Data Foundations & Responsible Data Handling

Why This Module Exists

Most non-profits are sitting on data they do not fully understand. Beneficiary records live in a case-management tool, donor history lives in a separate fundraising database, intake forms live in a filing cabinet or a shared inbox, and outcome numbers live in a spreadsheet that only one staff member knows how to update. None of it is labeled by how sensitive it is. When AI arrives, this is the soil it grows in — and bad soil grows bad outcomes.

The risk is not theoretical. The fastest-growing data risk for non-profits right now is "shadow AI" — staff and volunteers quietly pasting internal documents, case notes, or donor lists into free public AI tools to save time, with no approval and no record. The organization loses visibility, control, and the ability to audit what happened (ASU Lodestar, 2026). Public AI tools can retain what you type and use it to train future models. Real organizations have already leaked confidential material this way: Samsung engineers pasted source code into ChatGPT; the interim chief of the U.S. cyber-defense agency (CISA) uploaded sensitive documents to a public version of ChatGPT despite it being blocked for colleagues (IT Pro, 2025). For a non-profit, the equivalent mistake is an immigration status, a mental-health note, or a donor's financial details — and the people harmed are the people you exist to protect.

There is also an opportunity cost. AI can only help analyze outcomes, draft reports, or spot patterns if your data is "AI-ready" — meaning organized, consistent, and complete enough to trust. Most non-profit data is not there yet. The work in this module is the unglamorous foundation that makes everything in Modules 6, 7, and 8 actually possible.

Learning Objectives

By the end of this module, participants will be able to:

1. Build a simple data inventory that lists every place the organization holds data (beneficiary, donor, operational) and who owns each one.
2. Classify each data set by sensitivity (Public / Internal / Confidential / Highly Sensitive) and explain what "Highly Sensitive" means for the populations they serve.
3. State and apply the cardinal rule — never paste sensitive beneficiary or donor data into public AI tools — and recognize on sight what counts as sensitive.
4. Explain when consent, data-sharing agreements, and retention rules apply to AI use, and whether existing consent language covers AI.
5. Judge whether a data set is "AI-ready" (structured, consistent, complete, single source of truth) or not — and name the one fix that would help most.
6. Run a basic data-hygiene pass on a real system (deduplicate, standardize fields, agree on metric definitions) using their CRM or spreadsheet.
7. Recognize whether GDPR or HIPAA applies to their organization, in plain terms, without needing a lawyer in the room.
8. Use the three takeaway artifacts — the inventory & classification template, the "what data can/can't go into AI" rule card, and the consent & data-sharing checklist.

Session Agenda

Part A — Leads workshop + clinic (half-day, ~3.5 hrs): A6 IT/Data lead + A5 Operations

Part B — All-staff segment (45 min): A3 program staff + everyone

Core Content — Parts 1–4

Part 1 — Know What You Hold: Data Inventory & Mapping

You cannot protect, classify, or feed to AI what you cannot see. A data inventory is simply a written list of every place your organization holds data, what kind it is, who owns it, where it lives, and how long you keep it. It is the single most useful document this module produces — and most non-profits have never made one.

Walk your organization function by function and write down every data source: intake forms, case notes, the donor database, the email newsletter list, the volunteer roster, grant-reporting spreadsheets, the shared inbox, paper files in a cabinet. For each, capture where it lives (which system, cloud or on-premises or paper), who owns it (a named person, not "the team"), what's in it, and how sensitive it is. This is exactly the practice enterprise data teams call an "AI-ready data inventory" — you are doing the same thing, just lean (BigID, 2025; NonProfit PRO, 2025).

   DATA INVENTORY — what feeds the map
   ┌─────────────┬──────────────┬──────────────┬───────────────┐
   │ Source      │ Where it     │ Owner         │ Sensitivity   │
   │ (what)      │ lives        │ (named)       │ (P/I/C/HS)    │
   ├─────────────┼──────────────┼──────────────┼───────────────┤
   │ Case notes  │ Apricot      │ Program Dir.  │ Highly Sens.  │
   │ Donor gifts │ Raiser's Edge│ Dev. Director │ Confidential  │
   │ Newsletter  │ Mailchimp    │ Comms lead    │ Internal      │
   │ Annual rpt  │ Website      │ Comms lead    │ Public        │
   └─────────────┴──────────────┴──────────────┴───────────────┘

4D tie-in — Delegation: You cannot decide what to safely hand to AI (Delegation) until you know what each data set contains and how sensitive it is. The inventory is the precondition for every other decision in the kit.

Part 2 — Label the Risk: Sensitivity Classification, Ownership, Access & Retention

Once you can see your data, sort it by how much harm a leak would cause. A simple four-level scale is enough for almost every non-profit:

The "Highly Sensitive" row is the heart of this module. These are the categories that, if exposed, can cost someone their housing, their safety, their immigration case, or their dignity. Vera Solutions' responsible-AI principles put Privacy & Data Protection at the center: collect only what you need, anonymize or encrypt sensitive data, and follow the regulations that apply to you (Vera Solutions, 2024). Data minimization — collecting only what is absolutely necessary — is the cheapest protection you have (ASU Lodestar, 2025).

For each data set, also write down two more things:

• Access — who can see it, on a strict need-to-know basis with role-based permissions, reviewed periodically (ASU Lodestar, 2025).
• Retention — how long you keep it before secure deletion. AI-ready organizations enforce retention and keep only what is necessary, accurate, and current (Transcend, 2025; Striim, 2025).

   CLASSIFY → then apply the right protection
   Public ───────► share freely
   Internal ─────► internal access; default-deny external
   Confidential ─► role-based access · encryption · NEVER public AI
   Highly Sens. ─► strict need-to-know · consent check · NEVER any
                    external AI · elevated protection · audit

Do / Don't

Part 3 — The Cardinal Rule, Consent, and the Agreements Around Your Data

The cardinal rule, stated plainly: Never paste sensitive beneficiary or donor data into public AI tools. Candid frames it as the test that anyone can apply: do not enter personally identifying or confidential information, legal documents, passwords, or anything you wouldn't paste into a public website (Candid, 2025). Public AI tools can store your input and use it to train future models, which means you can lose ownership and control of that data the moment you hit send (ASU Lodestar, 2026). This is the one rule every staff member must know cold — it is the spine of the all-staff rule card (Part B and the Templates section).

Consent — does your existing language cover AI? Most non-profit consent forms were written before generative AI existed, so they almost never mention it. Good consent is freely given, informed, specific, documented, and obtained in advance — and you should explain in plain terms what data you collect, why, and how it is used, offer opt-outs for non-essential collection, and renew consent periodically (GDPR principles; ASU Lodestar, 2025). Before any AI touches beneficiary data, check: does the consent the person signed actually cover this use? If not, you need new language or you do not proceed. (Module 7 covers trauma-informed, culturally appropriate consent in depth.)

Does GDPR or HIPAA apply to you? (plain-language version)

• GDPR applies to any organization that holds personal data about people in the EU/EEA — regardless of where your office is. International programs, EU donors, or EU beneficiaries pull you in (Usercentrics, 2024; Foundation Group, 2025).
• HIPAA applies only if you are a "covered entity" — a health-care provider, health plan, or their business associate handling protected health information. A health clinic or hospital foundation is covered; a general social-services charity usually is not — but if you hold health data, treat it as Highly Sensitive regardless (Foundation Group, 2025).
• U.S. state laws are multiplying fast: by mid-2025, 13 states had comprehensive privacy laws, with more taking effect through 2025 (501c3.org, 2025). You do not need to memorize them — you need a named person responsible for staying informed (ASU Lodestar, 2025).

Data-sharing agreements with funders, partners, and contractors. Whenever data leaves your walls — to a government funder, a partner agency, a contractor, or an AI vendor — there should be a written agreement that spells out: who plays which role (data "controller" vs. "processor"), the exact purpose and the exact data shared, security requirements (encryption, access controls), breach-notification timelines, sub-processing rules, audit rights, and what happens to the data at the end (deletion or return) (ContractsCounsel, 2024; GDPR.eu DPA template). Vet every third-party provider's own data practices and put protective clauses in the contract (ASU Lodestar, 2025). For AI vendors specifically, the make-or-break clause is "this vendor will not train on our data" — that single line is what separates an enterprise tool from a consumer one (covered in depth in Module 3).

4D tie-in — Diligence: Consent checks, signed agreements, and the cardinal rule are Diligence in practice — verifying and taking ownership before data moves, not apologizing after.

Part 4 — "AI-Ready Data" and the Data-Hygiene Clinic

What "AI-ready" actually means. Data is AI-ready when it is structured (in fields, not buried in free-text or paper), consistent (the same thing is recorded the same way every time), complete (few blanks, minimal duplicates), and pulled toward a single source of truth rather than scattered across disconnected systems (Alteryx; Transcend; NonProfit PRO, 2025). For a non-profit, the practical version is: one place where each fact lives, recorded the same way each time.

NonProfit PRO's five practical steps to an AI-ready data foundation map cleanly onto lean teams (NonProfit PRO, 2025):

   1. ASSESS      → where does data live? siloed or unified? who owns quality?
   2. PRIORITIZE  → pick ONE use case / pilot, not everything at once
   3. CONSOLIDATE → reduce silos toward a single source of truth
   4. STANDARDIZE → clean: dedupe, fix fields, agree shared metric definitions
   5. GOVERN      → access controls, privacy rules, named ownership

The clinic — data hygiene on a real system. Whatever you use — Salesforce Nonprofit Cloud / NPSP, Blackbaud Raiser's Edge NXT, Bonterra Apricot, or just spreadsheets — the same problems show up: duplicate records, names in ALL CAPS, misspellings, inconsistent field values, and the same metric defined three different ways by three teams (Omatic, 2024). Apricot is built around the operational case record (intake, case notes, service history); NPSP organizes data into accounts, contacts, opportunities, and campaigns; Raiser's Edge NXT includes built-in validation tools. In every one, a basic hygiene pass means: merge duplicates, standardize key fields, fill or flag critical blanks, and write down one agreed definition for each outcome metric so "served" means the same thing across every project (Omatic, 2024; PairSoft).

Why this matters for M&E: AI-assisted impact analysis (Module 8) only works if your outcome data is tracked consistently across cohorts and time. Vera Solutions built Salesforce-based M&E tracking outcomes for millions of beneficiaries precisely because consistent, structured outcome data is the prerequisite for any analysis — by hand or by AI (Vera Solutions). A clean data catalog with shared metric definitions is what lets every team "reference the same shared understanding" (NonProfit PRO, 2025).

4D tie-in — Description: Clean, consistently-labeled data is what lets you describe a task to AI precisely. Garbage in, confident-garbage out — Discernment can only catch so much if the underlying data is a mess.

Hands-On Exercises

Exercise 1 — Build Your Data Inventory & Classification (Leads, ~35 min)

Instructions:

1. Using the template in the next section, list every place your organization holds data. Aim for 10–15 sources in the first pass.
2. For each, fill in: where it lives, named owner, what's in it, sensitivity level (P/I/C/HS), who can access it, and retention period.
3. Flag anything you can't answer as "unknown — to investigate."
4. Circle every row marked Confidential or Highly Sensitive. These are your "never goes near public AI" rows.

Facilitator answer key (sample):

Look-fors: every row has a named human owner; beneficiary data defaults to Highly Sensitive; at least one "unknown" is honestly flagged.

Exercise 2 — Data-Hygiene Clinic on a Live System (Leads, ~30 min)

Instructions:

1. Open your CRM or main outcome spreadsheet (a copy/sandbox, never production for a first pass).
2. Find and merge duplicate records for at least 3 contacts.
3. Standardize one messy field (e.g., fix ALL-CAPS names, or make "State" consistent: "AZ" everywhere, not "Arizona/Ariz./AZ").
4. Pick one outcome metric (e.g., "people served") and write one agreed definition. Note where teams currently disagree.
5. List the single biggest data-hygiene problem you found and assign an owner + date to fix it.

Facilitator sample solution: Common findings include the same donor entered twice with different spellings, "people served" counted as enrollments by one team and unique individuals by another, and date fields in three formats. The win is not finishing — it's that the team now sees the problem clearly and has agreed one definition and one owner.

Exercise 3 — The Sort-the-Data Game (All-staff, ~12 min)

Instructions: Read each item aloud; participants call out the sensitivity level and whether it can go into a public AI tool.

Facilitator note: The teaching moment is the easy ones (annual report) versus the hard ones (case notes). Reinforce: if it's about a real person we serve or a real donor, assume it cannot go into a public AI tool.

Templates & Takeaway Artifacts

Artifact 1 — Data Inventory & Sensitivity Classification Template

Rule of thumb baked into the last column: Public/Internal = usually yes (still no personal data about real people). Confidential/Highly Sensitive = no.

Artifact 2 — One-Page Rule Card: What Data Can & Can't Go Into AI Tools

 ┌─────────────────────────────────────────────────────────────┐
 │  THE ONE RULE: If you wouldn't paste it onto a public        │
 │  website, don't paste it into a public AI tool.              │
 ├─────────────────────────────────────────────────────────────┤
 │  ✅ OK TO USE WITH AI (low-risk, verify the output)          │
 │   • Already-public text (annual report, web copy)            │
 │   • General drafts with NO real names or personal details    │
 │   • Brainstorming, summarizing public/internal docs          │
 │   • Generic templates: job descriptions, policies, agendas   │
 ├─────────────────────────────────────────────────────────────┤
 │  ⛔ NEVER PUT INTO A PUBLIC AI TOOL                          │
 │   • Beneficiary case notes or anything about a real client   │
 │   • Immigration status · health/mental-health data           │
 │   • Financial distress, abuse, or safety details             │
 │   • Children's / minors' data                                │
 │   • Donor giving history, contact details, financials        │
 │   • Passwords, legal documents, contracts, board minutes     │
 ├─────────────────────────────────────────────────────────────┤
 │  WHEN UNSURE → don't paste it. Ask: __________ (data lead)   │
 │  Approved AI tools list lives at: ___________________________│
 └─────────────────────────────────────────────────────────────┘

(Sources: Candid 2025; ASU Lodestar 2026. Note: an approved enterprise tool with a "no training on our data" contract changes what's allowed — see Module 3.)

Artifact 3 — Consent & Data-Sharing Checklist

Consent

• Beneficiaries are told, in plain language, what we collect, why, and how it's used.
• Consent was freely given, specific, and documented (not buried in fine print).
• Existing consent explicitly covers AI use of the data — or we have new language.
• There is an opt-out for any non-essential collection.
• We collect only what we actually need (data minimization).

Data-sharing with funders / partners / contractors / AI vendors

• A written data-sharing or data-processing agreement is signed before data moves.
• It names roles (controller / processor), the exact purpose, and the exact data.
• It requires encryption, access controls, and a breach-notification timeline.
• It covers sub-processing, audit rights, and deletion/return at the end.
• For AI vendors: the contract says they will not train on our data.
• We have vetted the third party's own data-protection practices.

Regulation check (plain)

• We've noted whether GDPR applies (any EU donors/beneficiaries/staff).
• We've noted whether HIPAA applies (are we a health "covered entity"?).
• One named person is responsible for tracking changing state/national privacy laws.

Knowledge Check

1. (MCQ) A staff member wants to use ChatGPT to tidy up a client case note that mentions immigration status. What's the correct action?
   1. Remove the client's name first, then paste it
   2. Paste it — it's just grammar help
   3. Do not paste it into any public AI tool at all
   4. Ask the client for permission, then paste it
2. (Short answer) State the cardinal rule of this module in one sentence.
3. (MCQ) Which of these is correctly classified as Highly Sensitive?
   1. The published annual report
   2. A beneficiary's mental-health intake scores
   3. The staff meeting agenda
   4. A generic volunteer sign-up form
4. (Short answer) Name three things that make a data set "AI-ready."
5. (MCQ) Your organization runs an international program with beneficiaries in the EU. Which is true?
   1. GDPR doesn't apply because you're not based in the EU
   2. GDPR applies because you hold data about people in the EU
   3. Only HIPAA applies
   4. No privacy law applies to charities
6. (Short answer) Before sharing beneficiary data with a partner agency, what two things must you confirm?
7. (MCQ) What is "shadow AI"?
   1. An AI that runs only at night to save energy
   2. Staff using unapproved AI tools without oversight, often with sensitive data
   3. A backup copy of your AI tool
   4. AI used by competitors
8. (Short answer) In a data inventory, why does every data set need a named owner rather than "the team"?

Answer key

1. c — Confidential/Highly Sensitive data never goes into public AI tools; removing the name doesn't make it safe.
2. "Never paste sensitive beneficiary or donor data into public AI tools — if you wouldn't post it on a public website, don't paste it into AI."
3. b — health/mental-health data is Highly Sensitive.
4. Any three of: structured (in fields), consistent (recorded the same way), complete (few blanks/duplicates), single source of truth, properly labeled.
5. b — GDPR applies based on whose data you hold, not where you're based.
6. (1) A signed data-sharing agreement exists; (2) the beneficiary's consent covers that sharing.
7. b
8. Because ownership is what makes a rule enforceable and auditable; "the team" means no one is actually accountable, and access can't be reviewed when staff leave.

Facilitator Guide

Prep checklist

• Print the rule card (Artifact 2) for every all-staff participant.
• Ask leads to bring a copy/export of their main systems (CRM, outcome sheet) — never live production for the clinic.
• Have the inventory template open in a shared spreadsheet the group can edit together.
• Pre-fill one or two inventory rows as worked examples.
• Confirm who your "ask-when-unsure" data lead is, and fill it on the rule card.

Free/low-cost materials needed: A shared spreadsheet (Google Sheets / Excel), a printer for the rule card, and the open sector resources in Further Resources. No paid tooling required.

Timing: The leads workshop runs tight at 3.5 hrs — protect the clinic (Part 4); it's where behavior change happens. The all-staff segment must stay at 45 minutes; resist turning it into a lecture on regulation.

Common pitfalls:

• Over-engineering the inventory. Stop the group at "good enough, 10–15 rows." Perfection kills completion.
• Buying a tool to avoid doing the work. Redirect: the spreadsheet first, the platform later (if ever).
• Treating classification as IT's job. Program staff know best what's sensitive about beneficiary data — keep them in the room.
• Skipping the consent-covers-AI check because "we already have consent forms." Old forms rarely mention AI.

Discussion prompts:

• "Name one data set where you genuinely don't know who owns it." (Surfaces gaps fast.)
• "What's one thing a colleague might paste into AI without realizing the risk?"
• "Where do our teams define the same metric differently?"

Tailoring by audience:

• A6 IT/Data lead (or outsourced): Lead the clinic, own access/retention and vendor agreements. If support is outsourced, invite the contractor to this session and assign them the agreement-review actions.
• A5 Operations: Own the inventory upkeep and the quarterly hygiene pass.
• A3 program staff (all-staff segment): Focus only on the rule card and the sort-the-data game. They do not need the full governance detail — they need the one rule, cold.

Addressing fear and resistance (a real non-profit barrier): Two fears show up here. First, "I'll get in trouble for using AI wrong." Reframe: the rule card exists so no one has to guess — following it protects you as much as the people you serve. Co-create the "ask-when-unsure" path so staff feel supported, not policed. Second, "this is a lot of work we don't have time for." Reframe: a 45-minute inventory and a clear rule card prevent the kind of incident that costs days of cleanup and donor trust. Frame governance as the thing that lets the org safely use AI — not a barrier to it.

Outcome Scorecard

Measure with Kirkpatrick, weighting L2–L4 (knowing > feeling; doing > knowing):

Further Resources & Sources

Named sector frameworks & resources

• Candid — Getting started on a responsible AI use policy for nonprofits (the "when NOT to use AI" list; "don't paste anything you wouldn't put on a public website"): https://candid.org/blogs/how-to-create-responsible-ai-use-policy-for-nonprofits/
• ASU Lodestar Center — Responsible AI: Security and privacy tips for nonprofits (2026; shadow AI, approved-tool policy, don't enter confidential data): https://lodestar.asu.edu/blog/2026/04/responsible-ai-security-and-privacy-tips-nonprofits
• ASU Lodestar Center — 10 Ways to Safeguard Data in the Nonprofit Sector (2025; data policy, minimization, consent, access controls, vendor vetting): https://lodestar.asu.edu/blog/2025/01/10-ways-safeguard-data-nonprofit-sector
• Vera Solutions — Nine Principles of Responsible AI for Nonprofits (privacy & data protection, fairness/bias, human oversight): https://verasolutions.org/blog-nine-principles-of-responsible-ai-for-nonprofits/
• NTEN — AI for Nonprofits Resource Hub & courses (data governance, data culture, responsible AI policy): https://www.nten.org/learn/resource-hubs/artificial-intelligence
• NonProfit PRO — 5 Practical Steps to Build an AI-Ready Data Foundation (assess, prioritize, consolidate, standardize, govern): https://www.nonprofitpro.com/post/5-practical-steps-to-build-an-ai-ready-data-foundation-for-your-nonprofit/

Data hygiene & CRM

• Omatic — Database Cleanup Guide / Data Hygiene Best Practices (duplicates, standardization across NPSP, Raiser's Edge, Apricot): https://omaticsoftware.com/blog/start-cleanup-process-messy-nonprofit-database/ · https://omaticsoftware.com/blog/data-hygiene-best-practices-for-beginners/
• Vera Solutions — Salesforce-based M&E / outcome tracking at scale: https://verasolutions.org/

Privacy law (plain-language)

• Foundation Group — Navigating Donor Data Compliance: GDPR and Beyond (GDPR/HIPAA applicability to nonprofits): https://www.501c3.org/donor-data-compliance/
• Usercentrics — Why the GDPR Matters for Charities and Nonprofits: https://usercentrics.com/knowledge-hub/gdpr-for-charities/

Data-sharing & processing agreements

• ContractsCounsel — Data Sharing Agreement template/guide: https://www.contractscounsel.com/t/us/data-sharing-agreement-template
• GDPR.eu — Data Processing Agreement template (controller/processor clauses): https://gdpr.eu/data-processing-agreement/

AI-ready data & the "never paste" rule (with incidents)

• NonProfit PRO / Transcend / Alteryx / Striim — what "AI-ready data" means (structured, consistent, complete, single source of truth, retention): https://transcend.io/blog/what-is-ai-ready-data-the-complete-guide-for-enterprise-leaders · https://www.alteryx.com/glossary/ai-ready-data
• BigID — Why You Need an AI-Ready Data Inventory: https://bigid.com/blog/ai-data-inventory/
• IT Pro — CISA chief uploaded sensitive docs to public ChatGPT (real incident; why never to do it): https://www.itpro.com/security/data-protection/cisas-interim-chief-uploaded-sensitive-documents-to-a-public-version-of-chatgpt-security-experts-explain-why-you-should-never-do-that
• Wald.ai — ChatGPT Data Leaks and Security Incidents (2023–2026) (Samsung source-code leak, employee data-sharing stats): https://wald.ai/blog/chatgpt-data-leaks-and-security-incidents-20232024-a-comprehensive-overview