Home/Insights/AI in Finance

I Tested Claude for Excel: It Fixed My Broken P&L Model

By Ani BjörkströmPublished 16 August 2026Reviewed 17 November 202616 min video + articleAI in Finance

I Tested Claude for Excel: It Fixed My Broken P&L Model
▶ Watch the full 16-minute video · with chapters & transcript

AI IN FINANCE

Can Claude for Excel Fix a Broken P&L and Catch a Budget Breach?

In short: Claude for Excel flagged a travel-and-expense breach of 40.5% against an 8% policy limit — a SEK 42,000 overspend — and named the department driving it, in under a minute.

Key takeaways

  • Claude for Excel flagged a travel-and-expense breach of 40.5% against an 8% policy limit — a SEK 42,000 overspend — and named the department driving it, in under a minute.
  • On a P&L tab with three broken formulas, Claude diagnosed two divide-by-zero errors and one caused by a deleted reference column, then repaired all three on request.
  • Given 34 months of cash history, Claude built a full 2026 monthly revenue-and-cash forecast with adjustable growth and seasonality assumptions, including a manual flag for July as a Nordic factory-shutdown month.

In a hands-on tutorial, Ani Björkström, a Stockholm-based tech consultant who works with asset management and private equity firms, tests Claude for Excel against seven finance tasks that typically eat up a controller's or analyst's week. The workbook belongs to a fictive company, Northcomp AB — a mid-size Nordic distributor with 240 employees and SEK 310 million in annual revenue, operating across Sweden, Norway, Denmark and Finland. The data is synthetic, but the tasks — policy compliance, broken formulas, legacy models, messy GL exports, forecasting, valuation and board reporting — are drawn directly from day-to-day finance work.

How does Claude catch a policy breach inside a live spreadsheet?

Claude for Excel is opened from the Start ribbon and requires at least a Claude Pro account. Once active, it reads the formulas and data on whichever tab is currently open — so the video stresses staying on the correct tab before prompting. The first test asks Claude to check whether combined travel, hotel and meal spend stays under an 8% policy cap and to identify the biggest driver. Claude adds a policy-compliance breakdown directly into the sheet: total T&E spend of 58,155 against total operating costs works out to roughly 40% of the total — five times the limit — and it lists the spend by department (sales, management, operations, finance), pointing straight to where costs need to be reined in.

Can it actually repair a broken P&L, not just describe the problem?

On the P&L tab, three cells return errors — a recurring monthly headache for any finance team. Asked to find every broken cell and explain the cause, Claude identifies two divide-by-zero errors and a third caused by a missing reference column, then explains what to fix. Björkström fixes the first two manually to preserve understanding of the change, then asks Claude to fix the remainder directly — a shortcut recommended only outside production environments, with manual verification afterward non-negotiable.

What about legacy formulas and messy GL data nobody wants to touch?

A "legacy allocation" tab, built by a departed colleague and left untouched out of caution, allocates HR and facility costs across the four Nordic countries. Asked to explain the column E formula step by step and judge whether the logic holds up, Claude confirms the allocation logic is sound and walks through the calculation. Separately, a raw general-ledger export — with inconsistent department naming, negative amounts stored as text in parentheses, and likely duplicates — gets a full cleanup: Claude standardizes dates and department names, fixes the amount formatting, removes exact duplicates, and builds a GL Clean sheet plus a GL Summary sheet by month and account, leaving the raw export untouched as a source record.

TaskPrompt focusResult
T&E policy check8% cap on travel, hotel, meals40% of total cost, SEK 42,000 over budget, by department
P&L repairFind and fix broken cells3 errors diagnosed (2x div/0, 1 missing reference) and corrected
Legacy allocationExplain formula, judge soundnessLogic confirmed sound, step-by-step walkthrough produced
GL export cleanupStandardize dates, names, amounts; dedupeNew GL Clean and GL Summary sheets, source data preserved
2026 forecast & DCFMonthly forecast, seasonality, valuation, chartAssumptions, Forecast and DCF sheets plus a revenue chart by region

Is it ready for a CFO forecast or a board-ready DCF?

Using 34 months of cash history, Björkström asks Claude to build a monthly 2026 revenue-and-cash forecast, placing growth and seasonality assumptions on a separate, editable Assumptions sheet — noting that July is a factory-shutdown month in the Nordics, a detail Claude would not know without domain input. Claude returns an Assumptions sheet with adjustable inputs and a seasonality table, plus a live Forecast sheet. From there, it builds an indicative DCF valuation using the same historical financials and assumptions, and finally generates a bar chart of revenue by region on the Revenue tab, ready to be copied into a board presentation.

FAQ

How do you access Claude inside Excel?

It launches from the Start ribbon, requires signing in, and needs at least a Claude Pro account; once open, it reads the tabs and formulas of the currently active sheet.

Was the financial data in the demo real?

No — Northcomp AB is a fictive mid-size Nordic distributor created for the tutorial, with synthetic figures, not client data.

Should you let Claude edit a live model without review?

No. The video recommends using "ask before edit" rather than "accept all edits" in production, working from a copy of the workbook, and always reading Claude's reasoning and verifying outputs before presenting them to decision-makers.

Want this working inside your finance team?

Ani Björkström

Ani Björkström — founder of QvantX Sweden AB, a Stockholm consultancy building AI solutions for banks, asset managers and finance teams. Anthropic partner. Every article starts from a real client build, minus the confidential parts. LinkedIn →