Microsoft expands in-house MAI AI models across Microsoft 365 to boost efficiency

Microsoft is scaling the deployment of its proprietary MAI AI models across key productivity applications, combining them with frontier AI models to optimise enterprise workloads, lower inference costs and enhance task-specific performance through product-focused training.

Microsoft is accelerating the adoption of its proprietary MAI family of artificial intelligence (AI) models across major products, including GitHub Copilot, Excel, Outlook and other Microsoft 365 applications, as it looks to improve efficiency, lower AI operating costs and deliver stronger performance for enterprise users.

The strategy was outlined by Microsoft Chief Executive Officer Satya Nadella in a blog post published on July 23, where he introduced what he called the company's "Frontier Diffusion & Control" approach. The initiative focuses on combining frontier AI models with Microsoft's in-house AI capabilities to optimise performance for different enterprise scenarios rather than relying on a single large AI model for every task.

According to Microsoft, the latest generation of MAI models has been trained using reinforcement learning environments (RLEs) tailored to specific products. These environments enable the models to learn from real customer workflows instead of depending solely on benchmark datasets, allowing them to deliver improved results in applications such as GitHub Copilot and Excel.

Product-specific AI takes centre stage

Explaining the company's strategy, Nadella said AI systems should be optimised around business outcomes rather than the capabilities of an individual model.

"In a world where software has real marginal cost for the first time, how do we ensure frontier benefits are diffused across the entire ecosystem?" Nadella wrote. "The key is to optimise the cost-to-outcome frontier in real world context. In practical terms, that means using the right model for each task, and optimising the context, skills, tools, and agent harness around it."

Microsoft clarified that its MAI models are not intended to replace frontier AI systems developed by partners such as OpenAI and Anthropic. Instead, the company plans to use multiple models within a unified orchestration framework, assigning workloads based on complexity and performance requirements.

"In our products, frontier models from OpenAI and Anthropic are part of the orchestration system alongside MAI," Nadella said. "But the model is only one part of the hill-climbing system. Harness, memory, context, tools, skills, user interactions, etc. all shape the evals and performance of these agentic systems."

Lower AI costs with continuous optimisation

Microsoft said the new framework enables routine AI tasks to be handled by lightweight in-house models, reserving computationally intensive frontier models for more demanding applications. This approach is expected to reduce inference costs while maintaining, or even improving, output quality.

The company described its continuous optimisation process as "hill climbing," where AI systems evolve through ongoing product-specific evaluations and reinforcement learning driven by customer interactions. By keeping elements such as memory, context and product-specific tools outside the models themselves, Microsoft says it can replace or upgrade models without disrupting user experiences.

"The other key criteria to ensure that you are in control, is your evals should continue to hill climb even when any given model has been removed," Nadella said. "We build RLEs where models learn inside the product system and are rewarded for completing the tasks customers actually care about."

Microsoft said early deployments across GitHub Copilot, Excel and Outlook have delivered encouraging results, with MAI models outperforming general-purpose AI models in several specialised use cases while consuming significantly fewer tokens. The company also plans to extend the same framework through Microsoft Foundry, enabling enterprises to build AI agents using their own data, workflows and evaluation systems.