Executives used an internal FY27 strategy session to coach sellers against Claude and ChatGPT on price and security controls, one week after the company began quietly routing Excel and Outlook prompts to models it built itself
Microsoft spent part of an internal sales meeting this week teaching its sellers how to argue against the artificial intelligence companies whose models are currently running inside Microsoft's own software.
The session, held Tuesday and billed as the strategy kickoff for the fiscal year that started July 1, laid out competitive positioning against OpenAI and Anthropic, with Google also named in the room, according to Bloomberg, which reviewed internal materials and a transcript. The argument Microsoft wants its account teams carrying into customer meetings for FY27 rests on two planks: it costs less, and it comes with the governance controls that enterprise buyers have to sign off on anyway.
Executive Vice President Jay Parikh gave the framing its slogan. "Everyone else is selling parts," he told the room, while Microsoft delivers the assembled system end to end. That, he said, is the story the whole organization needs to carry through FY27.
Copilot Executive Vice President Jacob Andreou took it further and put a name on the board. According to Bloomberg's account, Andreou presented a side-by-side comparison of Copilot against Anthropic's Claude inside Microsoft's Office applications and characterized the rival model as slower and less accurate on that surface. He also told the room it lacked the security integrations enterprise-grade work requires, and said his team is focused on widening the gap.
Anthropic declined to comment on the report. Microsoft has not responded publicly to requests for comment from several outlets, including TechCrunch and Benzinga.
A battlecard is ordinary. The target is not.
Every large enterprise software vendor keeps a competitive deck with a rival's logo on it and a column of red marks running down the side. Sales enablement is a discipline, and knocking the competition is most of what it does.
What makes this session worth reading closely is the identity of the competition. OpenAI models have been the engine underneath Microsoft's AI product line since the Copilot brand existed. Anthropic's Claude models have been shipping inside Microsoft 365 Copilot and GitHub Copilot since Microsoft struck an expanded distribution arrangement in November 2025. Microsoft is a shareholder in both companies. It is coaching its sellers to talk down two of its own suppliers, one of which it owns roughly a quarter of.
That is not a contradiction so much as a disclosure of priorities. To a platform owner, model vendors are suppliers, and suppliers are replaceable. The deployment surface is the asset.
The cost argument has a number attached
Satya Nadella supplied the financial spine of the pitch. Per Bloomberg's report, the Microsoft CEO told the sales organization that the ability to track AI spending and shift work onto cheaper models will be the dominant customer conversation of the coming year.
He anchored it to a customer. Nadella pointed to a Unilever claims-processing system built on Microsoft's platform that swapped out one of the most advanced frontier models for a cheaper Microsoft alternative. A Microsoft spokesperson put the projected savings at roughly $300 million. Unilever did not comment on the figure, and no methodology behind it has been published.
The number is doing a specific job. It converts an abstract argument about model economics into a line item a CFO can picture, and it reframes the buying question away from which model tops the leaderboard toward how much of the leaderboard a company is paying for without needing it.
Nadella had already laid the intellectual groundwork in public, three days before the sales meeting. In a post on Sunday, July 12, he argued that enterprises buying AI face what he called a reverse information paradox: to get good output from a model, a company has to feed it the workflows, the corrections, the internal evaluations and the judgment that constitute its actual advantage. "You essentially pay for intelligence twice," he wrote, once in cash and once in proprietary knowledge handed to a vendor that keeps the learning.
Read alongside Tuesday's transcript, the essay looks less like philosophy and more like advance work. It gives the sales force a principled version of the same argument Andreou was making with a comparison slide.
The pitch trails a product decision Microsoft already made
Nobody commissions a competitive analysis to find out which model is better. You commission one after you know which model you are shipping.
On July 7, Bloomberg reported that Microsoft had started replacing OpenAI and Anthropic models with its own MAI systems inside Excel and Outlook, with tens of thousands of prompts each week now completed by internally built models, according to a person familiar with the work. Those apps had previously leaned more heavily on the partner labs. MAI still handles a small slice of total volume. The direction is the point.
The infrastructure for that shift landed at Build in June, where Microsoft AI chief Mustafa Suleyman introduced seven new MAI models spanning reasoning, coding, image generation, voice and transcription. Microsoft claimed one of them matched a prior-generation Anthropic coding model in human evaluations at lower cost. Published benchmarks released alongside the launch told a less flattering story, showing the new reasoning model trailing frontier systems from both labs by a wide margin.
Suleyman has been blunt about the objective. Microsoft pays a great deal of money to Anthropic, he said at the time, and the goal is to shrink that bill and eventually erase it.
The plumbing here matters more than the rhetoric. High-volume, low-complexity work (summarizing a thread, drafting a reply, formatting a sheet) is where inference bills compound, and it is also where a cheaper model is hard to distinguish from an expensive one. Route that traffic to MAI, keep frontier models for the hard cases, and the margin problem improves without the user interface changing at all. The customer sees the same Copilot.
The awkward part: Microsoft's own engineers voted the other way
In December 2025, Microsoft handed Claude Code licenses to thousands of employees across its Experiences and Devices organization, the group responsible for Windows, Microsoft 365, Outlook, Teams and Surface. Engineers, designers, program managers and staff who had never shipped a line of code all got seats.
It went well. Adoption climbed fast, and by The Verge's account the tool became popular enough to embarrass GitHub Copilot CLI, the in-house product it was being benchmarked against.
Microsoft wound it down anyway. Most internal Claude Code licenses were cut with a June 30 deadline, which is also the last day of Microsoft's fiscal year. In an internal memo, Executive Vice President Rajesh Jha framed the decision as convergence after a learning period, noting that building on Copilot CLI lets Microsoft shape the tool directly with GitHub around its own repositories and security requirements. People familiar with the decision told reporters that cost was a factor, and the calendar supports them.
Anthropic's models were not banned. They remain available through Copilot CLI and Microsoft Foundry, and inside Microsoft 365. What went away was Anthropic's interface, and with it the direct relationship between Anthropic and thousands of Microsoft developers.
That sequence is the sales pitch in miniature. Microsoft conceded the tool and kept the switchboard.
Why the timing is not accidental
Microsoft goes into this fiscal year with an investor problem that no amount of sales enablement fixes on its own.
The company has guided to roughly $190 billion in capital expenditure for calendar 2026, up about 61% year over year, including some $25 billion attributable to higher component prices as memory costs spike. Quarterly capex hit $30.88 billion in the March quarter, an 84% jump. The AI business is growing, with an annual revenue run rate above $37 billion, but Microsoft Cloud gross margin is compressing, and management has guided it lower again.
The stock has taken the difference. MSFT has fallen more than 20% over the past twelve months, closing Wednesday at $395.63 for a market capitalization near $2.94 trillion. Citi and Mizuho both trimmed price targets this week while keeping buy-side ratings, with both firms flagging capex digestion as the near-term overhang. Fiscal Q4 results land July 29.
Underneath the capex worry sits a sharper one that Wall Street has been calling the SaaS-pocalypse: the possibility that the labs themselves move up the stack and start selling the enterprise software that Microsoft, Salesforce, ServiceNow and Thomson Reuters have been selling for two decades. In that scenario, Copilot is a distribution wrapper with a markup on a competitor's API, and the moat is a habit.
A pitch built around owning the whole system is a direct answer to that fear, aimed at investors as much as at buyers.
What the Copilot versus Claude comparison actually establishes
Andreou's claim deserves to be read for what it is: an internal characterization, produced by the vendor, presented to its own sellers, on a surface that vendor builds and controls. No third party has published a comparison of the two systems inside Office apps. The latency and accuracy figures behind the slide are not public.
The security point is the most revealing of the three claims, because it is not a statement about the model at all. Security integration is engineering work, and the party that decides whether it gets done for a partner model inside Word is Microsoft.
The commercial context is unusual enough to state plainly. Microsoft committed up to $5 billion to Anthropic in November 2025 as part of the same arrangement in which Anthropic committed to purchase $30 billion of Azure compute, alongside Nvidia's pledge of up to $10 billion. Microsoft holds roughly 27% of OpenAI's public benefit corporation on an as-converted diluted basis against $13 billion in funding commitments. Every dollar Copilot spends with either lab is partly a dollar Microsoft gets back, and every dollar it stops spending is margin it keeps outright.
The April rewrite of the OpenAI agreement removed the last structural reason for restraint. Microsoft's license to OpenAI's intellectual property now runs to 2032 on a non-exclusive basis, OpenAI can serve customers on any cloud, the AGI clause that could have voided the arrangement is gone, and OpenAI's revenue share payments to Microsoft now run to 2030 under a total cap. Partners with no exclusivity owed to each other are just companies.
The market data explains why Anthropic in particular ended up on the slide. Menlo Ventures' enterprise survey put Anthropic at 40% of enterprise LLM spend against OpenAI's 27% and Google's 21%, up from 12% in 2023, and estimated Claude's share of enterprise coding workloads at 54% against 21%. Sellers do not get coached against companies that are losing.
Comments
Join the discussion and share your perspective.