Private AI for Nonprofits: Donor and Client Data
Grant drafts, donor letters and case notes each carry different risk. Shelters, health and legal aid programs face confidentiality rules that a cloud chatbot can break, and a local model costs less than expected.

A domestic violence program that receives federal VAWA money may not "disclose, reveal, or release any personally identifying information" about the people it serves, and the statute adds that this holds "regardless of whether the information has been encoded, encrypted, hashed, or otherwise protected" (34 U.S.C. 12291(b)(2)). An advocate who pastes a case note into a public chatbot to tidy it up has released that information to a third party. That single fact shapes how AI for nonprofits should be set up: the useful jobs are real, but some of the data behind them cannot leave the building, and encryption in transit does not change that.
Most nonprofits do three kinds of writing AI can speed up, and they carry very different risk. Treating them alike leads either to banning everything or to exposing clients.
Three jobs, three risk levels
| Job | Typical data | Risk if it leaves the building |
|---|---|---|
| Grant drafting | Program descriptions, outcomes, budgets, sometimes client stories | Low, unless client stories are included |
| Donor communications | Names, addresses, gift history, personal notes from staff | Moderate: donor trust, and in some cases a donor's request for anonymity |
| Case notes and client letters | Identities, locations, health, immigration or legal status | High, and for some programs governed by statute |
A practical rule follows from the table. Grant prose with no client detail can go to any tool the board has approved, cloud included. Donor records go to a local model or to a cloud tool under a contract that rules out training and limits retention. Client records go only to a model running on hardware the organization controls. This is a data-class routing rule, and the general pattern is laid out in splitting work between local and cloud models.
Client data: where the law draws the line
Which rules apply depends on what the organization does and who funds it. A few common cases:
Shelters and victim services
Beyond the VAWA text quoted above, the same section allows release only with the person's "informed, written, reasonably time-limited consent", or where a statute or court compels it. Programs funded through HUD homelessness grants face a related rule: victim service providers are prohibited from entering personally identifying information into the shared HMIS system and must use a separate comparable database instead, sharing only aggregate data with the local Continuum of Care (NNEDV Safety Net). If the shared community database is off limits, a vendor's chat service deserves the same caution; a model on a machine the shelter owns, with no outbound connection, avoids the question of release.
Substance use and health programs
Federal confidentiality rules for substance use disorder records, 42 CFR Part 2, apply to "federally assisted" programs. The definition is broader than many expect: a program counts if the IRS assists it "through the allowance of income tax deductions for contributions to the program or through the granting of tax exempt status to the program" (42 CFR 2.12). A 501(c)(3) running a treatment program may therefore be covered even with no federal grant at all. Clinics that bill insurance electronically are usually HIPAA covered entities on top of that.
Immigration and legal aid
Legal services organizations carry the confidentiality duties of the lawyers and accredited representatives who work there. Immigration files hold exactly what makes a client findable: addresses, family members, employers and status, so a disclosure can put someone at physical or legal risk.
None of this is legal advice, and a program's grant agreements may add their own terms. Read them before choosing tools; the confidentiality section of a federal award is often stricter than general privacy law.
Grant drafting
Grant writing is the least sensitive job, as long as client stories are kept out or properly de-identified. A local model can draft a needs statement from the organization's own data, rewrite last year's narrative to fit a new funder's word limits, and check a draft against the request for proposals for missing sections.
Funders are starting to set their own rules on AI. NIH, for example, states that it "will not consider applications that are either substantially developed by AI, or contain sections substantially developed by AI, to be original ideas of applicants", effective for applications submitted to the September 25, 2025 receipt date and later (NIH notice NOT-OD-25-132). Most nonprofits apply to foundations, but the lesson carries: check each funder's guidelines, use the model to edit and check rather than to originate the program idea, and keep the human author's draft as the source.
A client story in a grant needs more than a changed name. In a small town, "a mother of three who fled from the county's only trucking company" identifies someone. Remove the identifying combination of details as well as the name, and have the program director approve any story before it goes into a model of any kind.
Donor communications
Donor records hold more than an address: gift history, relationships, and notes from a major gifts officer about health or family. Anonymous donors expect that anonymity to hold inside internal tools too.
The obvious AI job is personalized thank-you and appeal letters. The volume is high and seasonal, which suits an overnight batch on a small machine. Take an illustrative year-end run: 1,200 donors, a letter of about 250 tokens each, and a model that generates 30 tokens per second on one request. That is 1,200 x 250 = 300,000 tokens, and 300,000 / 30 = 10,000 seconds, about 2 hours 47 minutes. Run overnight, the drafts are ready by morning on a single modest GPU.
Three rules keep it safe:
- Facts come from the CRM, not the model. Gift amounts, dates and the program a gift supported are inserted by the merge step. A model asked to "thank them for their generous gift" will sometimes invent a figure.
- Staff notes stay out by default. A notes field that mentions a donor's illness or divorce should not be in the prompt unless a person chose to include it for that letter.
- Every letter is reviewed by a person before it goes out, at least for major donors and for the first run of any new template.
Case notes on a modest budget
Case notes are where the confidentiality rules bite, and also where staff time goes. An advocate or case manager who writes eight notes a day at 12 minutes each spends 8 x 12 = 96 minutes on them. If a local model turns a few dictated lines into a draft in the program's format, and checking that draft takes 4 minutes, the same work takes 8 x 4 = 32 minutes, saving about an hour per person per day. Measure both numbers in the program's own pilot before relying on them.
The model's job is narrow. It reorganizes what the worker said; it does not add assessments, risk ratings or service recommendations. Any field it cannot fill from the dictation stays marked "not stated".
Hardware without a hardware budget
Nonprofits rarely have a line for a new GPU server, which makes donated and second-hand hardware attractive. Text drafting needs memory more than speed. A used workstation with a single 24 GB graphics card can run a capable mid-size model for a small team, and the inspection checks before trusting a second-hand card are covered in buying used GPUs for AI. For a team that only needs overnight batches, even a donated machine with no usable GPU can do the work slowly; what CPU-only inference handles is explained in running an LLM without a GPU.
Wipe and reinstall a donated machine before use, since its old drives may hold someone else's data, and keep it in a locked room: a box holding case notes is as sensitive as the file cabinet beside it.
Board-level policy
Because the risks sit in client data and donor trust, AI use is a governance question, and the board or its executive committee should approve the rules rather than leave them to whoever set up the tool. The staff-facing document itself, with approved tools, data classes and review rules, follows the structure in writing an AI acceptable use policy for staff. What the board adds is the part staff cannot decide for themselves:
- Grant terms. A list of current awards with confidentiality clauses, and a named person who checks each new award against the approved tools before money is accepted.
- Client notices and consent forms. If intake paperwork tells clients how their information is used, the board confirms the AI setup matches that promise, or the paperwork changes first.
- Donor privacy statement. The published statement on the website should still be true once donor records feed a model, including anonymous gifts.
- Volunteers and partners. Hotline volunteers and partner agencies are covered by the same rules as staff, and the client-data line is part of their orientation.
- An annual review. Grants, tools and laws change, so the board looks at the policy once a year alongside the audit.
For most programs the staff version fits on one page, with the client-data line in bold.
A sensible order for a first project: start with donor thank-you letters on a local machine, where facts come from the CRM and every letter is reviewed anyway. Move to case notes only once the review habit is in place and the program director has checked the setup against the program's grant terms.
Frequently asked questions
Can a nonprofit use a cloud AI service under a business agreement for client data?
Sometimes. A HIPAA-covered clinic can sign a business associate agreement with a vendor that offers one. A VAWA-funded program faces a rule about releasing information at all, so a contract with a vendor does not by itself fix the problem. The safest reading for victim services is to keep client data on hardware the program controls.
Can a donor or client ask what the AI system holds about them?
They can ask, and the answer should be easy to give. Because the model only drafts from records the organization already keeps, the data lives in the CRM or case file, not inside the model. Keep prompts and logs on the same retention rules so that answer stays true.
Is AI-assisted grant writing something funders will detect?
Detection is the wrong question. Some funders now ask about AI use or restrict it, so disclose it where asked and keep the model on editing, formatting and compliance checks.
What should happen to AI drafts and logs after a case closes?
Treat them like the case file. If the program destroys client records on a schedule, the model's inputs, outputs and any logs that contain client details should follow the same schedule.
Ready to Modernize Your Business?
Get your AI automation roadmap in minutes, not months.
Analyze Your Workflows →