Microsoft Copilot offers substantial privacy protections, but the answer depends on the account, app, and features you use. For in-house counsel comparing Copilot with ChatGPT, start with equivalent plans: organizational Copilot and ChatGPT Business or Enterprise both exclude business content from foundation-model training by default.
That still leaves the questions your legal team must answer before approving a workflow: who can access the information, what remains stored, which connected services receive it, and whether the contract permits that use.
GC AI is an enterprise legal AI platform built for in-house teams. We help lawyers review agreements against company standards, inspect the sources behind an answer, and turn findings into advice for the business. 2,200+ legal teams use GC AI, including TIME, SKIMS, News Corp, Riot Games, and Liquid Death.
Our co-founder and CEO, Cecilia Ziniti, was general counsel at Anki, Bloomtech, and Replit. That in-house experience informs the work GC AI supports: giving the business a usable answer while keeping counsel responsible for the legal judgment.
Copilot vs. ChatGPT Privacy Starts With the Account
A paid personal subscription is different from a company-managed workspace. The product name alone can also mislead: Microsoft now calls its organizational Microsoft 365 Copilot product Microsoft Copilot, while older names remain in some licenses and interfaces.
Personal Copilot and Personal ChatGPT
Microsoft changed its consumer Copilot documentation for the updated app released on August 18, 2026. Its current activity-history guidance says prompts, responses, and file contents in that app are not used to train foundation models. The older Copilot app follows separate guidance; the former training opt-out and 18-month history statement should not be applied to every current Copilot experience.
Copilot inside Microsoft 365 apps for home also has a no-foundation-model-training commitment. Those personal-account experiences still operate under consumer terms, rather than your organization's enterprise agreement.
For personal ChatGPT plans, including Free, Go, Plus, and Pro, eligible conversations may be used to improve models unless you turn off that use. OpenAI's data controls separate training, Memory, and deletion. Turning off model improvement leaves chat history in place.
For either product, a no-training setting or commitment does not establish that the account meets your company's requirements for confidential work. Approve the actual account and workflow before an employee uploads a customer contract.
Organizational Copilot, ChatGPT Business, and ChatGPT Enterprise
Compare the organizational offerings and the controls your company can actually configure. A data protection addendum (DPA) governs covered personal-data processing; Microsoft Graph connects organizational content such as emails, chats, and documents.
Offering | Training and Contractual Baseline | What Counsel Should Check |
|---|---|---|
Organizational Copilot Chat, signed in with Microsoft Entra | Enterprise data protection covers prompts and responses; no foundation-model training. The DPA and Product Terms apply. | Available retention and audit controls, uploaded files, agents, and web search. Enterprise data protection does not require the full Copilot license. |
Licensed organizational Microsoft Copilot, formerly Microsoft 365 Copilot | No foundation-model training on prompts, responses, or Microsoft Graph data. | Microsoft 365 permissions, Purview configuration, enabled models, and the additional data reached by each experience. |
ChatGPT Business | No training on business data by default. Covered business services have confidentiality terms; a DPA is available. | Workspace retention, administrator access, sharing, and connected apps. Do not assume every Enterprise control is included. |
ChatGPT Enterprise | No training on business data by default. Review the applicable business agreement and DPA. | Administrator-controlled retention, compliance access, role settings, and enabled features for the purchased deployment. |
Sources: Microsoft enterprise data protection, Copilot Chat protections, and OpenAI business privacy commitments, checked October 2, 2026.
A work email address is not the approval. Record the product, signed-in account, applicable agreement, allowed data, and enabled features. Keep API deployments separate: an API's retention arrangement does not automatically apply to a conversation in ChatGPT.
No Training Does Not Mean No Stored Records
Organizational Copilot stores prompts and responses as activity history. Microsoft Purview retention policies govern retained interaction records; deleting a visible chat is not proof that every compliance copy has disappeared.
For ChatGPT, saved conversations and uploaded files can have different lifecycles. OpenAI's retention guide explains that deleting a conversation does not delete a file still saved in Library. Temporary Chats may be retained for up to 30 days for safety; saving one converts it to a regular chat.
Ask IT to demonstrate retention and deletion for the exact workspace and the records your team creates. Check the chat, uploaded source file, exported assessment, shared copy, and any preservation requirement.
Relevant AI records may be subject to preservation and discovery, with privilege and other protections assessed separately. Counsel should include AI history in the company's preservation process and coordinate deletion when a preservation duty applies.
Who Can Read the Source and Who Can See the Answer?
Microsoft's privacy documentation says Copilot surfaces organizational content the user has permission to view. That makes existing SharePoint, OneDrive, and Teams permissions part of the review.
For example, an investigation folder shared too broadly may already be available to employees outside the matter team. Copilot can make information from accessible files easier to find; it does not repair the underlying access decision.
Ask IT and the document owner to:
Identify where privileged and restricted files live.
Remove unnecessary access and broad sharing links.
Apply appropriate sensitivity labels and encryption.
Test a representative user's access before expanding the rollout.
Then check the destination of the answer. Permission to retrieve a source does not authorize sending its summary to every colleague in a conversation or channel. Review the actual recipients and any file-sharing permissions before distributing the output.
Authorized compliance personnel can use Microsoft's search and Purview tools to review Copilot records. OpenAI's business privacy commitments describe Business administrators' access to workspace conversations and Enterprise administrators' access to conversation audit records through compliance tools.
Ordinary usage analytics do not, by themselves, provide transcript access. Confirm the enabled controls and employee notice with your workspace owner.
Web Search and Connected Apps Need a Separate Check
Organizational Copilot can generate Bing search queries from a prompt and relevant content. Those web queries follow different handling from the prompt and response: the Microsoft DPA and EU Data Boundary do not apply, although Product Terms add confidentiality and other protections.
Microsoft removes user and tenant identifiers from those queries and says they are not used to train foundation models or shared with advertisers. That does not make a query containing a confidential project name harmless. Decide whether web search is appropriate for the matter and configure the available administrator controls.
ChatGPT search can also send rewritten queries to search partners. Connected apps can receive relevant conversation context, and their own terms govern the data shared with them. Check each app's permissions and data handling before enabling it for legal work.
Review models and agents as well as connections. Microsoft identifies third-party model subprocessors and additional terms, and currently excludes Anthropic models from its EU Data Boundary commitments. Record which providers and features are enabled instead of writing that all Copilot data always stays inside one tenant or region.
Do These Protections Preserve Attorney-Client Privilege?
No product choice guarantees privilege. Confidentiality controls support the analysis, but counsel still needs to consider the communication's purpose, participants, applicable law, and how the tool was used.
In United States v. Heppner, the court rejected privilege and work-product claims for the defendant's Claude exchanges on the facts before it, including his self-directed use and the applicable privacy terms. Its discussion of counsel-directed use did not create a safe harbor for enterprise or legal AI.
ABA Formal Opinion 512 calls for technology competence and assessment of confidentiality risks, including when informed client consent is required. Review the service and the representation rather than treating encryption or a no-training promise as the whole answer.
Contractual confidentiality also needs its own review. A data processing addendum addresses personal-data processing; your NDA may separately restrict recipients, use, or disclosure of commercial information. Our ChatGPT confidentiality guide explains that contract question in more detail.
Where GC AI Fits in a Legal Team's Privacy Review
Organizational Copilot and ChatGPT Business or Enterprise can provide meaningful business-data protections. For substantive in-house legal work, we recommend evaluating GC AI on how it carries your team's standards and source checking through to a decision.
Our Services Agreement includes confidentiality obligations and a commitment not to use Customer Data to train generative AI models. Our security documentation describes encryption, segregated customer data, SOC reporting, and restrictions on model-provider training. It qualifies provider zero-data-retention arrangements as maintained wherever feasible.
Those provider arrangements are different from GC AI's own application storage. Your team should review the applicable agreement, deletion process, subprocessors, and enabled features for its use. On Teams and Enterprise plans, organization administrators can enable or disable model providers, with changes recorded.
Run an AI Vendor Review in GC AI
Start with a public or synthetic example while evaluating the product. Once your organization approves the workspace and data, use the same method on an actual vendor assessment:
Supply the review file: Put the vendor's agreement, DPA, privacy policy, subprocessor list, and your company's requirements into a GC AI Project. Our Projects documentation explains how related chats and files share context and how you control project access.
Ask for evidence-backed findings: Request a table covering training, retention, authorized access, onward disclosures, transfers, and deletion. Require the supporting provision, any exception, and "not established" where the documents do not answer the question.
Apply your legal positions: Our Playbooks apply defined standards to contract review. Use your approved positions to flag gaps in the agreement, such as a permitted-use clause broader than your company can accept.
Verify the analysis: Use our Exact Quote source-document view to inspect the cited language. Counsel should read the surrounding clauses, resolve inconsistent documents, and distinguish a contractual promise from a configurable setting.
Prepare the decision: Turn the checked findings into proposed changes and a short recommendation for the business. Our GC AI for Word supports review and redlining in the document; counsel approves the final language and saves the agreed requirements for future reviews.
This is a proposed evaluation workflow, not a claim that the software approves vendors or establishes privilege. The useful result is a review counsel can inspect, negotiate from, and explain.
Choose the Workflow Your Lawyers Can Defend
Use one representative vendor agreement and the same requirements to test each approved platform. Compare the completeness of its findings, whether its citations support its conclusions, and the work counsel must do to turn the result into a decision.
GC AI deserves that test when your team needs company-specific legal review and a source-linked path from the agreement to a recommendation. Ask for a demonstration using that task, then decide which matters and data your team will approve.
For Word-specific workflow comparisons, see our best legal AI for Microsoft Word guide.
Frequently Asked Questions
Is Copilot More Secure Than ChatGPT?
There is no single winner across every account and workflow. Compare organizational Copilot with ChatGPT Business or Enterprise, and personal Copilot with personal ChatGPT. Then assess the contract, retention, access, and connected services your company will actually use.
Can My Employer See What I Do on Copilot?
Authorized organizational administrators and compliance staff can access Copilot interaction records through supported search and compliance tools. That does not mean every manager can read every chat. Ask your administrator which roles have access and what your employee notice says.
Does Turning Off Training Delete My Chats?
No. A model-training control governs whether eligible content can improve a model. Chat history, saved files, Memory, and compliance records have separate controls and retention rules. Review each relevant copy before treating deletion as complete.
Should an In-House Team Consider GC AI Instead of Copilot or ChatGPT?
Yes, when the task needs company-specific legal standards, source-linked analysis, and a reviewed legal output. Compare GC AI using a representative agreement and your requirements. Confirm its contract and data controls, then judge how much work counsel needs to verify the findings and finish the advice.
Does Using GC AI for Word Keep Everything Inside Microsoft 365?
Do not assume that an add-in's location determines its data handling. GC AI for Word brings our legal review tools into Word, but you should assess GC AI's agreement, processing, storage, and subprocessors as part of the workflow. An integration does not inherit every Microsoft commitment automatically.
