription. The same FAQ says editing works only with Calculation Options set to Automatic, and Microsoft’s Premium page says the file must be saved to OneDrive with AutoSave on.

Edit mode runs on Windows, Mac and the web, with iPad and iPhone support rolling out. The Microsoft material cited here does not give a general minimum build; querying local files needs Windows Version 2601 or Mac 16.105 or later. The January 2026 post said Agent Mode was not yet available in the EU or UK.

Business pricing on Microsoft’s Copilot Business page, pricing checked November 2026:

• Microsoft 365 Copilot Business add-on: originally $21.00, now from $18.00 per user per month, paid yearly; $25.20 per user per month on a monthly subscription.
• Microsoft 365 Business Standard with Copilot: $23.50 per user per month, paid yearly.
• Microsoft 365 Business Premium with Copilot: $32.00 per user per month, paid yearly.

Microsoft lists the July 1, 2026, to December 31, 2026, discount, annual commitment and first-year-only terms on the same page. The enterprise add-on is $30.00 per user per month, paid yearly.

For recurring work, Microsoft offers scheduled prompts in Copilot Chat, capped at 10 per user, and Copilot Cowork, which creates and edits Excel spreadsheets and runs prompts on a schedule. Neither page says a scheduled run can update a specific existing workbook, so test that path before relying on it.

Claude for Excel and ChatGPT for Excel

Claude for Excel is generally available on Pro, Max, Team and Enterprise plans. Individuals install the Claude for Microsoft 365 listing from AppSource, and admins deploy it through the Microsoft 365 Admin Center.

The add-in runs on Excel on the web, Windows build 16.0.13127.20296 or later, and Mac 16.46 or later, but not on Excel 2016 or 2019 perpetual licenses, iPad or Android. Anthropic lists data tables and macros and VBA as unsupported, and after the September and October 2026 Windows updates some users lost chat history and uploaded skills in the add-in.

Claude Pro costs $20 per month billed monthly, or $17 per month with an annual subscription ($200 billed up front), on Anthropic’s pricing page, pricing checked November 2026.

ChatGPT for Excel launched in beta on March 5, 2026, and became generally available across all plans on May 5, 2026. The same Microsoft add-in also covers Word and PowerPoint.

Free and Go plans get limited usage, Plus and Pro draw on the plan’s agentic usage limit, and Business and Enterprise usage follows plan credits after a free preview that ended June 2, 2026. A typical task uses 5 to 20 credits per message on personal plans. Go costs $8 per month and Plus $20 per month on OpenAI’s pricing page, pricing checked November 2026.

OpenAI’s help page lists the current limits: spreadsheet chats sit apart from ChatGPT history and memory, VBA and macros may not be fully supported, and outputs may be incomplete. Direct workbook control runs through Codex in the ChatGPT desktop app, for Microsoft Excel only.

Pick whichever add-in matches the AI plan the company already pays for, since both tie usage to that plan.

Other Excel add-ins

AI Agent for Excel, ElyxAI, Shortcut and RowSpeak (Excelmatic) sell usage as IQ points, credits or subscriptions, pricing checked November 2026:

ElyxAI’s English listing names Plus and Pro plans without prices, which conflicts with the German listing. ElyxAI describes its product as one that “executes complete workflows - not just single actions”, and AI Agent for Excel suggests cell edits users apply with one click. Shortcut scored 59.25% as a Verified entry on the SpreadsheetBench leaderboard (October 16, 2025).

Run the weekly report once on a trial and read the credit meter before committing.

When to use an agent instead of VBA, macros or Power Query

VBA, Power Query and Power Automate still win when the source and steps never change; a multi-step AI tool or an AI employee wins when the report pulls from several tools and needs judgment. The matrix sets each approach against a task:

Microsoft calls VBA “desktop-centric” and says it has no Power Automate connector, so all supported VBA scenarios involve a user attending the macro. Application.OnTime can schedule a procedure to run at a specified time in the future.

• Office Scripts run on a schedule through Power Automate flows, need an enterprise or educational license, and the Run script action is capped at 1,600 calls per user per day.
• Power Query imports, shapes and refreshes data, and Microsoft documents scheduled refresh in products such as Power BI rather than in Excel itself.

The add-ins do not replace macros. Claude for Excel lists macros and VBA as unsupported, though practitioner Chris Webb showed Copilot generating Power Query M code in a 2025 post.

Keep Power Query for a stable source that only needs refreshing. Hand the report to Viktor when it draws on tools outside Excel and someone still rebuilds it every week.

Licensing and cost

Microsoft licenses Copilot per user on an annual commitment, the add-ins meter usage, and Viktor bills shared credits. Each model budgets differently.

Microsoft 365 Copilot: Copilot Business is an add-on for organizations with eligible Business plans (Business Premium, Standard, Basic, Apps for Business or Teams Essentials), supports up to 300 licensed users per tenant and needs an annual commitment. Copilot in the apps is included with the license, while advanced experiences such as Copilot Cowork are billed by usage.

Add-ins: Shortcut and ElyxAI sell credits, AI Agent for Excel sells IQ points, and Claude for Excel and ChatGPT for Excel draw on the usage limits of the buyer’s Claude or ChatGPT plan, with prices listed in the add-in sections above.

Viktor: Viktor’s pricing is usage-based, with credits shared across the whole workspace and no per-seat fee. He is free to start with up to $100 in credits and no credit card, and the Team plan is $100 a month for 40,000 credits for the whole workspace, with unused Team credits rolling over (pricing checked October 2026).

The Activity view shows credit use by person and task, each scheduled task shows the time it saved and its credit cost, and spend limits and auto top-ups with a monthly cap help control usage. Put the weekly report on a schedule: start free with up to $100 in credits, no credit card →

Accuracy limits and data security

AI tools speed up assembly, and review still decides what ships. Microsoft reported that Agent Mode scored 57.2% on all 912 SpreadsheetBench tasks, against 71.3% for humans, and the leaderboard lists that score as Unverified.

Where agents get financial models wrong

Each failure mode below comes with one mitigation drawn from the same body of practitioner and vendor guidance:

• Formulas that drift across periods: ICAEW says AI is notorious for changing formulas as it goes across a forecast; check that each row uses the same formula in every period.
• Hidden sheets, rows and columns: ICAEW advises asking the AI whether it added any and then confirming by hand.
• Contradictory information about external links: In one ICAEW case the AI reported external links and then gave a wrong answer to the check it was asked to run; check external links yourself and remove any that are not needed.
• Conflicting inputs from different systems: When two sources report different figures, such as revenue in the CRM and in the accounting tool, a tool without instructions has to pick one; ICAEW found that when conventions, thresholds or acceptable exceptions are not made explicit, the model will apply its own assumptions; name in the prompt which source takes precedence when figures conflict.
• Broken references: The SpreadsheetBench 2 authors named incorrect target-cell selection among the dominant bottlenecks for agents, and after rows, columns or tabs move, a formula can end up pointing to the wrong cell or range; inspect for #REF! errors and trace precedents and dependents before sign-off.
• Invented figures: In Wall Street Prep’s test, in an article updated June 29, 2026, Claude and Shortcut hallucinated historical line items that still added up to correct subtotals; trace each input line, not only subtotals, back to the filing.
• Dead numbers: ICAEW’s Ian Schnoor described an AI tool that built 1,000 unique formulas containing hard-coded values; search the formulas for typed constants.
• Plugs: Wall Street Prep found all the tools it tested used plugs instead of linking the statements; confirm the balance sheet balances without a plug, as ICAEW advises.
• Large, complex workbooks: On SpreadsheetBench 2, with an average of 11.8 worksheets per task, the authors reported 34.89% accuracy and 12.00% on debugging as the best overall score and the low end of debugging accuracy, respectively; Microsoft suggests splitting a formula-heavy spreadsheet into smaller documents.
• Different answers on each run: Schnoor warns that an AI tool given the same problem can return a different solution each time; the Journal of Accountancy recommends a control run on a small dataset with known flaws and a pre-calculated clean output.

Before a recurring report goes to the board, run the control dataset through it and compare.

A review checklist before trusting the output

The checklist below applies to every first run and every changed prompt:

• Totals tie out to the source system for the period: bank, ledger, Stripe or CRM.
• The balance sheet balances without a plug.
• Every input traces to a source, using Claude’s cell-level citations or Copilot’s source superscripts.
• Follow ICAEW guidance: keep formulas consistent across periods, with no dead numbers or overlong formulas.
• Hidden sheets, rows, columns and external links are checked by hand.
• Every material variance has a written reason.
• A duplicate of the file exists before editing, as OpenAI advises, or prior versions are available to restore.

Sign off only when every item passes, and run the full list on the first three scheduled runs.

What data leaves your tenant

Microsoft 365 Copilot keeps prompts and responses within the Microsoft 365 service boundary and states that “Prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs, including those used by Microsoft Copilot,” on its privacy page. Copilot shows each person only data they can already access, and a Purview DLP policy on a sensitivity label disables Copilot skills in a labeled Excel file.

Two exceptions apply. Web grounding sends short queries to Bing, and the EU Data Boundary doesn’t apply to them; Anthropic models in Copilot are currently excluded from the EU Data Boundary, and admins manage that setting in the Microsoft 365 admin center.

Third-party add-ins can send workbook content outside Microsoft’s service boundary, depending on the add-in and its configured endpoint. The ChatGPT listing says the app can send data over the Internet, and OpenAI notes Microsoft may be able to read workbook content under its Marketplace terms.

Claude for Excel is not currently ZDR-eligible, and Anthropic’s Indonesian help page says it does not inherit organisation retention settings. Admins can clear Let users access the Office Store and deploy only approved add-ins centrally.

Prompt injection is a documented risk.

• Microsoft published CVE-2026-26144 on March 10, 2026, an Excel flaw that “could potentially cause Copilot Agent mode to exfiltrate data via unintended network egress, enabling zero-click information disclosure attack.”
• PromptArmor showed hidden white-on-white text in an imported dataset driving a spreadsheet tool to exfiltrate data, and Anthropic’s docs say “Only use Claude for Excel with trusted spreadsheets” on the Claude for Excel page.
• Action1’s Alex Vovk advises restricting outbound network traffic from Office apps until patches are in place.

Viktor holds SOC 2 Type 1, with SOC 2 Type 2 and ISO 27001 in progress, and customer data is never used to train models, by Viktor or his model providers (viktor.com/security, October 2026). On Viktor’s security page, “money moves, code pushes and customer emails wait for explicit approval in Slack,” the same approval cards appear in Microsoft Teams, and admins can mark any other integration tool as needing approval or forbidden. API keys are added at run time by a gateway, so the model never reads them.

Review Viktor’s security details, including SOC 2 Type 1 and approval controls →

Getting started: hand off the recurring report

The first handoff takes one sheet, one prompt and one reviewed run. Follow these steps in order:

1. Pick the sheet rebuilt most often. A weekly report repays setup every week; a monthly one, once a month.
2. List every source and its owner. Name each tool, export and tab the sheet draws on, such as Stripe charges, QuickBooks expenses and HubSpot deals.
3. Write the prompt in steps. State sources, calculations, flags, charts, the summary and where the file gets posted, as in the five-step sell-through prompt above. Alternatively, screen-record the manual rebuild with narration; Viktor turns the recording into an editable skill and proposes a scheduled task that stays paused until someone approves it.
4. Review the first run against the checklist. Compare it line by line with last week’s hand-built version.
5. Schedule it and cap the spend. Approve the scheduled task, set a spend limit, and read time saved and credit cost on the task each week.

Pick the sheet today and write its source list before the next rebuild. Hire your first AI employee: start free with up to $100 in credits, no credit card →

FAQ

How do I get Agent Mode in Excel, and which license does it need?

Microsoft’s March 10, 2026 release notes identify Agent Mode as Edit with Copilot in Excel, the default when Copilot opens in Excel. The Copilot in Excel FAQ lists eligible routes: Personal or Family with an AI credits plan, Premium, a commercial Microsoft Copilot subscription, or a Copilot Chat-eligible business or enterprise subscription, with availability subject to tenant configuration.

Is workbook data used to train AI models?

No, for Microsoft 365 Copilot, whose privacy page excludes prompts, responses and Graph data from foundation model training. Anthropic does not train on commercial product data by default, while its consumer plans (Free, Pro and Max) may use chats if the user allows it, if a conversation is flagged for safety review, or if the user otherwise explicitly opts in to training; Viktor never uses customer data to train models.

Can an Excel AI agent be trusted with a financial model?

No, not without review. Anthropic does not recommend Claude for Excel for audit-critical calculations without verification, and Wall Street Prep found AI tools “hide errors in places humans usually don’t look”.

How should a recurring report prompt be written?

Name the sources, the calculation steps, the checks, the output file and the channel where it gets posted, in that order. Viktor runs that prompt as a scheduled task; for in-workbook tools, Microsoft’s scheduled prompts and Claude’s Cowork scheduled tasks are separate features from the add-ins.

Should I use an Excel AI agent or VBA?

Use VBA for fixed steps inside one desktop workbook, since Microsoft says supported VBA scenarios involve a user attending the macro. Use an AI tool when the report pulls from several systems or changes shape; Claude for Excel lists macros and VBA as unsupported, so the two approaches sit side by side rather than one replacing the other.