Microsoft routes some Word and Excel prompts to its in-house AI models (MAI)
A Bloomberg sourced shift inside Office reframes enterprise AI procurement and pressures OpenAI's and Anthropic's enterprise revenue mix.
A Bloomberg sourced shift inside Office reframes enterprise AI procurement and pressures OpenAI's and Anthropic's enterprise revenue mix.
Microsoft has begun routing a portion of Word and Excel prompts to its in-house MAI models instead of relying exclusively on OpenAI and Anthropic. The shift, first reported by Bloomberg on July 7 and corroborated the same day by Hindustan Times and PYMNTS, is the clearest sign yet that the world's largest enterprise AI buyers are pulling back from third-party dependencies.
For years, Microsoft advertised that large parts of Office 365 were powered by models from OpenAI and Anthropic, two of the leading third-party AI labs. The new behavior reverses that posture on the prompts it can serve cheaply with its own stack. PYMNTS described the change as "thousands of Office prompts" now resolving to in-house AI. Microsoft declined to share the prompt-share percentage, dollar savings, or quality comparison against OpenAI and Anthropic when reached for comment, so any specific number on this story should be read as a Bloomberg-sourced estimate rather than a company-confirmed figure.
At its Build conference in June 2026, Microsoft launched seven new MAI models, including an agentic coder and a text-to-image generator. Ten days before the Bloomberg report, the company shipped MAI-Code-1-Flash to Copilot Business and Copilot Enterprise through GitHub.
Microsoft 365 Copilot's Agent Mode and Office Agent launch in September 2025 marked the first time the company tied Word, Excel, and PowerPoint to a Copilot surface that could plausibly be served by Microsoft-owned models for at least routine tasks. Routing the underlying inference to MAI is the natural completion of that arc.
A Bank of America note circulated via Barron's warned that AI capex from Meta, Alphabet, and Amazon could spook public markets if it does not convert into measurable productivity. Inside the same reporting cycle, Amazon and Accenture have both pointed to cheaper AI stacks or in-house model work as a way to control inference bills, a posture some practitioners call "tokenmaxxing," or the practice of squeezing more output from every dollar of model spend.
Microsoft's posture inside Office is the highest-stakes test of that discipline. If the company's MAI models can serve the bulk of Word and Excel prompts at acceptable quality, every other Fortune 500 IT shop now has a reference architecture for buying less AI from OpenAI and Anthropic than they planned to a year ago. The pressure point lands hardest on OpenAI's enterprise revenue mix: many of those customers are still routing through Azure OpenAI Service, where Microsoft has historically booked the inference margin regardless of which model answered.
Three things will resolve whether the in-house route is a margin story or a capability one. Microsoft has not disclosed whether MAI is handling the easy prompts while OpenAI and Anthropic still serve the hardest ones, or whether quality parity is being measured against specific benchmarks rather than user satisfaction. The first external benchmark of MAI on Office workloads, or the first public dollar figure on Copilot inference cost, will decide.
The next concrete trigger is the rollout of MAI-Code-1-Flash to Copilot consumer tenants, which GitHub has signalled but not yet dated. A clean shipping milestone without quality complaints from enterprise developers would move the procurement calculus across the rest of the Fortune 500 with it.