AI Policy for Nonprofits: Client Records, Shelters, Donors, and Grants

In BDO's Nonprofit Standards Benchmarking Survey (n = 250 nonprofit leaders), 82% of surveyed US organizations reported using AI while fewer than 10% of surveyed nonprofits had a written AI policy (Dataconomy, Apr 8, 2025). That gap turns dangerous inside frontline human services. In February 2026, the British Association of Social Workers and the Ada Lovelace Institute studied 39 social workers across 17 local authorities. Researchers documented an AI note-drafting tool inserting "suicidal ideation" that a client never spoke (BASW, Feb 12, 2026). Meanwhile, top-ranking web downloads for an AI policy for nonprofits copy corporate HR templates and never mention VAWA, FVPSA, HIPAA, 42 CFR Part 2, HUD HMIS, or FERPA (Layer3Labs template). A 14-person housing charity or shelter cannot run on a startup handbook when volunteers and unpaid board members work on personal phones.

Small nonprofits carry higher legal sensitivity per employee than most businesses of equal size. A 12-person charity may hold Continuum of Care eviction records, youth files, major-donor estate notes, and federal grant budgets on one drive while relying on 30 volunteers and 11 outside board directors without single sign-on. Worse, in a non-scientific Philanthropy.org reader poll, 48% of respondents reported never checking AI outputs with another tool (Philanthropy.org poll).

Because ed.cafe is published by No Worker Left Behind, a Boston 501(c)(3) (EIN 88-3268312), we publish this policy ungated. See our governance model on the ed.cafe Council page. This guide shares operational information as of October 2026, not legal advice; confirm grant rules with your funder or counsel. Commercial firms without shelter or grant mandates can use our one-page AI use policy for a small business.

Why corporate AI templates fail on VAWA shelter records, FVPSA grants, and HUD HMIS data

Generic corporate policies allow staff to paste confidential files into "enterprise cloud tools" under standard vendor terms, a rule that violates federal domestic-violence statutes and risks physical harm or immediate grant loss.

Corporate privacy treats client files like a locked office cabinet where an outside vendor may enter after signing a routine NDA. Federal domestic-violence law treats a shelter's intake ledger like an unmarked safe house whose street address cannot go to third-party couriers even inside a sealed, encrypted envelope. Where the safe-house analogy breaks down is replication. Handing one paper ledger to a courier creates one copy. Pasting a "name-stripped" survivor intake narrative into a cloud chat tab creates server logs, safety queues, and third-party subpoena targets outside the shelter's custody.

Four statutory regimes govern human-service client records:

1. VAWA Confidentiality (34 U.S.C. 12291(b)(2)): Under the Violence Against Women Act, grantees and subgrantees shall not disclose, reveal, or release personally identifying information (PII) or individual service information, regardless of whether data is encoded, encrypted, hashed, or otherwise protected (US Code 34 U.S.C. 12291). Sharing requires informed, written, reasonably time-limited survivor consent or a narrow statutory or court exception. Consent can never be a condition of shelter or services. Because external AI chatbots and transcription apps transmit narratives to third-party servers, victim-service programs must ban entering survivor details, safety plans, or shelter addresses into external AI tools (NNEDV Safety Net).
2. FVPSA Confidentiality (42 U.S.C. 10406(c)(5)): Family Violence Prevention and Services Act grants enforce parallel strict bars against entering survivor PII into shared or public databases (US Code 42 U.S.C. 10406).
3. HUD HMIS (24 CFR 578.103) & Comparable Databases: Continuum of Care housing agencies must protect Homeless Management Information System (HMIS) records and never paste eviction, disability, or income files into chatbots (eCFR 24 CFR 578.103). Victim Service Providers are barred from shared HMIS altogether and must use isolated comparable databases.
4. HIPAA, 42 CFR Part 2, FERPA, and State PII Laws: Recovery and clinical agencies face HIPAA BAAs and updated 42 CFR Part 2 Substance Use Disorder enforcement (live since February 16, 2026); see using AI for case notes under HIPAA BAAs and Part 2. Youth programs receiving school records face FERPA (20 U.S.C. 1232g). Nonprofits holding Massachusetts residents' SSNs, IDs, or bank numbers (such as RAFT files) must maintain a Written Information Security Program under 201 CMR 17.00 (Mass.gov 201 CMR 17.00).

How to govern donor wealth screening, CRM exports, and AFP ethical standards

Uploading donor spreadsheets, home addresses, family notes, or gift histories into consumer AI chatbots or free browser extensions violates the Association of Fundraising Professionals (AFP) Code of Ethical Standards.

Standards 15 through 18 of the AFP Code classify donor and prospect lists as confidential organizational property. Fundraisers must never disclose privileged donor details to unauthorized third parties and must honor donor privacy choices (AFP Code of Ethical Standards). When a development staffer exports 1,500 rows from Bloomerang, Little Green Light, or Salesforce into a personal ChatGPT tab to ask "Which 30 donors should our ED take to lunch?", every street address, gift total, and family note lands on an unvetted consumer server.

Small development teams protect donor trust with three rules:

• Segment inside your CRM, draft blanks in AI: Filter gift tiers ($500-$2,499, lapsed 18+ months) directly inside your CRM's report builder. Ask AI only for blank thank-you letter templates ([Donor First Name], [Program], [Impact Stat]), then run mail merge locally.
• Require written DPAs for wealth screening: Vet prospect-research tools to confirm public SEC and property matches run under a Data Processing Addendum that bars pooling or reselling your donor list across other charities.
• Verify every major-gift letter: Chatbots routinely mix up two local donors or invent past board service. Every letter signed by an executive director or board chair needs a human check against the live CRM record.

What federal agencies and private foundations allow and ban in AI-assisted grant writing

Grant-seeking nonprofits face a clear split: most funders allow AI outlining and prose editing when humans verify every fact, whereas federal agencies strictly ban uploading grant proposals into AI tools during peer review.

On June 23, 2023, the National Institutes of Health issued Notice NOT-OD-23-149, prohibiting peer reviewers from using generative AI tools to analyze or critique grant applications because uploading proposals breaches peer-review confidentiality (NIH NOT-OD-23-149). The National Science Foundation (NSF) enforces the same merit-review confidentiality ban on reviewers while holding applicants fully accountable for proposal accuracy and originality (NSF Notice). Whenever your director sits on a United Way, municipal CDBG, or community-foundation panel, apply the federal rule: never upload another charity's grant application into an AI summarizer.

Test a safe workflow on a 12-person food-pantry charity applying for a $150,000 foundation renewal (assumptions stated):

1. Failure path: At 4:45 PM Friday, a freelance grant writer prompts a chatbot for "recent Roxbury food-insecurity rates and two composite client quotes." The model invents a fake city URL and fabricates quotes presented as real voices. On federal pass-through grants (2 CFR Part 200), false claims risk disallowance; on foundation grants, they wreck trust.
2. Safe 3-step path: Pull verified Census ACS tables and internal aggregate counts (4,120 households served in FY2025; zero PII) into a locked sheet first. Use an approved zero-training workspace tool solely to tighten prose around those locked numbers. Check every citation via our AI-assisted report verification routine and answer any funder AI-disclosure question truthfully.

Why volunteers and board members cause silent nonprofit data leaks

In a 15-person charity, paid staff are often outnumbered three-to-one by unpaid board directors, pro bono helpers, interns, and evening volunteers using unmanaged personal phones and laptops.

If your AI policy lives only inside an employee payroll handbook, the board treasurer pasting draft audit financials into a personal iPad and the Saturday volunteer translating intake forms on a phone app never see the boundary. Board members also join confidential Zoom executive sessions with consumer AI notetakers (Otter, Fireflies, Read.ai) set to auto-join and email transcripts externally. In all-party wiretap states like Massachusetts (M.G.L. c. 272 s. 99), unconsented recording carries civil and criminal liability. Enforce our meeting notetaker consent and bot-removal playbook and hand volunteers our guide on what not to paste into an AI tool at work.

The ed.cafe One-Page Small Nonprofit & Shelter-Safe AI Policy (Staff, Board & Volunteers)

Adopt the ed.cafe One-Page Small Nonprofit & Shelter-Safe AI Policy (Staff, Board & Volunteers) at your next board meeting. Attach the sign-off box to staff onboarding, grant-writer contracts, volunteer orientation, and annual board packets.

ORGANIZATION: _________________________________  EFFECTIVE DATE: ______________
POLICY OWNER (ED / Privacy Lead): _______________________  CONTACT: ___________

1. SCOPE (EVERYONE WHO TOUCHES OUR MISSION)
   Applies equally to all paid staff, interns, AmeriCorps/fellows, contractors
   (including freelance grant writers), volunteers, and Board members on any
   agency or personal computer, tablet, or smartphone.

2. APPROVED ACCOUNTS VS PERSONAL TOOLS
   - Approved Org Tools (Training OFF): [e.g., M365 Copilot / Workspace Gemini]
   - Personal / Free AI Accounts: Permitted ONLY for Green-Tier blank templates
     or public text. Never log into personal AI on shared intake computers.

3. TRAFFIC-LIGHT DATA TIERS (VAWA, HMIS & DONOR RED LINES)
   [RED - STRICTLY FORBIDDEN IN ANY EXTERNAL AI TOOL, EVEN IF "ANONYMIZED"]:
   a) Domestic-violence, sexual-assault, stalking, or shelter client details,
      locations, or safety plans (barred by VAWA 34 U.S.C. 12291(b)(2) & FVPSA
      42 U.S.C. 10406(c)(5), even if names are removed or text is encrypted).
   b) HUD HMIS housing files (24 CFR 578.103), clinical/HIPAA charts, 42 CFR
      Part 2 recovery notes, or youth FERPA records (unless a written BAA/QSOA
      + explicit ED and client consent are executed in advance).
   c) SSNs, government IDs, bank/RAFT numbers, immigration files, or passwords.
   d) Donor names, addresses, gift histories, or wealth notes (AFP Ethics Code).
   e) Other organizations' grant proposals reviewed on funding panels (NIH rule).

   [YELLOW - APPROVED ORG ACCOUNT ONLY + MANDATORY HUMAN CHECK]:
   - Internal SOPs, non-confidential schedules, aggregate de-identified service
     counts, and grant prose editing built strictly from pre-verified stats.

   [GREEN - SAFE FOR ANY TOOL]:
   - Blank bracketed outlines ("Write a blank 5-bullet DAP skeleton"), Excel
     formulas on dummy numbers, public law lookups, and generic appeal outlines.

4. CLIENT INTAKES, TRANSLATION & ZOOM NOTETAKER BOTS
   - Unapproved AI recording bots (Otter, Fireflies, Read.ai, phone voice apps)
     are BANNED in client visits, shelters, donor calls, and Board executive
     sessions; meeting hosts must remove unapproved bots immediately.
   - Live client translation must use approved interpreter lines ([LanguageLine]),
     never free consumer phone voice or chat apps.

5. GRANT WRITING, REPORTS & PUBLIC HONESTY
   - Humans must verify every statistic, citation, and URL against primary data.
   - Never use AI to invent "composite" client quotes presented as real people
     or publish unlabeled AI-generated photos depicting vulnerable clients.
   - Disclose AI writing assistance whenever a funder or government form asks.

6. 60-MINUTE NO-BLAME SAFETY REPORTING
   If Red-Tier data is accidentally pasted or recorded, tell [Policy Owner]
   within 60 minutes so logs can be purged and statutory steps checked.

ACKNOWLEDGMENT: Name: ___________________ Role: [ ]Staff [ ]Volunteer [ ]Board
Signature: _________________________________________ Date: ____________________

Pin the ed.cafe One-Page Small Nonprofit & Shelter-Safe AI Policy (Staff, Board & Volunteers) inside your volunteer binder and board portal. Four clear rules read before a shift protect survivors and donors better than a 30-page binder nobody opens.

Where this stops being true

One-page policies fit most 5-to-50 person charities, but specialized legal counsel and segmented IT enclaves must take over in five situations:

• Not legal advice: This guide shares operational governance education from No Worker Left Behind as of October 2026, not legal counsel. Subpoenas, statutory consent waivers, and breach notices require a licensed attorney.
• Multi-service agencies: Community action agencies running clinics (HIPAA + 42 CFR Part 2), VAWA shelters, and Head Start classrooms under one roof must isolate accounts so clinical BAA scribes never touch VAWA shelter programs.
• State executive contractors: Nonprofits holding state contracts may inherit public-sector mandates; Massachusetts Policy AI.001 v1.1 (January 31, 2025) covers state contractors, barring AI recording without agency counsel approval and requiring AI deliverable labels (Mass.gov AI.001 PDF).
• Unionized nonprofits: Where staff belong to SEIU, AFSCME, or UAW bargaining units, deploying AI monitoring or note tools can require collective bargaining.
• Cross-border NGOs: Charities operating in Europe or serving international refugees face GDPR transfer rules and EU AI Act high-risk obligations on aid-eligibility tools.

Questions people ask

Key terms

• VAWA Confidentiality (34 U.S.C. 12291(b)(2)): Federal rule barring victim-service grantees from disclosing individual or identifying survivor data, even if encrypted, without informed, written, time-limited consent.
• FVPSA (42 U.S.C. 10406(c)(5)): Federal mandate prohibiting domestic-violence programs from entering survivor PII into shared or public databases.
• HUD HMIS (24 CFR 578.103): Continuum of Care privacy rule governing homeless-service databases; DV providers use isolated comparable databases instead.
• AFP Code of Ethical Standards: Association of Fundraising Professionals rules requiring strict confidentiality over donor lists, gift histories, and prospect notes.
• NIH Notice NOT-OD-23-149: June 2023 federal directive barring grant peer reviewers from entering proposals or review drafts into generative AI tools.
• Zero-PII Blank-Template Workflow: Prompting AI solely for generic outlines with bracketed placeholders so zero client or donor bytes leave your database.

Sources

1. Dataconomy, "82 percent of nonprofits use AI (n = 250 BDO survey)" (Apr 8, 2025): Dataconomy
2. BASW / Ada Lovelace Institute, "Social workers risk legal action over AI inaccuracies (n = 39)" (Feb 12, 2026): BASW
3. Philanthropy.org, "Nonprofit AI Poll (non-scientific)" (current): Philanthropy.org
4. Layer3Labs, "Nonprofit AI Policy Template" (2026): Layer3Labs
5. US Code, "34 U.S.C. 12291(b)(2) (VAWA)" (current): US Code
6. US Code, "42 U.S.C. 10406(c)(5) (FVPSA)" (current): US Code
7. NNEDV Safety Net, "Agency Use Resources" (current): NNEDV
8. eCFR, "24 CFR 578.103 (HMIS)" (current): eCFR
9. AFP, "Code of Ethical Standards" (current): AFP
10. NIH, "NOT-OD-23-149 (Peer Review AI Ban)" (Jun 23, 2023): NIH
11. NSF, "Generative AI in Merit Review" (Dec 2023): NSF
12. Mass.gov, "201 CMR 17.00" (current): Mass.gov
13. Mass.gov EOTSS, "GenAI Policy AI.001 v1.1" (Jan 31, 2025): Mass.gov PDF