An office worker reviews analytics dashboards with a friendly humanoid robot assistant.
Microsoft says Copilot in Excel's finance scores are climbing. The company has now published the details of the test that produces them. In the first post of a new series called "Raising the Bar", the Excel team explains how FinanceBench works and the loop it uses to improve Copilot. The numbers are Microsoft's own, and no outside party has checked them. The grading design is still worth a close look, because it shows what Microsoft counts as good finance work from an AI.

An office worker reviews analytics dashboards with a friendly humanoid robot assistant. What FinanceBench is​

FinanceBench is one of more than 20 benchmarks Microsoft uses to measure Copilot in Excel. An earlier Microsoft post described it as a deep and challenging vertical benchmark focused on finance workflows, with finance-specific rubrics for correctness, structure, and auditability. That post also said it was developed with input and review from partners such as the Financial Modeling Institute, Microsoft Finance, and other finance professionals and customers.

According to the new post, the benchmark is curated from more than 1,000 eval cases. Finance professionals wrote and reviewed each one. The cases mimic multi-step work done by analysts, FP&A leads, and bankers:

  • three-statement models
  • discounted cash flow models
  • comparable market analyses
  • leveraged buyouts
  • driver-based forecasts
  • management reporting packs

Each case has four parts:

  1. Starting artifacts. A realistic model, data quirks included, plus any extra grounding data.
  2. Prompts. The task the user wants done, which can run over several turns.
  3. Expected outputs. A workbook or other artifact showing the ideal result.
  4. Rubrics. The itemized criteria used to grade the result.

Cases come from finance practitioners supplied by expert data partners, customer and forward-deployed engineer engagements, user research, and analysis of production signals. Microsoft also tags cases by workbook shape, source data, complexity, and features used. That lets it slice results for narrow scenarios, such as summarizing a revenue forecast in a large workbook with PivotTables.

How the grading works​

A single case typically has dozens of criteria. They go beyond whether the final number is right. They also cover formula relationships, model conventions, links between assumptions and drivers, and auditability, along with task completion, aesthetics, and Excel best practices.

Microsoft admits that writing good rubrics was often harder than building the rest of the case. It gives two examples of rubrics that failed:

  • Too strict. One early rubric demanded XLOOKUP. The model used INDEX and MATCH, which works equally well, and was marked wrong.
  • Too loose. A check confirmed a final figure but not whether it came from a linked formula or a number typed into the cell. Microsoft now checks both the method and the value.

Grading uses two mechanisms:

  • Codified graders check the workbook in code for deterministic items. Examples are whether a cell holds the expected value and whether a figure is formula-driven rather than hardcoded.
  • Agentic graders are AI models that inspect the workbook and judge softer criteria, such as whether the model is easy to follow or the formatting suits the task.

Criteria can be weighted before a final score is returned. Microsoft says finance experts review every case in Excel. They ask whether the expected output passes each rubric and whether each rubric fairly measures good work. Microsoft also runs tasks repeatedly to catch variation and re-grades identical outputs to check the grader agrees with itself. Criteria that return different verdicts for the same work are fixed.

The reported numbers​

  • From June to July, Copilot in Excel's pass rate on FinanceBench v2 kept climbing. Once it topped 80%, reaching 85.1%, Microsoft decided the benchmark needed to be harder.
  • Version 3 added fuller rubrics for aesthetics and formatting, based on production signals and failure analysis. It also added deeper expert review of multiple outputs.
  • Copilot in Excel has since improved its pass rate on the harder v3 benchmark by 31 points, through a mix of product changes, model upgrades, and skill optimization.
  • Work on v4 has already started.

Microsoft also tracks response quality, latency, and token consumption. Its reasoning is that a change that raises task success but sharply increases latency is not necessarily a win.

A worked example: the three-statement skill​

Microsoft walks through one fix, tested on 114 FinanceBench cases. Failure analysis found two problems:

  • Staged requests. Some users ask for a model in parts, such as the model now and a dependent schedule later. The old skill pushed to force the balance sheet to balance, so it over-completed the task. The skill now recognizes staged requests, leaves dependent cells blank and highlighted, and says what remains.
  • Interest calculations. The skill suggested beginning balances to avoid circular references, even when the task called for average balances. It now follows the convention set by the workbook or the user.

Across three-statement tasks, the updated skill improved pass rates by 12 points for GPT 5.6 and 7 points for Opus 5.5. These are Microsoft's figures. The post doesn't say whether "points" means percentage points, and it doesn't publish the starting or ending scores.

Both fixes are practical. Anyone who has built a three-statement model knows that circularity from average-balance interest is a real modeling choice, not a bug. A tool that silently swaps in a different convention creates extra work for the reviewer.

The agent-driven improvement loop​

Microsoft says more of this work now runs as automated loops, with human experts still deciding what good output looks like. There are three loops:

  • Learning from real usage. Agents analyze privacy-compliant production signals, such as recoverable errors, abandoned tasks, and dissatisfaction. They trace failures through Copilot's steps and check whether evals cover them. Problems that show up in both production and evals get fixed first.
  • Improving rubrics. Agents attempt each case several times, grade the results, and flag unfair or incomplete criteria. They propose revisions, and an expert reviews them.
  • Experimenting with changes. Agents propose prompt or skill changes and measure their effect on quality, latency, and efficiency. Microsoft says every change is reviewed before it ships, and that many experiments don't help but still rule something out.

The post doesn't describe the privacy methods or the thresholds used to pick signals.

Beyond Excel​

FinanceBench began with Excel work. Microsoft now runs it across Copilot in Excel, Copilot chat, and Cowork. It is also building cross-app scenarios. One example is a model built in Excel that feeds a PowerPoint pitch deck or a Word investment memo, using Microsoft 365 context, memory, and personalization.

This fits Microsoft's October 29 finance announcement. That announcement added more than 25 new finance skills and partner-built plugins across Microsoft Copilot. Per that post, accounting capabilities in Excel and Cowork are in Frontier, and finance skills in Copilot in Excel are generally available. The same post says Copilot's Excel context includes the workbook's change history. It can identify when edits were made and by whom, and it can reverse individual edits selectively.

What IT and finance teams should take from it​

Treat the pass rate as a vendor signal, not a guarantee. The score comes from Microsoft's private benchmark, graded partly by AI, with no public dataset or confidence intervals. Microsoft's own Copilot guidance, per the evidence reviewed for this piece, cautions that Copilot can make mistakes and that its output shouldn't drive sensitive decisions such as finance. A model passing most rubric items still needs a human review before it goes into a board pack.

The rubric list is a useful review checklist. The criteria Microsoft grades against make a sensible audit list for any AI-built workbook:

  • Are key outputs formula-linked rather than pasted values?
  • Does the interest calculation follow the stated convention, average or beginning balances?
  • Is an unbalanced balance sheet flagged instead of plugged?
  • Are unfinished staged items left visible?
  • Do assumptions connect to drivers?

Benchmark self-grading has limits. Microsoft acknowledges that if an expected output or rubric is wrong, scores can climb toward the wrong answer. Moving from v2 to v3 after hitting 85.1% also shows the benchmark is being tuned as the product improves. That is reasonable practice, but it means v2 and v3 scores can't be compared directly.

You can give feedback. Microsoft asks users to give a thumbs up or down, with a reason, on Copilot in Excel responses. It says this feedback feeds the loop.

Bottom line​

The post is a company blog and therefore promotional by nature. Even so, it shows an unusually concrete view of how Microsoft grades AI spreadsheet work. It rewards auditable, convention-following, formula-driven models rather than just right-looking numbers. For finance and IT teams deploying Copilot in Excel, those same standards are the right ones to apply when reviewing its output.

 

References

  1. FinanceBench: The loop that continually improves Copilot in Excel for Finance work Microsoft Copilot Community 2026-11-16T16:00:00+00:00
  2. FinanceBench: The loop that continually improves Copilot in Excel for Finance work | Microsoft Community Hub techcommunity.microsoft.com
  3. Microsoft Copilot for finance: Purpose-built for the way finance teams work | Microsoft Copilot Blog microsoft.com