Microsoft Plans to Integrate Non-OpenAI Models in 365 Copilot to Reduce Costs

According to Reuters, Microsoft is working to reduce its reliance on OpenAI in the productivity software market.

This effort is primarily focused on Microsoft's Microsoft 365 Copilot product, an AI assistant that was introduced alongside the productivity suite. Currently, Copilot is powered by OpenAI's technology, but Microsoft plans to integrate custom and open-source AI models into the assistant.

Microsoft 365 Copilot was first unveiled in March of last year and can automate common tasks in the flagship applications of the productivity suite. For example, the Copilot embedded in Word can summarize long documents and generate new ones, while the Excel version can suggest data visualization options.

Additionally, the assistant is integrated into some tools used by administrators to manage their company's Microsoft 365 deployments, including Purview, which helps prevent unauthorized use of business data. The built-in Copilot version provides guidance on how to use the tool and summarizes data breach alerts.

It is reported that Microsoft aims to lower operational costs by integrating new AI models into Copilot, potentially using the savings to reduce prices for customers. Improving Copilot's response time is another key objective.

One of the internally developed models that Microsoft may integrate into Copilot is Phi-4, which was launched earlier this month. With 14 billion parameters, it has significantly fewer than leading large language models (LLMs), resulting in lower operational costs. In a Microsoft evaluation, Phi-4 outperformed LLMs with five times as many parameters in a benchmark test comparing the mathematical abilities of AI models.

The report did not specify which open-source models might be integrated into Copilot. However, Meta Platforms Inc.'s Llama series of LLMs could be one of the algorithms under consideration. The Llama series is among the most advanced in the open-source ecosystem, outperforming some proprietary state-of-the-art models in certain tasks.

The latest addition to the Llama series, Llama 3.3, was released three weeks ago. It matches the output quality of the previous 40.5-billion-parameter Llama model but uses significantly less hardware. In Meta's benchmark tests, Llama 3.3 also generally outperformed OpenAI's GPT-4.

The productivity software market is not the only area where Microsoft is reducing its dependence on OpenAI. Earlier this year, its GitHub division's GitHub Copilot coding assistant began supporting large language models from Google LLC and Anthropic PBC. However, Microsoft told Reuters that "OpenAI remains a partner for us in cutting-edge models."

A Microsoft spokesperson added, "We combine various models from OpenAI and Microsoft, depending on the product and experience."