Not with consumer plans, and usually not with standard commercial business plans. An AI service can hold CUI only if it meets the same rules as anywhere else you keep CUI, and only if it fits inside the boundary you’ve documented. As of October 2026 there are real options that can: Microsoft 365 Copilot in GCC High, Azure OpenAI in Azure Government, and models you run on your own hardware.
Most of the companies asking this already have people pasting work into ChatGPT. The question isn’t whether to use AI; it’s where each kind of information is allowed to go.
Why the usual answer is no
AI tools aren’t a special category. A service that stores or processes CUI is a cloud service like any other, so DFARS 252.204-7012 applies: it has to meet security requirements equivalent to the FedRAMP Moderate baseline, and the provider has to support the clause’s obligations, including reporting cyber incidents to you so you can report within 72 hours, preserving evidence and allowing forensic access.
Consumer plans (free or personal ChatGPT, consumer Copilot) give you none of that. Commercial business plans give you better data terms, but usually not the contractual commitments 7012 requires, and they run outside the government cloud environments many contractors already use for CUI. Export-controlled technical data adds its own limits on where data is stored and who can access it.
What “FedRAMP authorized” doesn’t settle
Authorizations are a good sign, but they answer a narrower question than it seems. OpenAI’s ChatGPT Enterprise and API platform received FedRAMP Moderate authorization in April 2026, for use by federal agencies. That tells you about the provider’s security; it doesn’t, by itself, make a commercial contract between your company and the provider suitable for CUI.
For any AI service, get these answers in writing before CUI goes near it:
- Which offering and environment are covered, and is it the one you’re actually buying?
- Does your contract include the DFARS 252.204-7012 commitments: incident reporting, evidence preservation and forensic access?
- Where are prompts, files and outputs processed and stored, for how long, and are they used for training?
- If you handle export-controlled data, does the service meet those requirements too?
Options that can work
Microsoft 365 Copilot in GCC High
Available since December 2025 for organizations already on a GCC High tenant, with an add-on license. Prompts, retrieved content and responses stay inside the Microsoft 365 GCC High service boundary. Some features still trail the commercial version, and Copilot can reach anything a user can open, so permissions and labels have to be right first.
Azure OpenAI in Azure Government
Covered by Azure Government’s FedRAMP High authorization and in scope for DoD Impact Levels 4 and 5, with extra configuration for IL5. Best for building your own tools, such as search over engineering documents or drafting assistants, rather than a ready-made chat app.
Private models on your own hardware
Open-weight models running on a GPU workstation or server inside your enclave. Nothing leaves your boundary, and the system is documented like any other CUI asset. The trade-offs are hardware cost, upkeep and models that are capable but not the very largest.
Government versions of other AI services
New government offerings keep arriving. Each needs the same questions answered against your contract and your boundary; we check current terms rather than relying on last quarter’s answer.
Before anyone turns it on
- Fix oversharing. AI assistants surface whatever a user’s account can open. Clean up permissions on shared drives and sites first.
- Label CUI. Sensitivity labels and data loss prevention rules let the tools, and your people, tell CUI from everything else.
- Write the rule down. A short policy that says which kinds of information may go into which tool, in plain language.
- Company identities only. Single sign-on and multifactor authentication on every AI tool, and no personal accounts for company work.
- Update your documentation. An AI service or model server that touches CUI is in your assessment scope. Add it to your boundary diagram and System Security Plan.
What to do about the AI use already happening
Banning AI outright rarely works; people use it anyway, on personal accounts you can’t see. A better short-term answer is a sanctioned business-grade tool for non-CUI work, a clear rule that CUI never goes into it, and training on what CUI looks like. Then decide, deliberately, whether you need an option inside your boundary for CUI work.
If you’re still getting the boundary itself under control, start there. Our article on getting caught up on NIST SP 800-171 covers scoping, and our AI integration page shows how we approach AI for regulated data.
Questions we hear
Is Copilot Chat in our commercial Microsoft 365 tenant okay for CUI?
Generally no. CUI belongs in the environment you’ve documented for it; for most contractors using Microsoft 365, that means GCC High. Copilot in GCC High is the version designed for that boundary.
If a vendor is FedRAMP authorized, are we covered?
Not by itself. You still need the DFARS 7012 commitments in your contract, an environment that matches what was authorized, and the service documented in your boundary and System Security Plan.
Are local models good enough to be useful?
For many everyday tasks, such as summarizing documents, drafting, searching internal files and helping with code, yes, on a capable GPU workstation or server. They won’t match the very largest cloud models on every task, which is why the choice depends on the work.
Does adding AI change our CMMC scope?
If it processes CUI, yes. It becomes part of the system being assessed, so plan it into the boundary from the start rather than discovering it during an assessment.
Sources: DFARS 252.204-7012; Microsoft: Microsoft 365 Copilot in GCC High; Microsoft: Azure OpenAI in the FedRAMP High authorization for Azure Government; Microsoft: Azure OpenAI at DoD IL4 and IL5; OpenAI: FedRAMP Moderate.