Free course: AI for nonprofit programs

Most program and case-management teams are stretched thin: case managers writing up notes after hours, intake coordinators re-keying the same form twice, program directors squeezing service-design work into whatever time is left. AI will not fix that, but it can take a real bite out of the paperwork and drafting that eats a caseworker’s week.

This course is about direct service and case work, not grantmaking. If your title is “program officer” and you decide which grants to fund, that is a different job with a different course. This one is for case managers, program directors, intake coordinators, and the staff who handle referrals, follow-up, and day-to-day service delivery.

Because this work touches people who are often relying on you and have few other options, safety runs through every section below: what AI can genuinely help with, how to prompt it well, which tools are worth knowing, what can go wrong, and a concrete plan to get started. It takes about 20 minutes, and you do not need to buy anything to follow along.

*️⃣ Note
If you want the general foundations first, how AI actually works, the broad opportunities and risks across an organization, start with our free course on AI for nonprofit organizations. This course assumes you just want to get to work on programs and case management, so it skips the general theory.

What can AI do for nonprofit programs and case management?

The fastest way to find your own uses is to walk through your real service-delivery pipeline (intake, case planning, referral, follow-up, program design) and notice where AI fits. Here is the landscape, grouped by the work you already do. If you want to browse more widely afterwards, our database of the most common AI use cases in nonprofits covers every department.

AI for case notes and documentation

First-draft case notes from a session, with consent, in your agency’s format (SOAP, DAP, or BIRP), reviewed and signed off by the worker before it enters the record.
Case-history synthesis before a meeting, turning months of notes into a short brief of patterns, open actions, and flags, instead of a re-read squeezed in beforehand.
Session and interview transcription, with documented consent, so a searchable text record replaces handwritten notes.
Team meeting and case-conference summarization, turned into decisions and action items so staff can participate instead of scribing.
Policy and funder-guideline summarization, condensing a long document into a short brief of what actually applies to your program.

AI for plain-language materials and translation for participants

Plain-language rewrites of dense eligibility notices, consent forms, and service agreements, checked by a person before any participant sees them.
Translation of program materials into the languages your participants actually speak, with a bilingual reviewer checking the output.
Captioning and subtitling of recorded trainings, orientation videos, and program events.
Reading-level rewrites so materials work for participants who read at a lower level, or for whom English is a second language.

AI for program design and drafting

Logic model and theory-of-change first drafts, from a plain-language description of a program idea, validated by staff against the W.K. Kellogg logic model guide and real evidence before anyone relies on the causal chain.
Data-collection form and survey drafting, which staff then review for clarity and bias.
Funder outcome-report drafting, from raw data and notes, fact-checked line by line before it goes out.

AI for service, eligibility, and benefits research

Evidence-based program-model research, cross-checked against a source like the What Works Clearinghouse before it shapes a program design.
Community resource-landscape scanning, a first-pass map of services, gaps, and partners in your service area, with every listing verified directly.
Regulatory and compliance research, a first summary of licensing or reporting rules that apply to a program, verified against the primary source.
Peer program benchmarking, how comparable organizations structure a similar program, as a starting list to investigate by direct outreach.

AI for participant-facing communication (handle with extra care)

A routine FAQ chatbot answering questions like documents needed or opening hours, with every question outside a short, vetted list handed to a person.
Appointment reminders and follow-up sequences, sent by text or email to reduce no-shows and free front-desk time.
Text-message support for hard-to-reach participants without reliable internet, built for low literacy and low bandwidth.

This is also where the highest stakes in program work sit. In 2023, a chatbot run by the National Eating Disorders Association recommended calorie counting and a daily calorie deficit to people seeking eating-disorder support, and the organization took it offline within days (CBS News, NPR). Never run an unsupervised chatbot for crisis, safety, or clinically sensitive topics, whatever time it would save.

AI for participant-data analysis (the highest-stakes category)

Waitlist and caseload triage support, a tool that suggests a priority category from presenting need, always as a first-pass suggestion a person reviews, never a decision.
Participant feedback theme and sentiment analysis, grouping open-text survey responses into themes, paired with a human read of a real sample, since AI can miss cultural context in feedback from diverse communities.
AI-assisted document and ID processing at intake, extracting fields from uploaded documents instead of hand-keying them, with anything low-confidence flagged for human review.

Predictive risk scoring deserves a specific warning. The best-documented example, Allegheny County’s Family Screening Tool for child-welfare referrals, showed a pattern of disproportionately flagging Black children for investigation and drew scrutiny from the U.S. Department of Justice’s Civil Rights Division (PBS NewsHour), despite being built with university partners. Treat any risk or triage score as one input a trained human reviews, never an automatic decision, and do not build one in-house.

New projects AI makes possible for small program teams

Everything above is work you already do. The bigger opportunity is work you currently skip because it was never affordable.

Give every case worker a same-day case-history brief before each meeting, instead of a re-read squeezed in beforehand or skipped entirely.
Answer every routine participant question instantly, instead of letting it queue at a busy front desk.
Offer key materials in every language your participants speak, not only the one or two you can staff for.
Catch a program’s declining engagement early, while you can still act on it, instead of first seeing it in a funder report months later.

⬆️ Important
Before you hand any task to AI, run a quick gut-check: what happens if it is wrong, is a vulnerable person on the receiving end, and does it need deep context about a real person or a real human relationship to do well? When a vulnerable person is directly on the receiving end (a participant in crisis, someone whose benefits depend on the answer), the bar is much higher, and for some uses the answer is simply no.

AI agents for nonprofit programs

Everything above treats AI as an assistant: you ask, it answers, you decide what to keep. Agentic AI goes a step further. You give an agent a goal, and it plans and carries out the steps itself across your files, your apps, and the web, pausing for your approval on the actions that matter. Instead of prompting turn by turn, you delegate a whole task and review the result.

Part of what makes agents more capable is that they can do more than write text back to you: they can connect to your other tools (through connectors, the emerging MCP standard, or APIs), create and edit real files, and run saved routines called skills. The most capable general-purpose options right now are knowledge-work agents like Claude Cowork or ChatGPT Work. Most major AI vendors are shipping one, so check what your current tools already include before paying for a new seat.

This is a fast-moving space, so treat the list below as a few examples, not the full set. Keep every one of these on non-sensitive, staff-facing work that a person reviews before it goes anywhere near a participant’s case:

Turn approved session notes into a first-draft case-summary pack. Hand an agent a folder of already-reviewed, de-identified notes and get back a structured summary, instead of assembling one by hand.
Build a plain-language service-info kit from your policies. Point an agent at your program manuals and eligibility rules and get back a first-draft set of plain-language explainers a person then checks for accuracy.
Get a scheduled funder-reporting brief from your outcomes data. An agent gathers last month’s outcome numbers into a short draft summary on a schedule, with every figure traced to its source file.
Connect to a resource directory to draft referral options. An agent searches a maintained resource directory and drafts a short list of candidate referrals for a worker to verify and choose from, never a referral sent to a participant on its own.

⚠️ Warning
Because an agent acts on its own, it carries risks a chatbot does not. The main ones to know, among others:
It acts, so mistakes ship. Never let an agent make an eligibility, triage, or referral decision, or touch identifiable participant data, without a human controlling the outcome. Keep an approval gate on anything that sends, posts, or acts on a case, and start any agent in read-only mode.
Hidden instructions can hijack it (prompt injection). A web page or document the agent reads can contain instructions it may follow without you realizing. Give it only the access a task needs, nothing more.
Connected tools and add-ons widen what can go wrong. Every tool, connector, or skill you grant an agent is more it can reach and more code you are trusting, so connect only what a task needs and install skills or plugins only from sources you trust.
Wider access means wider exposure of sensitive case data. An agent reaching into your case-management system, drive, or inbox can touch far more sensitive participant data than a single pasted prompt, so keep it out of identifiable case data until you have real experience and firm guardrails in place.

Our fuller guide to AI agents for nonprofits covers what they can do today and the guardrails to set, and agent skills, plugins, and connectors explains how to extend one safely.

How to write better AI prompts for nonprofit programs

This is the part most people skip, and it is the difference between “AI writes a generic case note” and “AI saves a case worker real time”. These habits matter far more than which tool you pick.

Build a reusable program context pack

Write this once and reuse it forever: your mission and what your program does, who you serve and in what circumstances, your house style (reading level, tone, anything you always avoid saying), and a standing rule to never invent eligibility rules, dollar amounts, deadlines, or resources, and to flag anything a human must verify.

Prompting techniques that actually improve output

Show, do not describe. Pasting an anonymized case note you were happy with teaches the AI your format better than any description.
Ask for options, not an answer. “Give me three plain-language versions of this notice at a sixth-grade reading level” beats one guess.
Say what you do not want. “Do not change any deadline, dollar amount, or eligibility rule from the original” is a constraint that sharply improves accuracy.
Give it a role and an audience. “You are helping a case manager write for participants who may read at a sixth-grade level” produces noticeably different work than an unframed request.
Never guess about a specific person’s case. Describe the situation in general terms and let a human fill in what is actually true for that participant.
Make it critique itself. Ask the AI to review its own draft against a checklist (is anything unclear, does it invent a detail not in the source) before you read it.
Save what works. A prompt that produced a good case-note draft is an asset. Keep a shared document of working prompts so your team stops reinventing them.

If you want to go further on this, our guide to prompt and context engineering covers the techniques above in much more depth.

✅ Example

A reusable prompt skeleton for program and case work:

# ABOUT OUR PROGRAM
> Mission and program: [what the program does and for whom]
> Who we serve: [population, common languages, common situations]
> House style: [reading level, tone, anything we always avoid saying]

# REQUEST
[e.g. Rewrite the eligibility notice below in plain language at a sixth-grade
reading level. Explain what it says, what the person needs to do, and by when.]

# SOURCE MATERIAL
> [paste the approved text or notes. No identifying participant details.]

# CONSTRAINTS
Do not change any deadline, dollar amount, or eligibility rule from the
original. Do not invent any detail not in the source. Flag anything unclear
so a human can check it.

The best AI tools for nonprofit programs (and how to get them cheaper)

You do not need to buy anything to start. A free general assistant covers most of what is above. Every tool below is an example within its category, not a recommendation to adopt: the field moves fast, and the practice matters more than the vendor.

General assistants for drafting, summarizing, and research: ChatGPT, Claude, or Gemini. Start here.
Case-management platforms with AI features: CaseWorthy, Apricot by Bonterra, or Casebook, if your existing platform has added an AI layer worth turning on.
Session note-taking and transcription: Mentalyc is purpose-built for clinical and case documentation; general meeting notetakers like Otter.ai or Fireflies.ai suit internal meetings better than direct client sessions, since they often lack a signed data agreement for clinical content on their standard plans.
Referral matching and resource directories: Findhelp or your local 211 operation, built on a directory your organization or a trusted network actually maintains.
Accessibility and translation: DeepL or Google Translate for first-pass translation, Amara for captioning, both followed by a bilingual or human review.
Client communication and chatbots: LiveChatAI or Twilio.org‘s nonprofit program, always with a visible path to a real person.
Scheduling and waitlists: Calendly or WaitWell for appointment reminders and no-show reduction.

For a fuller, regularly reviewed list, see our AI tools for program management guide, or the broader best AI tools for nonprofit organizations.

*️⃣ Note
Before you pay for anything, check the nonprofit programs. TechSoup offers discounted and donated software to eligible organizations, and Google for Nonprofits includes Google Workspace, free translation, and other tools at no cost for eligible organizations. Many AI tools also have nonprofit pricing that is not advertised on their pricing page, so it is worth asking.

AI risks for nonprofit programs: what can go wrong and how to prevent it

None of this is a reason to avoid AI. It is the reason to set a few guardrails first, because in program work a mistake can reach a person who is relying on you, not just your budget or your reputation. Our database of key AI risks for nonprofits covers each of these in more detail, with mitigation strategies.

Client data exposure through consumer AI tools. Case notes, health details, immigration status, and family-safety information are the input here, not the exception, so privacy is not optional. Never put identifiable participant data into a free or personal-tier AI account.
A confident but harmful answer reaching a participant directly, as the NEDA chatbot case above shows. Restrict any client-facing bot to a narrow, tested set of vetted answers with a visible human escalation path.
Over-reliance on AI-drafted case notes and plans without real review. Newer staff especially can sign off on a weak draft without catching the error. Require a documented human sign-off before anything AI-drafted enters the official record.
Vendor data-handling and retention practices you never vetted. A 2020 ransomware breach at a major nonprofit-software vendor exposed data belonging to constituents of more than 13,000 organizations and led to a $49.5 million multistate settlement (BleepingComputer). Ask any vendor in writing where data is stored and what happens if you cancel.
Algorithmic bias in predictive risk-scoring and triage, covered above. Treat any score as one input a human reviews, never a decision, and demand an independent bias audit before adopting one.
A stale referral or resource directory. A directory only a general model’s training data backs will confidently send someone to a shelter that closed. Only connect a client-facing tool to a directory your organization or a trusted network actively maintains.
Recording or transcribing a session without real consent. Never record a client session without documented, explicit informed consent that names the AI tool.
Erosion of the human relationship at the center of case work. Reserve AI for behind-the-scenes support (drafting, summarizing, scheduling) rather than the emotionally significant parts of a relationship.
Undisclosed AI use eroding trust. Tell participants where AI is used in their own case, and give them a real, no-penalty way to decline it.

⚠️ Warning
One of the most useful things you can do early is write a short no-go list: the handful of things your organization decides in advance never to do with AI. A reasonable starting list for a program team:
Never let an eligibility, triage, or referral tool make a final decision with no human review.
Never deploy a general-purpose or unsupervised chatbot for crisis, safety, or clinically sensitive topics.
Never record or transcribe a client session without documented, explicit informed consent that names the AI tool.
Never treat an AI risk or triage score as anything more than one input a trained human reviews.
Never let an AI-drafted case note or plan enter the official record without a documented human sign-off.

Next steps: how to start using AI in your programs team

Reading about AI changes nothing on its own. Here is a realistic order of work, from this week to the next few months.

Pick one repetitive, low-stakes task and try it this week. A plain-language rewrite of a notice, or turning a recorded team meeting into a summary, is the usual best starting point.
Build your context pack (mission, who you serve, house style, never-invent rules). Save it somewhere shared. Every future prompt gets better from this one document.
Write your data rule: the short list of what nobody pastes into a consumer AI tool. Case notes, health details, immigration status, and family-safety information are the usual entries.
Agree a short no-go list with whoever owns program decisions, using the list above as a starting point.
Name who reviews AI output before anything reaches a participant. This one decision is what turns “we should check it” into something that actually happens.
Run a two-to-four-week pilot on that first task. Write down a real baseline number and what “working” means before you start, and track edit time, the amount of correction each output needs, since it is the metric that matters most here. Our checklist for new AI pilots and projects walks through what to decide before you start.
In your first month, expand to a handful of uses rather than everything at once, and keep the low-risk, back-office ones (intake automation, plain-language rewrites, meeting summaries) ahead of anything participant-facing.
In months two and three, write a one-page AI policy covering your data rule, your no-go list, who reviews what, and how you tell participants when AI is involved in their case. Our AI policy template gives you a ready-to-adapt starting point.
Bring your colleagues in. Show two or three concrete before-and-after examples from your own pilot, including a frontline case worker’s honest reaction. That persuades far better than an article about AI.
Set a review date every few months. Tools, prices, and capabilities change quickly, so a use that was too risky to try may be worth revisiting, and your policy will need updating.
Go deeper when you are ready to do this properly. Our full AI course for nonprofit programs covers the complete use catalog, tool selection, the full risk and disclosure framework, team roles and buy-in, and how to run and measure a pilot safely. We also have equivalent courses for communications, fundraising, operations, HR, and leadership.
Get help if you would rather not do it alone. We offer AI consulting for nonprofits, including a free first consultation to talk through your situation and where AI would genuinely help, and where it should stay off the table for now.

Whatever you do next, the first step matters more than the plan. Pick one low-stakes task, try it, and decide with evidence rather than opinion.