Our July 2026 Excel comparison used a deliberately messy synthetic workbook. The evidence review explains the judging limitations and why figures and task coverage need separate checks.
Evidence review: 4 October 2026. The retained scoring record is dated 31 July 2026. This is a review of that historical exercise, not a new test of current products. We have corrected the earlier description of fully blind, independent scoring and removed unverified model versions and the headline winner.
The workbook and the tasks
The exercise used a synthetic management workbook containing deliberately inconsistent labels, mixed date formats, numbers stored as text and other defects. A written answer key existed alongside the workbook. No real client financial data is reproduced here.
Tasks covered aggregation, charting, formula explanation, cleaning the workbook, calculating a margin and identifying a management concern. These require different checks. A chart can be legible while its source total is wrong; an accurate calculation can still leave much of the requested work unfinished.
The retained files include the original workbook, one working copy per tool, the answer key and a detailed scoring record. The record says that one copy per tool was used across the tasks. That differs from the planned fresh copy for each task, so earlier changes could have affected later results.
What the scoring record reports
The contemporaneous record reports that Copilot returned an incorrect regional total on the aggregation task and that Claude addressed inconsistent labels. It also records Claude cleaning only one of four worksheets in the clean-up task, while Copilot covered all four. Those are useful observations about accuracy and coverage.
However, the record itself says that its objective-result summary was derived from the judge's rationales and needed confirmation against the raw task outputs before publication. The retained task packets still contain output placeholders. A final workbook is not a complete record of each earlier answer. We therefore do not present the exact error amount or the overall score as independently reverified findings.
This distinction also applies to the management question. The notes report both tools identifying the intended issue, but the judging conditions limit what can be concluded about their relative performance.
How judging actually happened
The scoring record explicitly describes a fixed assignment: Output A was Copilot and Output B was Claude throughout. All six tasks were evaluated in one judge conversation. Tool names were stripped, but the mapping was not independently randomised for each task. An early judgement could therefore affect later scores.
More seriously, the notes say that the answer to the final management question appeared in an earlier evaluator note. By the time that task was scored, the judge had already seen the intended answer. Calling that final evaluation blind was incorrect.
The original article reported a 5-to-1 result with mean scores of 8.5 and 6.4. Those numbers were recorded, but the design does not support treating them as a robust overall ranking. We have removed them from the headline and excerpt. A language-model judge is also capable of error; choosing a different vendor does not establish neutrality.
What to check in an AI-assisted spreadsheet
Reconcile totals to source rows. Confirm how inconsistent labels and missing values were handled, especially where two categories look identical on screen. Check whether the answer uses the requested basis: the average of monthly percentages and a margin calculated from combined totals can be different measures.
For clean-up tasks, inspect coverage across every requested worksheet. Ask the reviewer to record the cells changed and any assumptions. A tool should not quietly invent a replacement for a genuinely unknown value.
Keep the original workbook and compare it with the saved result. An attractive summary or a plausible explanation is not enough to establish that the underlying cells are correct. Assign a person who understands the figures to approve the output before it is used in a management report. A worked monthly reporting pack demonstration on made-up figures shows a variance table and draft commentary alongside the questions a reviewer still needs answered.
Limits of this historical comparison
The version and licence fields were not completed in the retained scoring record. The files support discussing the test design and its reported observations, with the limitations above. They do not establish current product superiority, measured savings or a reproducible performance benchmark. No new runs were performed for this review.
The Word review provides a narrower example where a change can be checked directly in the retained source and output files.
Choosing tools for your team now
These are historical observations, reviewed on 4 October 2026. They do not establish which current subscription or model will work best in your organisation. Microsoft says the Copilot experience depends on licensing, application and tenant configuration. Anthropic documents a Microsoft 365 connector for Claude, whose setup and permissions also matter. Check the actual experience available to your staff before buying another tool.
Our current Claude vs Copilot guide starts with the work your team needs to do and the systems you already have. AI at Work also compares Claude and Copilot for specific kinds of work. For help assessing a specific task, talk to us. Training materials, source files and answer keys are supplied privately as part of an agreed commercial engagement.