Claude for Nonprofits: Data handling and your AI use policy

What you'll have at the end

Your organization has Claude seats. You may be just starting out or mid-rollout. Some may be sitting unused, and some staff may already be using personal AI accounts you can't see. Training and use cases are part of the answer, but the first step is to put in writing how to use AI safely at work and what the guardrails are. This tutorial is that foundation, in three phases.

By the end, you'll have:

1. Answered "how is my data used by Claude?" in writing, with links to the public pages your executive director (ED), board or legal counsel can check for themselves.
2. Written a one-page AI use policy that leadership has signed and that gives staff clear guardrails around AI use.
3. Invited every staff member to your organization's Claude account and told them what the policy allows and what it doesn't. AI use moves off personal accounts and onto the account you govern.

This isn't product training or a settings walk-through: for those see Claude Academy and the Claude Enterprise Administrator Guide, and if you haven't set up your account yet, start with Getting started with Claude for nonprofits.

This tutorial covers both Enterprise and Team plans. It's written for Team. Where Enterprise differs, an On Enterprise callout says so.

Before you start

Who this is for. You, the admin in charge of your organization's Claude settings, holding the Owner or Primary Owner role (roles below). If Organization settings > Members says you're an Admin, pair up with an Owner. Throughout, "you" means you, the admin, and "staff" means everyone at your organization who uses Claude.

Who you need. On Team: mostly you. You'll still send three questions to staff in Phase 2, and whoever leads the organization signs the half-page policy at the end.

Setup time. Team: about 30 minutes of your time, and a few days for staff to answer three questions.

Phase 1 - Understand how Claude handles your data

Nonprofits ask us the same six questions about their data. Each is answered below in plain English, with a link to the public page behind it. This summarizes Anthropic's public documentation and isn't legal advice. If anything here differs from your organization's agreement with Anthropic (for most Team plans, the Commercial Terms you accepted at signup), the agreement governs.

Note your organization's answers as you go, as you will use them in Phase 2.

Team vs Enterprise

Default data handling is the same on both plans: no training on your data by default, same default privacy, same security certifications (SOC 2, ISO 27001, ISO 42001). Both plans have single sign-on (SSO), domain verification and just-in-time (JIT) provisioning (accounts created the first time someone signs in).

1. What data goes into Claude?

The data you already have. Claude works with what staff give it: what they type or paste, files they upload (spreadsheets, PDFs, documents), and, if you turn connectors on, what it can read from tools like Google Drive or Microsoft 365. It can't reach anything you haven't connected, and with connectors on, it pulls in only what's relevant to what someone asks.

In practice that looks like a development director uploading last year's gift spreadsheet and asking which lapsed donors to call first, or a program lead pasting in survey responses to find the themes. Start there: the spreadsheets, documents and notes staff already handle every day.

Most of this information is already covered by rules you have today.

What's new is the conversation history: what staff asked and what Claude answered. The next five questions cover it: whether it's used to train models, who can see it, how long it's kept and where, and what changes when you connect other tools.

2. Is our data used to train Claude?

No. By default, Anthropic does not use conversations (including inputs and outputs), files or connected data from Team or Enterprise accounts to train its models. This is the same on both Team and Enterprise plans. Learn more at our Privacy Center and in the Commercial Terms of Service.

The exception you control: when a staff member clicks thumbs-up or thumbs-down on a response, that conversation is submitted to Anthropic as feedback and may be used to improve Claude (see Manage user feedback settings). Owners can turn it on or off, so check how yours is set.

3. Who can see our conversations?

• Colleagues. Other staff see a conversation only if the person who had it shares it with them. Sharing a project (a workspace with its own instructions, files and chats) shares its instructions and files, not the chats in it. Each person's chats stay private by default. Owners decide whether projects can be shared or made public (visible to everyone in your organization, not outside it) at all.
• Owners and Admins can't browse conversations in the app either. A few routes do reach content, and you decide who holds them. Both plans: the Primary Owner can run the organization-wide data export of all conversations.

• Memory, if you turn it on, is private to each person, though it's included in organization data exports. See memory controls for Team and Enterprise plans.
• At Anthropic. Anthropic may process data for internal processes such as safety-related review, product support or incident response. Review the Data Processing Addendum for details.
• Outside your organization. People outside your organization can see only artifacts (something Claude built in a chat, like an interactive page or tool), and only if your settings allow it. Owners control outside sharing, by email invitation and by link. Email invitations are on by default on Team and off on Enterprise; you'll set both in Phase 3. See Share artifacts.

4. How long is data kept, and can we delete it?

• Conversations stay in your account until the staff member deletes them (or, on Enterprise, your retention period removes them). Once deleted, they leave that person's history immediately and Anthropic's back-end systems within 30 days, except where the law requires longer retention or as described below.
• Content that automated systems flag under Anthropic's Usage Policy (Anthropic's own rules for what Claude may be used for, separate from your policy) can be kept for up to two years, and the classification itself (a score, not your content) for up to seven years. This applies to flagged content, not everyday conversations.
• On Team, there's no auto-delete: conversations stay until staff delete them.

Learn more about how long Anthropic stores your organization's data.

5. Where does our data live, and who else handles it?

• Data is stored in the United States. By default, processing may route through data centers in the US, Europe, Asia and Australia (Enterprise plans on usage-based billing can restrict inference to the US). See where Anthropic's servers are located.
• Anthropic uses a published list of sub-processors: cloud hosting, support, billing and similar vendors. Anthropic posts changes to the Trust Center and gives advance notice of new sub-processors under its Data Processing Addendum; you can subscribe to the Trust Center news feed.
• Anthropic holds SOC 2 Type II, ISO 27001:2022 and ISO/IEC 42001:2023 (AI management) certifications. Reports and certificates are available from the Trust Center, some on request. See what certifications Anthropic has obtained.

Review security details at the Anthropic Trust Center.

6. What happens when we connect Claude to our other systems, or use skills?

Connectors (Google Drive, Gmail, Microsoft 365, Slack, your CRM via a custom connector, and others) let Claude read from a system when it's relevant to what a staff member asks, and take actions in it if you allow them.

• In your connected systems, Claude only sees what the user has access to. Take Google Drive: if a staff member can't open a file in Drive, Claude can't open it for them either. Each person signs in to the connected tool with their own account. Two exceptions you'd set up on purpose: a custom connector (one your IT team or a vendor builds) that uses one shared login reaches whatever that login can access, and with enterprise-managed authorization IT authorizes a connector once for everyone. Making a connector available doesn't hand Claude anyone's login.
• Claude retrieves what's relevant at the time of the request. Whatever it retrieves becomes part of that conversation and follows the same rules as questions 2-5: not used for training by default, visible to other staff only if the conversation is shared and to admins as described in question 3, same retention. The connected service itself still handles your data under its own terms. See how the Google Workspace connectors handle your data.
• Owners decide which connectors are available at all, what each may do (always allow, needs approval or blocked, per action) and, on Enterprise, which groups get which. Desktop extensions in the Claude Desktop app are open by default until you turn on the allowlist. You'll set all of this in Phase 3.

Skills are packaged instructions and reference files that teach Claude how your organization does a specific task, such as "draft a grant report in our format." Anything you put in a skill is treated like anything else you give Claude: not used for training by default, visible only where it's shared and to admins as described in question 3. Owners decide who can create, share and publish skills; see provision and manage skills for your organization.

What this means for donor, client and health data

• Donor and client information: most of it can go into your organization's Claude account within your policy.
  • Donor data such as names and giving history: yes, within your policy.
  • Client-identifying information such as names and case notes: also within your policy, and the Phase 2 template makes it an ask-first item.
  • Health and disability details: leave them out unless your organization has the HIPAA-ready version of Claude (Enterprise plans only); if protected health information (PHI) is central to the work you want help with, read "Covered by HIPAA?" below first.
  • Immigration status: handle it like health details.
• In every case, by default none of this is used for training, it's visible only to the staff it's shared with and to admins as described in question 3, and your organization is responsible for complying with the laws that apply to its data (see the Commercial Terms and the Data Processing Addendum).
• Covered by HIPAA? If you're a clinic or a health plan, or you handle health records for one, protected health information (PHI) needs a Business Associate Agreement (BAA) with Anthropic. That's available with the HIPAA-ready configuration on Enterprise plans, and not on Team. On Team, keep PHI out of Claude entirely and use it for everything else: grant writing, board reports, policies, de-identified program summaries. If PHI is central to the work you want help with, contact our sales team about Enterprise.

• Not sure whether HIPAA applies to you? That depends on the services you provide and who you provide them for. Anthropic can't make that determination for you. Check with your legal counsel or compliance lead before client health details go in.
• When in doubt, ask one question: would I be comfortable if a colleague saw this in a shared project? If not, leave it out or strip the identifying details first. Staff can apply that test without legal training.

Phase 2 - Write your AI use policy

If you don't have a written AI policy yet, that's normal. Write something short first and treat it as a living document. You're not starting from scratch: as Phase 1 said, your existing confidentiality rules still apply, and this policy mostly says how they carry over to Claude.

Don't try to cover every team and every scenario in one pass. Start where the need is clearest (one program team, one type of data, one connector) and add to it as staff bring you new use cases and as Claude changes. Holding out for one comprehensive policy is how organizations stall, and the next product update would reopen it anyway.

Step 1 asks staff before you draft anything, so the policy answers their real questions in their own words.

Step 1. Run a three-question staff check-in (give people 3-4 days to answer)

Before you write anything, find out what's already happening. Leadership and staff usually see AI use differently, and NTEN and The Bridgespan Group's State of Nonprofit AI Adoption and Governance report (2026) describes the gap. A short check-in closes most of it and gives you raw material for the policy.

Send the questions below to all staff (Team) or through team leads (Enterprise), as an anonymous form:

Step 2. Make six decisions

Work through these with your signer. Each decision becomes a short section of the policy. Under each one is example text from a fictional organization, Riverbend Community Services, that you can adapt.

1. Purpose - why we use AI. Frame it around the time it gives staff back.

2. What we encourage, what needs a conversation first, and what's off-limits. The staff check-in answers go here (see Step 3). Lead with what you want staff to use Claude for, and keep the not-permitted list short. A long list of limits reads as a warning, and staff stop using the ask-first route.

3. How Claude handles our data. A few lines, lifted from Phase 1. Say whether feedback is on or off.

4. A person checks the output. Every version of your policy needs this line. Checking well is a skill: the AI Fluency for nonprofits course teaches it as Discernment. Point staff there.

5. Which tools. Use the organization's account. Name the connectors and skills that are on and tell staff to use them: that's where most of the time savings are. Say whether memory is on.

6. Policy owner, Primary Owner, signer, review date. Name the policy owner and the Primary Owner on the Claude account.

Step 3. Let Claude draft decisions 2 and 5 from what staff said

Staff find a policy easier to recognize and follow when it uses their own words. You're an Owner and this is the account you govern, so it's fine to use it before the policy is signed. Open your organization's Claude account, paste in the staff check-in responses and run this prompt:

Review what comes back with your signer before it goes in the policy.

Step 4. Get it signed

Send your draft policy from Step 3, plus Phase 1 of this tutorial for the facts behind it, to your signer, and ask for a yes or comments by a specific date. If legal or your board wants the underlying documents, the claims in Phase 1 link to Anthropic's public pages.

For a starting point, use the AI use policy template from Step 2. No Claude account yet? Use the Google Docs version instead. If you want a fuller policy document, NTEN's AI for nonprofits resource hub has two downloadable policy templates (NTEN's and ANB Advisory Group's) plus community examples. The Integration lesson in AI Fluency for nonprofits walks you through drafting one step by step. Most organizations under 50 staff won't need more than one page. Larger ones usually start from one of the NTEN templates.

Phase 3 - Roll it out and keep it current

You have a signed policy. Now set the handful of settings that enforce it, then tell staff.

Step 1. Set the few settings that enforce the policy

Most of these are under Organization settings.

The policy itself. Let staff ask it. If your nonprofit is verified, the Claude for Nonprofits plugin is installed by default and includes an AI use policy lookup skill, so staff can ask Claude "can I use Claude for this?" and get an answer from your rules.

• Save the signed policy where staff can already reach it, such as a shared Drive folder or your intranet.
• The first time someone asks, Claude asks where the policy lives: they point it at the document through a connector or drop it into the chat, and if memory is on, Claude remembers it for next time.
• The plugin needs Skills and Cowork (the mode where you hand Claude a whole task and it works through it across your files and connected tools) turned on for your organization (see Manage plugins for your organization).
• To save staff that first step, an Owner can add a line under Organization settings > Organization and access > Organization instructions, for example: "Our AI use policy is the document called AI use policy in the shared Policies folder".

Both plans, if staff use Claude Code: have developers sign in to Claude Code with the organization's account, so the policy covers developers as well as chat. IT can restrict logins to your organization with Claude Code managed settings.

Step 2. Tell staff

1. Invite everyone you have seats for, not just the people who asked. The policy says "use our account, not personal ones," so everyone needs an invitation or that line is empty.

• Each invitation takes up a seat the moment you send it, so check your count under Organization settings > Organization and access (Total seats) first (see purchase and manage seats on Team or Enterprise plans, or on Enterprise ask your Anthropic contact).
• Fewer seats than staff? Invite in waves: start with the teams whose check-in answers showed the most use, tell everyone else when their turn comes, and keep the policy organization-wide from day one.
• For a structured pilot, the launch section of the Enterprise Administrator Guide covers champions and phased rollout.
• Send the invitations and the note today; the domain work below can follow when IT is ready.
• Both plans: two ways to add people, and they differ when seats run out.
  • By email (Organization settings > Members): each pending invitation uses one of your purchased seats.
  • By invite link (Organization settings > Organization and access; on by default on Team, off on Enterprise, not available with SSO) or organization discovery (staff on your domain are prompted to join when they sign up): once your purchased seats are used, each new joiner can add a paid seat automatically (see Find and join a Team or Enterprise organization).
  • To stay inside your count, invite by email. See manage members on Team and Enterprise plans.
  • Staff whose email is on another domain need that domain added to your allowed domains (the email domains permitted to join, in Organization settings) first.
• Both plans, to stop new personal accounts on your domain: verify your email domain (whoever manages your website domain adds one DNS record; it can take up to two days to show as verified) and turn on Restrict organization creation. After that, nobody can create a new personal Claude account with a work address; they join yours by invitation, invite link or organization discovery instead.

2. Communicate, more than once. Send the note below today, then say it again at your next all-staff meeting, ask managers to raise it with their teams, and pin the policy where staff already look. To start, copy Riverbend's and swap the brackets (drop the ask-Claude line if the plugin isn't on for your organization):

Step 3. Keep it current

A week after the invitations go out, open Organization settings > Members: resend anything still pending, then open Analytics from your user menu to see who has started. Then put a 90-day reminder in your calendar. Revisit sooner if:

• Anthropic changes how it handles data. Follow the Help Center release notes and in-product notifications. For the source documents, click Subscribe to updates on the Trust Center, and keep an eye on the Privacy Center and the Commercial Terms.
• You want to connect a system that holds client or donor data
• Staff keep asking the policy owner the same question, and the policy doesn't answer it

What's next

Enterprise admins: continue with the Claude Enterprise Administrator Guide for configuration in depth. Everyone: send staff the Claude for Nonprofits session for staff, A Staff Guide to Making Claude Useful on Day One, along with the note above. Questions after reading this: the session for admins, An Admin's Guide to Data Handling and AI Use Policy, walks through this tutorial with live Q&A.

Learn more

• Getting started with Claude for nonprofits: nonprofit pricing, verification and the nonprofit connectors, from Claude Academy.
• AI Fluency for nonprofits: the full course from Anthropic and GivingTuesday. Its Integration lesson is the policy-drafting exercise referenced in Phase 2.
• Claude 101: the self-paced introductory course on Claude Academy, for staff once they have their invitation.
• Is my data used for model training? and How long do you store my organization's data?: the two Privacy Center articles to forward to whoever signs.
• Anthropic Trust Center and the Data Processing Addendum: certifications, sub-processors, security documentation and the contractual data terms.