The slide that ranked a rival
A procurement lead in London spent March running Claude and Copilot side by side on the same contracts, then wrote a careful recommendation. In July a Microsoft account team arrived with a slide that settled the same question in one line: the rival was slower, less accurate, and missing the proper security integrations. The slide was faster than the bake-off, and it cost nothing to produce.
Why it matters: that line did not come out of a laboratory. It came out of an internal sales meeting, and it was written to be repeated to you.
What was claimed, and who claimed it
Bloomberg reported that Microsoft used an internal meeting this week, billed as a strategy session for the new fiscal year, to train its salespeople to compare rival AI products unfavourably with its own. The rivals named were OpenAI, Google and Anthropic.
Jacob Andreou, the executive vice president for Copilot, presented a direct comparison with Anthropic's Claude and said that inside Microsoft's office applications the rival model was "slower and less accurate, and lacked the proper security integrations". Jay Parikh, an executive vice president, handed the sellers their framing: "Everyone else is selling parts - we're selling the full end-to-end system. That's the story that we all need to get out there and tell in FY27."
Read the second quote again. It is not a claim about performance. It is an instruction about narrative, and it names the fiscal year it was built to serve.
The second fact that changes the first
A separate report this month found that Microsoft has been removing OpenAI and Anthropic models from flagship applications including Word and Excel and replacing them with its own, and gave cost as the motive. Set the two facts side by side and the comparison stops looking like a product finding.
A vendor that is already replacing a third-party model to protect its own margin has a reason to rank that model low, and it briefed its sellers to do exactly that. The ranking and the substitution point in the same direction, and that direction is Microsoft's cost line rather than your workload.
Copilot may genuinely be better inside Word. Nothing in that meeting establishes it, and the party with the largest stake in the answer is the only party that ran the test.
Yes, but: every vendor sells against its rivals
Yes, but: competitive selling is not a scandal, and treating it as one is its own kind of error. Sales enablement has ranked rivals in every category for as long as there have been categories, and Microsoft is entitled to argue that an integrated system beats a pile of assembled parts.
What changed is the shape of the product. When the thing being ranked is a model that your vendor also resells, hosts, and can swap out beneath your applications, the comparison stops being a pitch and becomes a preview of which model you will end up running, chosen on someone else's margin.
Four questions before the claim reaches your evaluation
Ask which model version was tested, on which task set, on what date, and who ran it. A comparison missing any of those four is decoration, and any account team can supply them within a day if the test actually exists.
Then take the question out of the meeting room entirely. Run two weeks on your own documents, your own retrieval, your own security review, and score it yourself. The only benchmark that outlives a vendor's fiscal year is the one you ran on your own work.
The bottom line: put model identity in the contract. Name the model, require written notice before any substitution, and keep one fallback you have actually tested. The pitch is rewritten every July; your dependency should not be.
Read next: Two AI Giants Just Became Your Integrator | Match the Model Tier to the Job
