Introducing Microsoft 365 Copilot Tuning

Microsoft

May 19, 2025

AI tools powered by out-of-the box LLMs and retrieval augmented generation (RAG) may not always understand your business in terms of specific processes, terminology, and style. They can provide results that are based on publicly available datasets and lack contextual awareness. Many AI solutions providers have turned to fine-tuning to address this. While fine-tuning offers a solution, making it truly enterprise-ready is usually time consuming, complex, and intensive.

Microsoft 365 Copilot Tuning offers customers a new way to unlock the value of fine-tuning for organizations without the cost and complexity of other solutions. Now, makers can use the low-code tooling in Microsoft Copilot Studio to take advantage of highly automated fine-tuning “recipes” that can use your enterprise data to train models to assist with domain-specific tasks. Once training is finished, agents, built in Agent Builder with just a few clicks, can then tap into these fine-tuned, task-specific models and easily integrate them into the Microsoft 365 apps you work with every day, like Chat, Teams, Word, SharePoint and Agreements Solution.

Our turnkey approach

The approach we’re taking to fine-tuning streamlines many complex steps typically associated with task-specific fine-tuning. One of the key innovations that powers Copilot Tuning is a breakthrough data collection method that tackles a familiar challenge - the need for clean, high-quality domain specific data, identified by subject matter experts (SMEs). Our approach to data preparation, described in this peer-reviewed ICML-accepted paper, enables SMEs to clean an entire dataset while labeling less than 10% of its contents.

This innovation can reduce the need for code development and data science work. The result is more democratized access to fine-tuning; makers with domain expertise in a business task or process can now tap into its power. For example:

Data selection is simple. With just a few clicks makers can elect to use files stored in SharePoint or soon, Microsoft Graph-connected data
Data ingestion and processing are eyes-off, meaning at no point during either of these processes will anyone have access to the contents of the files used to fine-tune the model.
Access Lists (ACLs) are built-in. The fine-tuned model preserves access lists associated with the training data and automatically checks for alignment between who can access the data and who can use the model – to minimize the possibility of data oversharing.
Pre-defined tasks are included in our approach and require no manual coding by the maker. At this time, we have three task types: document generation, document summarization, and expert Q&A generation.

We’ll continue to evolve the technology powering Copilot Tuning. For example, in the coming months we’ll make it possible to add more and more data sources and continue to expand our library of tasks.

Use cases & pilot examples:

We’re introducing Copilot Tuning with three pre-built “recipes” that focus on specific tasks including Expert Q&A, Document Generation and Document Summarization. There are broad practical applications for each.

Expert Q&A: Build a Customer Service Knowledge Agent trained on technical support materials from your SharePoint repositories that can instantly provide rich, detailed answers to your support reps’ questions, even when internal jargon and organizational acronyms are used.
Document Generation: Create a Proposal Writer agent. Train Copilot using your company’s archive of successful RFPs to assemble complex first drafts written in your company’s approved format and tone.
Summarization and Analysis: Build a Project Analyst agent that can summarize large volumes of project documentation to produce executive or customer status summaries that extract insights specific to key metrics.

We’ve been piloting Copilot Tuning with several customers, including the global consultancy Ernst & Young, Canadian law firm McCarthy Tétrault, and US agricultural cooperative Land O'Lakes. It’s exciting to see how they’re using Copilot Tuning to drive AI innovation that builds on the rich expertise developed over years in each of their businesses. Here’s what they have to say:

We are thrilled to collaborate with Microsoft on the Copilot Tuning project. By integrating a Tax-domain fine-tuned LLM with our enterprise knowledge and the expertise of our Tax Advisors through M365, we are bringing an enhanced tax service to the market. This synergy improves service quality and significantly advances tax and legal research with relevant knowledge and intelligence readily available in M365 where people are already working.

Marna Ricker
Global Vice Chair – Tax
Ernst & Young

We are excited about what can be created with our talent and data combined with the technology and platform securely provided by Microsoft. Our collaboration can already produce high-quality first drafts of agreements, factums, and other complex legal documents, underpinned by our firm’s own precedents and knowledge. The future of legal will be defined by partnerships like this.

Matt D. Peters
Partner, National Transformation Lead
McCarthy Tétrault

Within just weeks of initiation, Land O’Lakes and Microsoft delivered a finetuned AI model—proof that when deep ag expertise meets worldclass AI innovation, speed and impact follow. This milestone showcases the power and agility of our alliance. Together, we’re reimagining what’s possible in our ag businesses and with our customers and member-owners.

Jeremy Lembeck
Director, Data & Analytics
Land O’Lakes, Inc.

About the Copilot Tuning Early Access Program

Today we’re introducing Copilot Tuning through an Early Access Program (EAP). Customers with more than 5,000 Microsoft 365 Copilot seats may participate. It’s important to note that EAP candidates will undergo additional screening to ensure they have scenarios that align with the three pre-configured tasks currently (document summarization, Q&A generation and document generation). We may also offer support for initial implementation work to maximize the benefits of this program.

To learn more about this EAP, including engagement terms and billing model, please visit our FAQ page.
If your organization is interested in the EAP but does not meet the minimum seat criteria, please work with your Microsoft account manager to find out more details about your organization’s potential participation.

We anticipate expansion of this EAP in the coming months and will provide updates to our developer and maker communities along the way.