SOP-25 · Technology · v1.0

AI Tool Request & Approval

Owner Role

IT Lead (with Operations and Legal input)

Approver

Executive Director

Effective Date

[YYYY-MM-DD]

Review Cadence

Annual (with quarterly review of approved tools)

Last Reviewed

[YYYY-MM-DD]

1. Purpose

To bring AI tool adoption inside [Org Name]'s vendor risk, data protection, and operational governance — not outside it. Staff are using AI now, with or without permission; consumer products were designed for individual use, not for organizational data. This procedure makes adoption deliberate: identify the use case, evaluate the tool, configure it correctly, train staff to use it well, and audit usage. Failure to do this lets sensitive data flow to free consumer AI services that retain, train on, or expose it.

2. Scope

This SOP covers: any AI tool used in [Org Name]'s work, including:

• Generative AI assistants (ChatGPT, Claude, Gemini, Copilot, etc.).
• AI features embedded in existing tools (Salesforce Einstein, Microsoft 365 Copilot, Google Workspace Gemini, Notion AI, etc.).
• AI-powered transcription / meeting note tools (Otter, Fireflies, Granola, etc.).
• AI-powered code assistants (Copilot, Cursor, etc.).
• Image / video generation tools.
• AI-based analysis tools (data analysis assistants, summarization, translation).
• Custom AI built on top of provider APIs (e.g., custom GPTs, Claude Projects, internal tools using OpenAI/Anthropic/Google APIs).
• AI use within third-party vendors' products (vendor must disclose).

This SOP does not cover: general-purpose ML embedded in spam filters, search engines, or consumer device features that do not process [Org Name] data outside the device.

Applies to: all staff, contractors, board members; any party using AI tools on [Org Name] business or with [Org Name] data.

3. Definitions

• Generative AI: systems that produce text, images, code, audio, or video output based on prompts.
• Foundation model: the underlying large model (e.g., GPT-4, Claude, Gemini) accessed directly or through an application.
• Training data: data the AI provider may use to improve their models. Most consumer AI services train on user inputs by default unless you opt out or use enterprise terms.
• Hallucination: AI-generated content that is plausible-sounding but factually incorrect or fabricated.
• Approved tool: an AI tool reviewed and authorized for use at [Org Name] for specific use cases with specific data classifications.
• Data classification: sensitivity level (Public, Internal, Confidential, Restricted) per the Data Classification Policy.

4. Procedure

4.1 Default rule

1. No AI tool is used with [Org Name] data unless approved. Period.
2. Use of consumer AI accounts (free ChatGPT, free Gemini, etc.) with [Org Name] data is prohibited regardless of intent.
3. Personal exploration of AI tools on personal accounts with no [Org Name] data is permitted; do not paste [Org Name] documents, donor information, employee data, financials, board materials, or program data into personal AI accounts.

4.2 Request

Any staff member proposing to use an AI tool for [Org Name] work submits an AI Tool Request to the IT Lead with:

1. Tool name, vendor, and link.
2. Proposed use case (specific — "summarize meeting transcripts" not "AI for productivity").
3. Data type and classification proposed for use.
4. Frequency / scale of use.
5. Number of users.
6. Whether existing approved tool can serve the same use case.

4.3 Evaluation (by IT Lead)

Vendor due diligence per SOP-03 — vendor onboarding tier appropriate to data classification.

AI-specific evaluation:

• Data handling: does the provider train on customer inputs by default? Is there an enterprise tier or opt-out that prevents training? Is data retained for review by humans? For how long?
• Tenancy: is data isolated per customer (enterprise) or commingled (consumer)?
• Encryption: in transit and at rest.
• Geographic processing: where is data processed? Cross-border implications.
• Sub-processors: who else handles the data?
• Compliance posture: SOC 2, ISO 27001, GDPR, HIPAA (if applicable).
• DPA available: signed Data Processing Agreement.
• Output ownership and copyright: who owns generated content; any restrictions on commercial use; indemnification for IP claims.
• Hallucination controls: are outputs labeled as AI-generated? What guidance exists for fact-checking?
• Auditability: can [Org Name] retrieve usage logs?
• Termination provisions: what happens to data on termination?

Risk classification:

• Low-risk: approved AI processing only Public or Internal data with no PII.
• Medium-risk: Internal data with limited PII; commercial enterprise tier required.
• High-risk: Confidential data, donor/employee/beneficiary PII, financial data; requires enterprise tier with strong DPA, signed agreement, and ED approval.
• Prohibited: Restricted data (e.g., minor PII for orgs serving children, health data subject to HIPAA, payment card data) unless tool has specific certification.

4.4 Decision and configuration

1. IT Lead approves, denies, or requests additional information.
2. Approved tools added to the AI Approved Tools Register with: tool name, approved use cases, approved data classifications, configuration requirements, account/license type required (consumer accounts NOT permitted for org data), approval date, renewal date.
3. Configuration enforced — IT Lead provisions through SSO where possible; verifies training opt-out where applicable; configures admin settings.
4. License procurement through standard procurement (no personal credit card reimbursement for AI tools).
5. DPA executed for vendors processing PII.

4.5 Specific use case guidance

• Donor communication drafting: AI may draft, human must review and edit; final donor-facing content is human responsibility.
• Meeting transcription / notes: ensure attendees are notified that AI is recording; consent required in two-party-consent states (CA, FL, IL, MD, MA, NH, PA, WA).
• Confidential discussion (board, legal, personnel, fundraising strategy): transcription / AI participation typically NOT permitted. Default off.
• Financial analysis / decision support: AI as input to human decision; AI does not make financial decisions; outputs verified.
• Performance reviews / employment decisions: AI must not be sole basis; bias risk; legal exposure.
• Beneficiary-facing AI (chatbots, intake): very high scrutiny; bias and access implications; involve program leadership and legal.

5. Records & Retention

7. Roles & Responsibilities

• IT Lead: Owns the procedure; evaluates requests; maintains Approved Tools Register; runs discovery audits.
• Executive Director: Approves high-risk tools; approves AI Use Policy; communicates board.
• Operations Director: Manages vendor relationship and DPA execution.
• Legal counsel: Reviews DPAs and AI-specific terms; advises on bias and use restrictions.
• HR Lead: Owns AI-in-employment guidance; manages training rollout.
• All staff: Comply with policy; do not use unapproved tools; report shadow AI; verify AI outputs.
• Audit/Finance Committee: Reviews annual AI posture report.
• Board: Annual AI risk awareness; approves AI Use Policy.

8. Related Documents

• SOP-03 Vendor Onboarding
• SOP-13 Incident Reporting
• SOP-14 Data Breach Response
• SOP-17 IT Access Provisioning
• SOP-24 Cybersecurity Annual Checklist
• AI Use Policy · Acceptable Use Policy · Data Classification Policy
• NIST AI Risk Management Framework
• State two-party consent recording laws

This sample shows SOP-25 in its full form, exactly as delivered in the library — including the Common Pitfalls section that buyers consistently call out as the most valuable part of each document.