How to Use AI for Program Outcome Reports With De-Identified Data

Can a nonprofit use AI to write program outcome reports?

Yes, if the AI tool only ever sees de-identified data. The safest workflow aggregates and strips client records inside your case management system, then gives the AI tool only the summary numbers and approved narrative.

That matters because outcome reports are built from the most sensitive records a nonprofit holds: case notes, service dates, health and housing status, and school or court involvement. Federal grant rules, and for some programs HIPAA, require you to safeguard that information, and a prompt to an outside AI tool is one more place it can go.

What rules apply to client data in outcome reports?

For federally funded programs, the Uniform Guidance applies. 2 CFR 200.303(e) requires recipients and subrecipients to "take reasonable cybersecurity and other measures to safeguard information," including protected personally identifiable information.

The Uniform Guidance defines protected PII in 2 CFR 200.1 as an individual's first name or first initial and last name in combination with other information, including Social Security numbers, date and place of birth, and "criminal, medical and financial records, educational transcripts." That list reads like the inputs to a typical outcome report.

Programs that are HIPAA covered entities or business associates have a second, more specific standard for de-identification, covered below. For other nonprofits, it is a useful, published benchmark for what "de-identified" should mean.

What is de-identified data?

De-identified data is data that can no longer reasonably be linked back to a person. NIST SP 800-188, published in September 2023, defines de-identification as "any process of removing the association between a set of identifying data and the data subject."

HIPAA gives two concrete methods in 45 CFR 164.514(b). Under Expert Determination, a qualified expert applies statistical methods and documents that the risk of identifying anyone is very small. Under Safe Harbor, the organization removes 18 types of identifiers and has no actual knowledge that what remains could identify someone. Safe Harbor's list includes:

• Names, including names of relatives and household members
• Street address, city, county, and ZIP code (the first three ZIP digits may stay only if that area has more than 20,000 people)
• All elements of dates except the year, such as intake, exit, and birth dates
• Ages over 89, which must be grouped as 90 or older
• Phone numbers, email addresses, Social Security numbers, case or account numbers, and any other unique identifying number or code

Why isn't removing names enough?

Because small programs make people easy to spot. HHS's de-identification guidance gives examples of "actual knowledge" that defeats Safe Harbor, including a unique occupation or a widely publicized event tied to one person. In a small community program, "the one client who moved from shelter to homeownership in March" may be identifiable to anyone who knows the program.

Free text is the other trap. HHS says identifiers must be removed wherever they appear, including narrative fields, and that narratives can identify someone through context alone. Case note excerpts and success stories are where most re-identification risk sits in an outcome report.

HHS also says there is no universal minimum cell size; the right threshold depends on who will see the data. NIST SP 800-188 adds that tools which only mask personal information may not be enough, and describes governance such as a Disclosure Review Board that approves de-identified data before release.

How do you build a de-identified outcome report with AI, step by step?

Do the de-identification before AI is involved, and treat the AI tool as a writing assistant for numbers you have already cleared. The steps below are a starting point; adjust them to your funders' requirements and your program's data.

• Pull the report data from your case management system as aggregate counts and rates only.
• Remove or generalize every Safe Harbor-style identifier: dates to year or quarter, locations to county or larger, ages to bands.
• Set a minimum count for any reported group, and combine or suppress smaller groups. Write the threshold down so it is applied the same way each time.
• Delete free-text fields entirely. If you want a client story, write it separately, get the client's consent under your own policy, and keep it out of the AI prompt.
• Have a second person review the de-identified table before it goes into any AI tool. A program director who knows the clients is the best check.
• Use an AI account your organization controls, with your data protection settings applied, rather than a personal or free consumer account.
• Ask the AI tool to draft narrative from the approved table only, then check every number in the draft against the source.
• Keep the approved table, the reviewer's sign-off, and the final report together in the grant file, and keep them for as long as your document retention policy requires for that grant.

What should the AI tool itself be allowed to do?

Draft, summarize, and reformat. It should not have connectors to your case management system, shared drives that hold client files, or email inboxes where referrals arrive. Each of those connections can pull identified data into a prompt without anyone deciding to put it there.

Your AI policy should say this in one line: client-level data never goes into an AI tool, and outcome reports are drafted from approved, de-identified tables. If your organization already has an AI policy, add the outcome report workflow as an example of permitted use.

What should a program director do this quarter?

Pick the next funder report due and run it through the workflow above once. Write down the minimum group size you used, who reviewed the table, and which AI account drafted the narrative. That single documented run becomes the template for every report after it, and it gives you a clear answer if a funder asks how client data is protected in your reporting.

How s90 helps

s90 supports nonprofits and social service agencies across Los Angeles and Orange County. We set up AI tools under your organization's accounts, apply guardrails that keep client data out of prompts, and alert you when staff start using new AI apps. If you want help putting a de-identified reporting workflow in place before your next grant cycle, we're happy to talk.