Microsoft AI: The Internal Emancipation That Divides

The context, in plain terms

Under the direction of Mustafa Suleyman, head of its AI division, Microsoft unveiled its family of seven in-house AI models, named MAI (Microsoft AI), on June 2, 2026, at the Build 2026 conference in San Francisco. Starting in early June 2026, the company began gradually replacing Anthropic's Claude and OpenAI's GPT models with its own MAI models for certain specific tasks within Excel and Outlook. The decision is currently 'in progress' of deployment.

This strategic transition aims to reduce an estimated annual expenditure of $500 million for the use of Anthropic's models. According to a source cited by Bloomberg, tens of thousands of AI queries per week in Excel and Outlook are now processed by MAI models. Microsoft states that an MAI model optimized for Excel achieves GPT-5.4 performance while consuming ten times fewer resources. Mustafa Suleyman publicly confirmed in June 2026 the goal of entirely eliminating dependence on Anthropic models, and MAI models are already integrated into GitHub Copilot, with imminent deployment for the Teams transcription tool.

However, several uncertainties remain around this decision. No full implementation date for 100% of queries in Outlook and Excel has been communicated. Microsoft has also not announced any pricing changes related to MAI deployment in the French market, nor a specific expanded deployment date. Collaboration with OpenAI and Anthropic continues for certain services like Azure and Copilot, suggesting a gradual rather than abrupt transition, without visible disruption for the end user. Finally, the performance of the MAI model for Excel, although affirmed by Microsoft as equivalent to GPT-5.4, has not yet been validated by independent tests published in the consulted business press sources.

The Range of Reactions: Shared Support, but Marked Hostility

The Kapari test bench tested Microsoft's decision with a simulated panel of 56 voices, revealing a range of shared reactions. Within this panel, 23 voices express support for the decision, recognizing the strategic relevance of such technological emancipation. However, 26 voices declare hostility, showing reservations or a clear rejection of this transition.

Between these two poles, 7 voices are in doubt, signaling unresolved questions or a need for further clarification before taking a stance. This distribution of reactions, with slightly majority hostility and a significant block of doubt, indicates that the decision, while offering opportunities, faces significant resistance that requires particular attention before any progress.

Fault Lines: Unexpected Support and More Reserved Internal Voices

The analysis of reactions reveals interesting fault lines within the panel. A notable signal is that of a voice which, although belonging to a group generally inclined to distrust the decision, declares itself favorable. This profile, a GitHub product director, is a valuable lever for action: carefully listening to this unexpected voice would allow understanding the concrete arguments that convinced them and identifying unexpected benefits to highlight.

Furthermore, another signal highlights a divergence between internal and external perceptions: the test bench observes that 'the house is colder than the street.' This means that internal voices react less warmly to the decision than external voices. This divergence underscores the importance of formulating distinct messages, specifically addressing the concerns and expectations of internal collaborators before communicating to the external public, to ensure better internal support.

The Dominant Friction and a Tipping Point

The dominant friction identified by the Kapari engine is doubt about execution. This is the main obstacle to defuse before further exposing the decision. An important signal in this regard concerns employee representatives: although this group represents only 12% of the simulated panel and does not have a predominant numerical weight, their voices are identified as 'noisy.' This 'weighting' signal indicates that their doubts about the feasibility and implications of execution can amplify the dominant friction and destabilize general perception. It is therefore crucial to gather their reactions and address their concerns to avoid a blocking point.

The stability of the verdict, calculated as 'Adjust' over three independent passes of the engine, reinforces this observation. The persistence of this verdict, despite repeated simulations, indicates that the friction points are structural and not circumstantial. This confirms that doubt about execution is indeed the dominant obstacle to address as a priority and robustly, and that the decision cannot move forward without clarification and strengthening of the implementation strategy.

Verdict: Adjust the Decision to Address Doubts About Execution

The verdict to Adjust the decision, accompanied by a High reception risk, directly stems from the shared range of reactions and the dominant friction identified. The simulated reactions are split between support and hostility, with a clear blocking point from employee representatives regarding execution. Adjusting does not mean questioning the strategic direction, but rather refining the method and communication to maximize the chances of success.

To move forward, the suggested path includes several concrete actions based on the identified signals. It is recommended to listen carefully to unexpected voices, such as that of the GitHub product director, to understand concrete levers of support and integrate them into overall communication. Furthermore, it is imperative to design distinct messages for internal and external audiences, as reactions from 'the house' are less warm than those from 'the street,' requiring a specific approach for collaborators. Finally, the decision must be refined by defusing doubt about execution, particularly by gathering reactions from employee representatives, even if their numerical weight is low, to prevent their concerns from becoming a major and noisy blocking point. The stability of the verdict over several passes confirms the relevance of these adjustments.

Questions about this case

Related cases