--- "title": "Can we use AI on documents covered by an NDA?" "description": "A practical guide to using AI on confidential documents covered by NDAs, data room terms and client agreements, including what to check in the contract, which tool settings matter and how to set guardrails for agents." "section": "AI Engineering" "sidebarTitle": "Using AI on NDA documents" "published": "2026-11-05T14:12:54.875Z" "updated": "2026-11-05T14:12:54.875Z" --- ## Can we use AI on documents covered by an NDA? It depends on what the agreement says and which tool you use. Many NDAs limit who may see the information and require it be protected, so pasting it into a personal or consumer AI account is usually a bad idea. Using it in a business or enterprise account that does not train on your data, keeps data within agreed limits and lets you control retention is often acceptable, but only if the agreement does not forbid sharing with service providers. When in doubt, read the clause, ask the counterparty, or use a setup that keeps data inside your own environment. ## Who this is for General counsel, COOs, deal teams, advisors and anyone at a firm who regularly receives confidential documents from clients, counterparties or data rooms and wants to use AI on them without breaching an agreement. ## The short version - Read the agreement first. The words that matter are usually about who may receive the information, which "representatives" are allowed and how it must be protected. - Personal and free AI accounts should be off-limits for anything covered by an NDA. - Approved business accounts with no training on your data, sensible retention and access controls are the normal baseline. - Agents and integrations need tighter limits than chat, because they can read and move more data. This page is practical guidance, not legal advice. For a specific agreement, ask your lawyer. ## What to look for in the agreement Confidentiality agreements vary, but a few clauses decide most AI questions: - **Permitted recipients.** Many NDAs let you share with "representatives", which may include advisors, contractors or service providers. Some lists are narrow and only name employees and professional advisors. If AI vendors are not clearly covered, treat that as a question to resolve. - **Purpose limits.** Most NDAs allow use only to evaluate a deal or deliver a service. Using the information to train or improve someone else's model is outside that purpose, which is one reason no-training settings matter. - **Standard of care.** Many agreements require you to protect the information at least as carefully as your own confidential data. If your firm already approves a tool for its own sensitive data, that helps your case. - **Return and destruction.** If you must destroy the information when talks end, you need to know where copies live, including chat histories, uploaded files and any indexes an agent built. - **Data room terms.** Virtual data rooms often have their own click-through terms that limit downloading, copying or sharing. Check these as well as the NDA. If the agreement is silent or unclear, a short email to the counterparty asking whether use of an approved enterprise AI tool is acceptable often settles it. ## Which tool settings matter When you assess an AI tool for NDA material, look for these, and record the answer with the date you checked, because vendor terms change. As of November 2026, the major vendors offer business plans where customer data is not used for training by default, but plan names and terms differ, so check the current terms yourself. - **No training on your data**, stated in the business terms, not just a settings toggle. - **Retention controls**, so you know how long prompts, files and outputs are kept and can delete them. - **Access controls**, such as single sign-on, workspace permissions and the ability to remove users. - **Data location**, if your agreements or regulators care where data is stored. - **Audit logs**, so you can show who accessed what if a counterparty asks. Our page on [keeping client data safe with AI](/docs/keeping-client-data-safe-with-ai) covers vendor questions in more detail, and [local vs cloud AI](/docs/local-vs-cloud-ai) explains when keeping models inside your own environment is worth the effort. ## Simple rules for people Most breaches of confidentiality come from habits, not systems. Give your team a short set of rules: 1. Only use the firm's approved AI accounts for confidential material. Never personal accounts. 2. Check the deal or client folder for any AI restrictions before uploading. 3. Upload only what the task needs. Summarizing one contract does not require the whole data room. 4. Do not paste confidential details into tools connected to the open web unless they are approved for it. 5. When a deal ends, delete chats and files tied to it along with the other copies. Put these in your [AI acceptable use policy](/docs/write-an-ai-acceptable-use-policy) and cover them in training with real examples. ## Guardrails for agents and integrations An agent that can search your document system, email or a data room can see far more than one person pasting into a chat. That makes engineering choices matter: - **Scope access by deal or client.** Give an agent access to one project folder, not the whole drive. - **Respect existing walls.** If your firm keeps information barriers between teams, the agent's permissions must follow them. - **Keep an index you can delete.** If the agent builds a search index of documents, make sure it can be wiped when an agreement requires destruction. - **Log what it reads and sends.** Logs help you answer counterparty questions and catch mistakes. - **Keep a person in the loop** before anything leaves the firm, such as an email to a counterparty. See [how AI agents connect to CRM and email](/docs/connect-ai-agents-to-crm-and-email) and [maintaining AI systems](/docs/maintaining-ai-systems) for how to keep these controls working over time. ## Example: reviewing a data room A firm evaluating an acquisition receives access to a data room under an NDA that permits sharing with representatives who are bound by confidentiality. A sensible setup is to confirm that the firm's enterprise AI account meets the NDA's standard of care, download only the documents needed for the review into a deal-specific folder, give the review agent access to that folder only, and delete the folder, chats and index when the process ends. This is an illustrative setup, not a description of a specific client. ## Common mistakes - **Assuming "enterprise" means allowed.** A good tool does not override a narrow agreement. - **Forgetting the copies.** Chat histories and agent indexes are copies, and destruction clauses apply to them. - **Giving agents broad access** because it was easier to set up. - **No record of decisions.** If a counterparty asks, you want to show what you checked and when. ## When you do not need us If your firm already has approved enterprise AI accounts, a clear policy and lawyers who review NDAs for AI use, you may only need to add the habits above to training. Bring in help when you want agents working on confidential documents. ## How Ruby helps - **AI training:** teaching teams safe habits for confidential material, using your own policy. - **AI strategy and consulting:** setting rules that map agreement types to approved tools. - **AI engineering:** building agents with scoped access, logs and deletion built in. If you want to use AI on confidential documents without guessing, [book a call with Ruby](https://rubyadvisory.com/book-a-call). ## Related pages - [How to keep client data safe when using AI](/docs/keeping-client-data-safe-with-ai) - [Is local AI or cloud AI better for privacy?](/docs/local-vs-cloud-ai) - [How to write an AI acceptable use policy](/docs/write-an-ai-acceptable-use-policy) - [AI for private equity due diligence](/docs/ai-for-private-equity-due-diligence)