How to use AI for grant writing without leaking donor data

What donor and client data ends up in a grant draft?

A grant draft can hold more donor and client data than most small teams expect. A single proposal can pull from the donor database, the case management system, the finance system and staff memory, and all of it lands in one working document that someone is tempted to paste into an AI grant writing tool.

Typical examples include a list of major donors for a matching gift section, gift amounts and pledge dates for a sustainability narrative, a client success story with a first name, age and neighborhood, an outcomes spreadsheet exported with client IDs still attached, and a budget that shows individual staff salaries. Each piece looks harmless inside the draft. Pasted into a personal AI account, it becomes information the organization no longer controls.

Which California rules apply when a nonprofit uses AI?

California's customer records law, including its reasonable security rule, reaches most charities, while the CCPA usually does not. The CCPA's definition of a covered business in Civil Code section 1798.140(d) is limited, with narrow exceptions, to entities organized or operated for profit, so most charities fall outside it.

The customer records chapter is different. Section 1798.80(a) defines a business as a group "however organized," whether or not it operates at a profit. That pulls nonprofits into section 1798.81.5, which requires a business that owns, licenses or maintains personal information about California residents to keep reasonable security procedures and practices appropriate to the nature of that information.

Section 1798.82 adds the breach notice duty. Notice is required when unencrypted personal information is, or is reasonably believed to have been, acquired by an unauthorized person, and the statute now sets a 30 calendar day deadline after discovery. The defined data elements are specific: a name combined with a Social Security number, a government ID number, a financial account number with its access code, medical information, health insurance information, biometric data or genetic data, or an online login with its password. A donor's name and gift amount alone are not on that list, but a client story that pairs a name with a diagnosis can be. Whether a particular AI upload counts as an unauthorized acquisition depends on your facts, and the counsel review list below covers it. Our pieces on AI exposure and the California breach law and responding when donor or client data reaches an AI tool walk through that analysis.

What do federal grants add?

Federal awards add a direct safeguarding duty and, at some agencies, rules about AI authorship. For awards issued on or after October 1, 2024 under the revised Uniform Guidance, 2 CFR 200.303(e) requires recipients and subrecipients to take reasonable cybersecurity and other measures to safeguard information, including protected personally identifiable information and information the agency or pass-through entity designates as sensitive. 2 CFR 200.1 defines protected PII. Client data from a federally funded program sits squarely inside that duty, including when staff reuse it to write the next proposal.

Some funders now regulate AI use in the application itself. In NIH Notice NOT-OD-25-132, released July 17, 2025 and effective for applications submitted for the September 25, 2025 receipt date and later, NIH said it will not consider applications substantially developed by AI, or with sections substantially developed by AI, to be the original ideas of applicants. The notice says AI detected after an award may be referred to the Office of Research Integrity and can lead to disallowed costs, withheld future awards, suspension or termination. Most community nonprofits never apply to NIH, but the notice shows where funders are heading. Read the AI and data terms in every request for proposals and grant agreement before the writing starts.

Which AI grant writing tasks are safe for a small team?

Using AI for grant writing carries the least risk when the task starts from public, invented or de-identified material. The danger comes from what goes into the prompt, so the safest uses never require pasting donor or client records.

• Outlining a response from the funder's published request for proposals and scoring rubric.
• Tightening a need statement your team wrote from public sources such as census data or published research.
• Cutting a section to fit a character limit or reading level.
• Checking a draft against the funder's stated priorities for gaps.
• Summarizing your organization's own published annual report or website.

Higher risk tasks need a different approach: donor lists for match reporting, client stories with identifying details, case notes, outcome exports with client IDs, and salary-level budget detail. Do that work in the systems where the data already lives, or de-identify it first.

How do you tell client stories without identifying clients?

Strip identifiers before the text reaches any AI tool, and keep the client's consent to share the story on file. Replace names, turn exact ages into ranges, remove schools, neighborhoods, employers and dates, and combine details from several clients when the story allows it. A reader should not be able to recognize the person, and neither should someone who knows them.

The same approach works for outcome data. Report counts and percentages, suppress small cells that could single someone out, and keep the identified export inside the case management system. Our guide to AI and de-identified outcome reports covers the mechanics. Programs that handle health information under HIPAA have their own de-identification standard and business associate rules, covered in our piece on HIPAA and AI for nonprofit health programs.

What kind of AI account should grant writers use?

Grant writers should use an account the organization owns, signed in with organization credentials, under business data terms. A personal ChatGPT account or other free plan leaves the nonprofit unable to see what was shared, unable to delete it when a grant writer or volunteer leaves, and dependent on consumer terms that may allow the provider to use conversations to improve its models. Check those terms for any tool before approving it.

An organization account also gives you sign-in logs, the ability to remove access the day someone departs, and a single place to apply data rules. Our pieces on approving free AI tools without giving away donor data and giving staff AI on a nonprofit budget compare the practical options.

What can a small team do this quarter?

Start with a few decisions that fit on one page and do not require new software.

• Pick one approved AI tool for grant work and require staff, contractors and volunteers to use it through organization accounts.
• Write down what never goes into a prompt: donor names with gift amounts, client names, case notes, Social Security numbers, health information and staff salaries. Add it to your one-page AI policy.
• Build a de-identified story bank and a summary outcomes sheet that grant writers can use freely.
• Read the AI, confidentiality and data security terms in each active grant agreement and upcoming request for proposals.
• Review sign-in records for AI tools nobody approved, using the steps in our shadow AI detection guide.
• Keep a short note in each proposal file on how AI was used, so you can answer if a funder asks. For the security questions funders put on applications, see our piece on grant application cybersecurity questions.

What should your counsel review before staff use AI on grant work?

Run this review before your team adopts an AI tool for grant work, and again when a new grant agreement or request for proposals arrives. Bring these decisions to your own counsel:

• What reasonable security means for your AI setup under Civil Code section 1798.81.5. Bring your list of approved AI tools, their business data terms and a summary of the donor and client data staff handle.
• Whether a past upload of client data to a personal AI account triggers notice under Civil Code section 1798.82 and its 30 calendar day deadline. Bring what was entered, the account type and the provider's terms at the time.
• Which information your federal awards or pass-through entities designate as sensitive under 2 CFR 200.303(e). Bring each award agreement and any subaward terms.
• Whether a funder's AI rules, such as NIH Notice NOT-OD-25-132, limit how your proposals may be drafted. Bring the request for proposals, the grant agreement and your notes on how AI was used.

How does s90 help nonprofits use AI for grant work?

s90 sets up organization AI accounts with single sign-on, data loss prevention rules that flag donor and client data before it reaches an unapproved AI tool, and alerts when staff sign up for new AI apps. If your team wants to use AI on proposals without guessing where the data goes, we are happy to walk through what that looks like for an organization your size.