AI Data Security

Why nonprofit donor data requires a different conversation about AI

When a hospital evaluates an AI tool, HIPAA creates a legal framework for the conversation. When a financial institution evaluates AI, SOC 2 certifications and regulatory requirements define the minimum bar.

Nonprofits have no equivalent mandate. Which means the burden falls on development directors - people hired for their relationship skills and sector knowledge, not their data architecture expertise - to evaluate AI tools against a standard that nobody has clearly defined for them.

This guide is an attempt to fill that gap. It explains the specific risks that apply to nonprofit donor data, the technical safeguards to look for, and the questions to ask any AI vendor before connecting them to your CRM.

What can actually go wrong with donor data and AI

Training data contamination

Many AI tools use user data to improve their models. If a nonprofit's donor records - names, giving history, personal notes - are used as training data, that information can surface in AI outputs for completely unrelated users. This is not hypothetical; it is a documented risk with consumer AI tools. The safeguard to look for is architectural data isolation, not a policy promise. Ask vendors specifically whether your data is ever used in model training - and whether that guarantee is contractual or architectural.

PII in generated content

AI tools that generate donor communications, reports, or briefings from raw CRM data can inadvertently include personally identifiable information in content intended for public use - newsletters, grant reports, social posts. This is particularly risky when the generation happens in a single step without a redaction layer. The safeguard: automatic PII detection that runs before content is surfaced to the user, not after.

Hallucinated donor history

AI tools that generate from context without grounding their output in retrieved records can produce plausible-sounding but entirely fabricated donor history. A briefing that says "Margaret has expressed interest in your capital campaign" when no such conversation ever occurred is worse than no briefing at all - it creates false confidence and potential donor relationship damage. The safeguard: retrieval-augmented generation (RAG) with mandatory citation. Any AI tool used for donor intelligence should be able to cite the specific record, note, or document that its answer came from.

Five questions your board chair will ask - and how to answer them

1. Does our donor data ever train your AI models?

Good answer: No - your data lives in an isolated environment and is never used to train any model, shared or proprietary. This is an architectural guarantee, not a policy commitment.

Watch for: We take data privacy very seriously and comply with all applicable regulations.

2. What happens to our data if we cancel the service?

Good answer: Your data is deleted from all systems within a specific timeframe, and you receive a full export before deletion.

Watch for: Vague commitments without specific deletion timelines or export guarantees.

3. How do you detect and handle PII in AI-generated content?

Good answer: PII detection runs automatically before any content is surfaced to your team. Donor names, contact information, and sensitive details are flagged and can be redacted with a single action before publication.

Watch for: Manual review requirements, or post-generation redaction that depends on user judgment.

4. Can you show us an audit trail?

Good answer: A demonstrable log of every AI interaction - what was queried, what was retrieved, what was generated, by whom, and when. This is the equivalent of bank-level transaction logging applied to AI interactions with donor data.

Watch for: Logging that captures queries but not retrieved content or generation parameters.

5. Where do your AI answers come from?

Good answer: Every answer is grounded in retrieved documents or records from your own data. We can show you the specific note, email, or CRM record that each answer came from.

Watch for: Generation without citation, or answers that reference "our AI's knowledge" rather than your organization's actual records.

What PII redaction means - and what it doesn't

The phrase "PII redaction" gets used loosely. In practice it covers two distinct technical operations and one critical architectural choice.

Detection is the act of identifying PII inside a body of text - names, emails, phone numbers, addresses, financial details. Detection alone does nothing protective; it only flags what is present.
Redaction is the act of removing detected PII or replacing it with placeholders before the content is stored, indexed, or shown.
Write-time vs. query-time determines when detection happens. Write-time detection runs before data enters the AI's index, so sensitive content never becomes searchable. Query-time detection only intercepts PII at the moment of display - useful, but the data has already been indexed and may persist in vector stores, logs, or caches.

For nonprofits, the categories of PII that matter most are wider than the standard tech-industry list: donor names, contact information, giving amounts, family information, health disclosures captured in gift notes, and immigration or program-eligibility details that staff sometimes record in free-text fields.

Go deeper

AI PII Redaction for Nonprofits (Whitepaper)

A technical deep-dive on threat models, 4-layer detection pipelines, write-time vs. query-time enforcement, and a vendor evaluation checklist.

Read the whitepaper →

A template board talking point on AI adoption

If your board asks about AI tool adoption, here is language you can adapt for your context:

"We evaluated [tool] specifically on three criteria: data isolation, PII handling, and answer grounding. Their architecture ensures our donor data never trains shared models, PII is detected and redacted automatically before content leaves the system, and every AI-generated answer is cited to a specific record in our CRM or files. We have a full AI Policy Pack available for board review."

The "AI Policy Pack" referenced above is documentation that reputable AI vendors should be able to provide - covering their data handling practices, contractual guarantees, audit trail capabilities, and model training policies. Request it before signing any agreement.

About Gratefully

How Gratefully handles nonprofit donor data security

Gratefully was built from the ground up for nonprofit donor data. A few specifics:

Your donor data never trains any AI model - this is an architectural guarantee enforced by environment isolation, not a policy position.
PII detection runs at write time - before data enters the knowledge graph - and again before content is surfaced to users.
Every AI answer is retrieved from and cited to a specific record in your organization's own data. If the information isn't in your records, Gratefully says so.
All AI interactions are logged in a full audit trail - queryable by your team and shareable with your board.
We provide a free AI Policy Pack - board-ready documentation of all data handling practices and contractual guarantees.

Start your free trial or request a security review Get the AI Policy Pack →

Glossary: AI security terms for nonprofit leaders

PII (Personally Identifiable Information): Any data that can identify a specific person - names, emails, phone numbers, addresses, giving amounts tied to an individual, family details, or sensitive disclosures recorded in donor notes.
RAG (Retrieval-Augmented Generation): An AI architecture where the model retrieves specific records from a trusted source before generating an answer, then cites those records. RAG dramatically reduces hallucination because answers are grounded in real data rather than the model's internal patterns.
Data isolation: An architectural separation that ensures one organization's data cannot enter another organization's environment, training set, or AI outputs. Strong isolation is enforced by infrastructure, not by policy.
Audit trail: A complete, queryable log of every AI interaction - the prompt, the records retrieved, the response generated, the user responsible, and the timestamp. Essential for board oversight and incident response.
Write-time detection vs. query-time detection: Write-time detection scans and redacts PII before data is written to the AI's index, so sensitive content never enters the searchable system. Query-time detection only catches PII at the moment of display - by then it has already been indexed and may persist in vector stores or caches.
Training data contamination: When customer or user data is absorbed into a model's training corpus, allowing fragments of that data to surface in responses to unrelated users. A documented risk with consumer AI products.
Model fine-tuning: The process of further training a base model on additional data. Nonprofits should ask whether their data is ever used for fine-tuning - even a vendor that doesn't train base models may fine-tune on customer data unless explicitly prohibited.
Hallucination (in AI context): When an AI generates plausible-sounding but factually fabricated content. In donor work, hallucination can manufacture conversations, gifts, or interests that never existed - creating false confidence and reputational risk.