AI assistants like Microsoft 365 Copilot are arriving in federal offices fast, and many agencies can now try them at little or no cost. The tools are powerful. But turned on without preparation, they will find and surface every document your organization has ever overshared.

---

The Story – The Assistants Are Here

In 2025, the federal government moved from experimenting with generative AI to buying it. OMB’s April 2025 memo M-25-21 directed agencies to accelerate their use of AI, name Chief AI Officers and publish AI strategies. A few months later, GSA’s OneGov agreement with Microsoft made Microsoft 365 Copilot available to eligible federal customers at no cost for the first year. Similar deals with other AI providers followed. For many agencies, the question is no longer whether to deploy an AI assistant, but how quickly.

That speed is the risk. An assistant like Copilot works by reading what the user is allowed to read: email, Teams chats, SharePoint sites, OneDrive files. It doesn’t create new access. It makes existing access easy to use. A user who asks “summarize everything about the reorganization” may get a tidy summary drawn from a folder that was shared with “everyone” five years ago and forgotten.

Fix the Permissions First

Most organizations have years of oversharing built up in Microsoft 365: sites open to the whole tenant, sharing links that never expire, departed employees’ files still accessible, and sensitive spreadsheets sitting in team folders. Before AI, that content was hidden by obscurity. With an AI assistant, obscurity is gone.

A responsible rollout starts with cleanup. Review who can see what in SharePoint, Teams and OneDrive. Remove “everyone” and organization-wide access where it isn’t needed. Apply sensitivity labels to protect controlled and sensitive information, and use data loss prevention policies to keep it from leaving where it belongs. Records management matters too: prompts and AI-generated content can be federal records, so retention policies need to cover them.

Then roll out in stages. Start with a pilot group, watch what the assistant surfaces, fix what it finds, and expand from there. Training is the other half. Users need to know what the assistant is good at, that its answers must be checked, and what information should never go into a prompt. A one-page acceptable use policy, written in plain language, does more than a long one nobody reads.

What We Recommend

At InfoStructures, we provide IT program management, Microsoft 365 administration and training, and technical writing for our federal customers. Our approach to AI assistants is the same as any major rollout: assess first, fix the foundation, pilot, train, and measure. That means a permissions and oversharing review, sensitivity labels and retention policies in place, a clear usage policy, hands-on training for each wave of users, and a feedback loop so problems found in the pilot are fixed before everyone has access.

AI assistants can save real time on email, meetings and document work. The organizations that get that value without a data incident are the ones that do the unglamorous preparation first. Do you know what your AI assistant would find if you turned it on tomorrow?