Our July 2026 Word comparison shows why document checks matter. Review what was removed, factual accuracy and time to a usable result before choosing a tool.
Evidence review: 4 October 2026. This article originally presented a six-task winner and a speed comparison. The retained files support the document-removal observation below, but the run log does not contain completed timings or the full scoring record. We have removed the overall winner, precise speed claims and unverified model labels. This is a historical case, not a fresh benchmark.
What the exercise covered
The July exercise compared Claude and Copilot working in Word. Its written protocol covered executive summaries, rewriting for a different audience, identifying inconsistent figures, turning commitments into actions, a short senior-reader summary and document clean-up.
The protocol required a separate document copy for each tool and task. It also recorded a contaminated first attempt where tools shared a document; those outputs were excluded by the protocol. That distinction matters: a tool that sees an earlier tool's work is no longer starting from the same material.
We retained the source documents and individual run files. The written plan also specified blind AI scoring, but a plan does not establish that every judging step happened as intended. The retained run log has blank time, iteration and final-output fields. We cannot use it to substantiate a claim that one product was consistently faster.
A document can look cleaner while losing information
The source for the clean-up task contained an internal note asking for a figure to be checked, along with an embedded exchange about a draft figure. Neither passage appears in the retained cleaned document from either tool.
That is a directly checkable difference between the source and the saved outputs. It does not establish what either tool said in its chat pane, because the saved documents do not contain that conversation. The earlier wording that the tools removed material without telling anyone went beyond the retained evidence.
For a working document, a review note may be necessary even if it should disappear from the finished version. The reviewer needs to know whether the underlying question was resolved before accepting its removal. Making a paper easier to read does not settle an outstanding factual check.
How to assess a Word result
Read the changed document beside the source. Check whether qualifications, unresolved comments and decision conditions survived. Use Word's comparison or version history where available to locate changes, then have the document owner decide which are acceptable.
For summaries, trace the important figures back to their source. A shorter document can preserve every number and still omit the condition that changes the decision. For action plans, distinguish an owner named in the source from one inferred by the assistant.
Time the complete task through human review. A quick first draft that needs extensive correction may cost more time overall. Conversely, a carefully reasoned answer may be less useful for a small repeated edit if it interrupts the person doing the work. This is a measurement to make in your own workflow; the surviving records do not support a comparative timing claim here.
What this case can establish
The retained files support a specific lesson about checking removed content. They do not establish an overall Claude or Copilot winner, a reliable model ranking or a productivity saving. Licence details, application builds and a complete timing and judging record would be needed for a reproducible performance comparison.
The Excel article explains a different evidence issue: a detailed scoring record still needs to disclose how judging actually happened.
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. 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.