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

The effort is said to focus on the tech giant’s Microsoft 365 Copilot product. It’s an artificial intelligence assistant that ships with the eponymous productivity suite. Microsoft is reportedly working to integrate custom and open-source artificial intelligence models into the assistant, which is currently powered by technology from OpenAI.

Microsoft 365 Copilot made its debut last March. It automates common tasks in the productivity suite’s flagship applications. The version of Copilot embedded in Word, for example, can summarize lengthy documents and generate new ones, while the Excel version suggests ways to visualize data.

The assistant is also built into some of the tools that administrators use to manage their companies’ Microsoft 365 deployments. One of those tools is Purview, which helps prevent workers from using business data in an authorized manner. The embedded version of Copilot provides pointers on how to use the tool and summarizes data leak alerts.

According to today’s report, Microsoft hopes that incorporating new AI models into Copilot could reduce the cost of operating the assistant. It’s believed the company may use any savings it achieves to lower prices for customers. Improving Copilot’s response times is another priority. 

One of the internally developed models that Microsoft could reportedly build into Copilot is Phi-4. Introduced earlier this month, it features 14 billion parameters, a fraction of the number in frontier LLMs. That means the model cost significantly less to run.

In a Microsoft evaluation, Phi-4 outperformed an LLM with five times as many parameters on a benchmark that compared AI models’ math prowess.

Today’s report didn’t specify which open-source models could find their way into Copilot. However, it’s probable that Meta Platforms Inc.’s Llama LLMs are among the algorithms under consideration. The LLM series is one of the most advanced in the open-source ecosystem and outperforms some proprietary frontier models across certain tasks.

The newest addition to the lineup, Llama 3.3, rolled out three weeks ago. It provides comparable output quality as a previous-generation Llama model with 405 billion parameters, but uses only a fraction of the hardware. It also mostly outperformed OpenAI’s GPT-4o across the benchmarks that Meta evaluated.

The productivity software market is not the only area where Microsoft has started reducing its reliance on OpenAI. Earlier this year, its GitHub unit’s GitHub Copilot coding assistant received support for LLMs from Google LLC and Anthropic PBC. However, Microsoft told Reuters that “OpenAI continues as the company’s partner on frontier models.”

“We incorporate various models from OpenAI and Microsoft depending on the product and experience,” a Microsoft spokesperson added. 

Image: Microsoft