AI Fluency for Every Role (the 4D Foundation) + Champions Network

Why This Module Exists

Almost every non-profit now touches AI — and almost none of them get much from it. The 2026 Nonprofit AI Adoption Report (Virtuous + Fundraising.AI) found roughly 92% of organizations using AI in some form, but only ~7% reporting major improvement in capability, with ~81% using it individually and ad hoc and just ~4% having documented, repeatable workflows (directional). The report's authors call this the "efficiency plateau": the barrier is not access to tools, it is the absence of shared habits and systems around them.

That gap is a people problem, not a technology problem — which is exactly what Assessment Section 4 measures. The most common reason AI stalls in a non-profit is not the software. It is staff who never learned how to direct it, never learned how it fails, and quietly stopped using it after one bad output. Add the sector's structural reality — lean teams, high turnover, tight budgets, and genuine fear (about a third of nonprofits cite staff resistance or ethics as a barrier) — and a single trained "AI person" who leaves takes the whole capability with them.

This module closes that gap two ways. First, it gives every staff member the same foundational fluency: a plain-language picture of what generative AI is, how it fails, and the four habits that make collaboration safe and useful. Second, it stands up an AI Champions network — a small group of respected peers who spread adoption from demonstrated value rather than top-down mandate, run a learning loop that feeds policy updates, and bake fluency into onboarding so it survives turnover.

Learning Objectives

Every staff member — by the end of this module, participants will be able to:

1. Explain in plain language what generative AI is and how it fails (hallucination, bias, and "confident-but-wrong") — without technical jargon.
2. Apply the 4D fluency habits — Delegation, Description, Discernment, Diligence — to a real task from their own role.
3. Decide what to hand to AI versus keep human using the organization's do-use / do-not-use rules (Delegation).
4. Write a clear, context-rich prompt and rewrite a weak one into a strong one (Description).
5. Critically evaluate an AI output for accuracy, bias, and relevance before trusting it (Discernment).
6. Verify, disclose, and take ownership of AI-assisted work, and recite the cardinal data rule: never paste sensitive beneficiary or donor data into public AI tools (Diligence).

AI Champions (A7) — additionally able to:

7. Run a grassroots adoption playbook: recruit by invitation, demonstrate value, win over fence-sitters, and constructively involve skeptics.
8. Operate a learning loop that collects staff issues and proposes policy/cheat-sheet updates, and embed fluency in onboarding so it survives turnover.

Session Agenda

A stacked, mostly-self-paced design built for busy, lean teams. Do not deliver it all in one block.

Core Content

Part 1 — What generative AI actually is (and how it fails), in plain language

A generative AI tool (a "large language model," or LLM — like Claude, ChatGPT, or Microsoft Copilot) is, at heart, a very sophisticated prediction engine. It has read an enormous amount of text and learned the patterns of how words follow one another. When you ask it something, it predicts a fluent, plausible-sounding response — one word at a time. It is not looking anything up in a database of facts, and it is not "thinking" the way a person does.

That single idea explains all three of the failure modes every staff member must recognize:

HOW LLMs FAIL  (the three you must know)
┌───────────────────┬───────────────────────────────────────────────────────────┐
│ Hallucination     │ It invents facts, citations, statistics, or quotes that     │
│                   │ sound real but are false. It is predicting plausible text,  │
│                   │ not retrieving truth.                                       │
├───────────────────┼───────────────────────────────────────────────────────────┤
│ Bias              │ It reflects patterns in its training data — which can carry │
│                   │ racial, gender, language, and socioeconomic bias against    │
│                   │ the very people you serve.                                  │
├───────────────────┼───────────────────────────────────────────────────────────┤
│ Confident-wrong   │ It states wrong answers in the same calm, polished, certain │
│                   │ tone as right ones. Fluency is NOT accuracy.                │
└───────────────────┴───────────────────────────────────────────────────────────┘

Non-profit example. A program coordinator asks an AI tool for "three peer-reviewed studies showing our after-school model reduces truancy." The tool returns three studies — titles, authors, journals, years — all formatted perfectly. Two of them do not exist. The tool hallucinated them, and presented them with total confidence. If that had gone into a grant report unchecked, it would have been a fabrication in a funder document.

Part 2 — The 4D fluency spine

This whole training kit — all eight modules — is anchored on the 4D AI Fluency framework from Anthropic + GivingTuesday's AI Fluency for Nonprofits (built on the AI Fluency framework by Profs. Rick Dakan and Joseph Feller). Four habits, in order, for every role:

   DELEGATION  ──►  DESCRIPTION  ──►  DISCERNMENT  ──►  DILIGENCE
   what to hand      how to ask         how to judge      how to own
   to AI vs. keep    clearly, with      the output for     it: verify,
   human             context            error & bias       disclose, sign
        └────────── Delegation–Diligence loop ───────────┘
                 └─ Description–Discernment loop ─┘

The two inner loops matter: you describe → discern → re-describe until the output is right, and you delegate → stay diligent so you never lose ownership of what AI helped produce.

Delegation — decide what to hand to AI, and what to keep human

The official framework breaks this into goal/task awareness, platform awareness, and task delegation. In plain terms: know your goal, know what this tool can and cannot do, and split the work wisely. In a non-profit, Delegation is first a safeguard: high-stakes decisions about real people — eligibility, benefits, crisis or safety messaging, mental-health support, case notes with sensitive data — stay human. (Module 7 covers the full do-not-use list.)

• Example: Delegate to AI — drafting a volunteer newsletter, summarizing meeting notes, turning bullet points into a tidy paragraph. Keep human — deciding which families get a limited number of housing slots.

Description — prompt clearly, with context and constraints

Officially: product, process, and performance description. Plain version: tell the tool who it's for, what you want, in what format, and what constraints apply. A vague prompt gets a vague answer; a specific prompt with context gets usable work. This is the single skill that most separates frustrated staff from productive staff. (See the before/after artifact below and Part 4's exercise.)

• Example: Not "write a thank-you letter." Instead: "Draft a warm, 150-word thank-you to a first-time $50 donor to our youth literacy program. Mention their gift funds books for one child for a year. Plain, sincere tone, no jargon, 6th-grade reading level."

Discernment — critically evaluate the output

Officially: product, process, and performance discernment. Plain version: read it like a skeptical editor. Is it accurate? Is it biased against anyone you serve? Is it actually relevant, or just fluent? This is where you catch hallucinations and confident-wrong answers before they cause harm.

• Example: AI drafts intake-form language. A discerning reader notices it assumes everyone has a permanent address and a smartphone — a bias that would exclude the unhoused clients the program exists to serve.

Diligence — verify, disclose, and take ownership

Officially: creation, transparency, and deployment diligence. Plain version: you are accountable for anything you send out, whether or not AI helped. Fact-check before release. Disclose AI assistance where your policy or a funder requires it. And follow the cardinal data rule.

Part 3 — The do-use / do-not-use rules (co-created, not handed down)

Every staff member must be able to state the organization's rules before doing real work with AI. The point is not a long policy document — it is a short, memorable list that people actually internalize. A starting template (your Champions and leadership will tailor it):

Part 4 — Free, sector-built places to learn this (so you don't build from scratch)

You can deliver the whole foundational tier on free, non-profit-specific courses. Pick one primary course and supplement.

Part 5 — The AI Champions Network (the A7 companion track)

A single trained "AI person" is a turnover risk, not a strategy. The durable model is a small network of champions — respected peers across roles — who spread adoption from demonstrated value. Grassroots evidence is strong: organizations that picked a handful of natural champions, gave them time and air cover, and ran monthly peer-sharing sessions saw voluntary adoption climb (one cited case reached ~68% in six months) as fence-sitters watched peers get real value (directional).

Why grassroots beats mandate. With AI specifically — where fear, skepticism, and real uncertainty drive resistance — a top-down mandate doesn't just fail, it backfires. Culture can't be commanded. It spreads when a colleague doing your job shows you something that saved them an hour.

THE CHAMPIONS MODEL
   Recruit (invite, don't assign)        2–4 natural champions across roles
            │                            curious · willing to experiment · trusted
            ▼
   Equip within guardrails               approved tools · protected time · air cover
            │
            ▼
   Demonstrate value                     monthly "here's what I tried" peer shares
            │                            quick wins first, honest failures welcome
            ▼
   Spread to fence-sitters               social proof from peers; permission to fail
            │
            ▼
   Embed & preserve                      cheat sheets · prompt library · onboarding
            │                            recorded sessions survive turnover
            ▼
   Learning loop ──────────────────────► staff issues → proposed policy/cheat updates

Selecting champions. Look for the ~10–15% who show three traits: curiosity about AI, willingness to experiment, and existing social capital (peers already trust them). Recruit by invitation and self-selection, never by assignment. Spread them across functions — a fundraiser, a program lead, an ops person — not just "the tech-comfortable" ones.

Overcoming fear and resistance. Create psychological safety: it's fine to admit you don't understand AI, fine to experiment and fail, and fine to thoughtfully not adopt a tool. Leaders should model the learning curve out loud ("here's where Copilot confused me this week"). And don't dismiss the skeptics — involve them: appoint resisters to draft the do-not-use list, set output-quality standards, or stress-test prompts. That turns resistance into ownership of the very safeguards they care about.

Part 6 — Low-cost formats and a realistic time-to-competency

Fluency is built through practice, not a single sitting — the 4D framework is explicit that it develops over repeated use, not overnight. For a lean, high-turnover team, sequence a stack of mostly-free formats rather than one big event:

A realistic time-to-competency. Set expectations honestly so people don't quit after one frustrating session:

Day 0       ~90 min self-paced course + 30-min kickoff
Week 1–2    First real-but-low-risk task (draft, summary, reformatting) with the cheat sheet at hand
Week 3–6    Two to three workflows feel natural; staff start contributing to the prompt library
Month 2–3   "Aware" → early "Practicing": at least one documented, repeatable workflow per team

Hands-On Exercises

Exercise 1 — Rewrite a weak prompt (Description)

Scenario: You need a short social post announcing your food bank's holiday drive.

Task: Start from the weak prompt, then rewrite it using Description (who it's for, what you want, format, tone, constraints).

Weak prompt: Write a social media post about our food drive.

Facilitator sample solution (strong prompt):

Debrief: The strong version names audience, length, platform, dates, tone, the specific ask, an impact line, and constraints. That is Description. Note how much less editing the strong version's output needs.

Exercise 2 — Catch the failure (Discernment + Diligence)

Scenario: A staff member used AI to draft a paragraph for a grant report. Read this draft critically.

Task (individually, then discuss): List everything you would verify or flag before this goes to a funder.

Facilitator answer key:

• Hallucination risk: The citation (Henderson et al., 2023) and journal may not exist — verify it independently; AI invents plausible references.
• Confident-wrong numbers: "1,200," "340%," and "100%" must be checked against your actual M&E data. AI will state any number confidently.
• Bias / overclaim: "100% improved" is implausible and could mislead a funder — a fabrication risk even if AI produced it.
• Diligence / ownership: Whoever submits this is accountable, full stop. Verify every fact, then decide whether your policy requires disclosing AI assistance.
• Data rule: Confirm no identifiable beneficiary data was pasted into a public tool to generate this.

Exercise 3 — Delegation sort (5 minutes, table groups)

Task: Sort these tasks into Delegate to AI, AI-assisted with heavy human review, or Keep fully human: (a) summarizing a 20-page report; (b) deciding which client gets the last shelter bed tonight; (c) drafting a donor thank-you; (d) writing case notes about a domestic-violence survivor; (e) translating a general event flyer; (f) screening intake forms to flag for staff follow-up.

Facilitator answer key: Delegate — (a), (c), (e). AI-assisted with heavy review — (f) AI flags, human decides. Keep fully human — (b) high-stakes allocation, (d) sensitive beneficiary data + judgment. Use this to surface why — tie each to Delegation as a safeguard.

Templates & Takeaway Artifacts

Artifact A — The 4D Habits one-page cheat sheet (print and pin)

┌─────────────────────────────────────────────────────────────────────────┐
│  THE 4D HABITS — before any AI work          [Org name] · v1.0           │
│  "AI amplifies your mission — it does not replace your judgment."         │
├─────────────────────────────────────────────────────────────────────────┤
│ 1. DELEGATE   Should AI do this at all? High-stakes decisions about real │
│               people stay HUMAN. Check the do-not-use list.              │
│                                                                           │
│ 2. DESCRIBE   Tell it: WHO it's for · WHAT you want · FORMAT · TONE ·     │
│               CONSTRAINTS. Vague in = vague out.                         │
│                                                                           │
│ 3. DISCERN    Read it like a skeptical editor. Accurate? Biased against  │
│               anyone we serve? Actually relevant? Fluent ≠ correct.      │
│                                                                           │
│ 4. DILIGENCE  Verify every fact, number, name, citation. Disclose if     │
│               required. YOU own the output. Sign off knowingly.          │
├─────────────────────────────────────────────────────────────────────────┤
│ ⛔ CARDINAL RULE: Never paste beneficiary/donor data into a PUBLIC AI    │
│    tool. Use approved tools only. De-identify first. When in doubt,      │
│    leave it out.                                                          │
│ HOW IT FAILS: Hallucination · Bias · Confident-but-wrong.                │
│ Stuck or unsure? → Ask your AI Champion: [name / channel]                │
└─────────────────────────────────────────────────────────────────────────┘

Artifact B — Good vs. weak prompts (before/after set, keep adding to the shared library)

Artifact C — AI Champion Charter (template)

AI CHAMPION CHARTER — [Org name]
1. PURPOSE
   Spread safe, useful AI habits from demonstrated value (not mandate),
   run the learning loop, and preserve fluency against staff turnover.

2. CHAMPIONS (2–4, across roles — not just the tech-comfortable)
   Name | Role | Function they represent
   ____ | ____ | __________________________
   Selection: invited & self-selected · curious · experimental · trusted by peers.

3. MANDATE (what champions DO)
   • Model the 4D habits and the do-not-use list in their own work
   • Run monthly peer "here's what I tried" shares + bi-weekly lunch-and-learns
   • Grow the shared prompt library; maintain the one-page cheat sheets
   • Onboard new hires on AI fluency (embed in onboarding checklist)
   • Run the learning loop (Artifact D) → propose policy/cheat-sheet updates
   • Support skeptics constructively; never shame non-adoption

4. WHAT CHAMPIONS DO **NOT** DO
   • Are not IT/security owners and not the AI police
   • Do not approve new tools or override the do-not-use list (that's leadership)
   • Do not handle sensitive data on others' behalf in public tools

5. RESOURCES & AIR COVER (leadership commits)
   • Protected time: ~[2–4] hrs/month, shielded from productivity penalties
   • Access to approved tools (Module 3) · a sandbox for safe practice
   • A standing channel + quarterly 30-min review with leadership

6. CADENCE & MEASURES
   • Monthly share · quarterly learning-loop review → policy v[N.N]
   • Tracked: prompt-library entries used · documented workflows · onboarding
     completion · confidence shift (see Outcome Scorecard)

7. TERM & SUCCESSION
   • [12] months, renewable. Each champion names/trains a backup so the
     capability survives departure. Recorded sessions kept in [location].

   Approved by: ____________ (ED/Leadership)   Date: __________

Artifact D — Learning Loop intake (staff report issues → propose policy updates)

AI LEARNING-LOOP INTAKE — submit anytime to [channel/form]
Date: ______   Submitted by (optional): ______   Role: ______

1. What were you trying to do with AI?  ______________________________

2. What happened? (the issue, surprise, win, or risk)  _______________

3. Type:  [ ] Hallucination/wrong fact  [ ] Bias concern
          [ ] Data/privacy worry        [ ] Tool limitation
          [ ] A win worth sharing       [ ] Unclear rule / gap in policy

4. Did anything sensitive get exposed?  [ ] No  [ ] Possibly → escalate now

5. Your suggestion (cheat-sheet tweak, new rule, prompt to add, training):
   ____________________________________________________________________

— CHAMPIONS REVIEW (monthly) —
Theme/pattern: ___________  Action: [ ] Update cheat sheet  [ ] Propose
policy change to leadership  [ ] Add to prompt library  [ ] Lunch-and-learn topic
Owner: ______  Target date: ______  Policy version bumped to: v____

Knowledge Check

1. (MCQ) A "hallucination" in an AI tool means: (a) the tool crashes; (b) the tool invents plausible-sounding but false information; (c) the tool refuses to answer; (d) the tool is offline.
2. (MCQ) Which task should stay fully human? (a) summarizing a long report; (b) drafting a thank-you note; (c) deciding which family gets the last shelter bed; (d) reformatting bullet points.
3. (MCQ) The cardinal data rule says you should never: (a) use AI at work; (b) paste sensitive beneficiary/donor data into a public AI tool; (c) write long prompts; (d) disclose AI use.
4. (Short answer) Name the four 4D habits in order and one sentence on each.
5. (Short answer) Rewrite this weak prompt into a strong one: "Write a newsletter intro."
6. (MCQ) "Confident-but-wrong" means: (a) the AI is rude; (b) wrong answers are stated in the same polished, certain tone as correct ones; (c) the AI always admits mistakes; (d) the AI only fails on hard questions.
7. (Short answer) Why does the grassroots Champions model work better than a top-down AI mandate in a non-profit?
8. (Short answer) Give two ways the organization preserves AI fluency against high staff turnover.
9. (MCQ) A skeptical staff member is best handled by: (a) ignoring them; (b) mandating tool use; (c) inviting them to help write the do-not-use list and quality standards; (d) removing their tool access.

Answer key

1-b · 2-c · 3-b · 4 — Delegation (decide AI vs. human), Description (prompt clearly with context/constraints), Discernment (critically evaluate output for error/bias), Diligence (verify, disclose, take ownership). · 5 — strong prompt names audience, topic, length, tone, key points, and constraints, e.g., "Write a 120-word, upbeat intro for our monthly donor newsletter highlighting our 500th meal served; plain language, ends with a thank-you, no jargon." · 6-b · 7 — AI provokes fear/skepticism; mandates backfire; adoption spreads through peer-demonstrated value and social proof, and culture can't be commanded. · 8 — any two: one-page cheat sheets, shared prompt library, recorded sessions, AI fluency in onboarding, named champion backups/succession. · 9-c.

Facilitator Guide

Prep checklist

• Pick ONE primary self-paced course (Anthropic + GivingTuesday, or NetHope + Microsoft) and send the link ~1 week ahead.
• Finalize a draft do-use / do-not-use list and the cardinal data rule (co-create with staff/champions — don't hand it down).
• Print Artifact A (4D cheat sheet) for everyone; pre-fill the champion name/channel.
• Prepare ONE live failure demo in an approved tool (ask for a fake citation; show the hallucination).
• Confirm the approved-tool list (Module 3) so no one practices in an unapproved tool with real data.

Free/low-cost materials needed: a screen for the live demo, printed cheat sheets, access to one approved AI tool for the sandbox, a shared doc/folder for the prompt library and the learning-loop intake. No paid tooling required.

Timing: keep the live kickoff to 30 minutes — it sets tone and rules, not depth. Depth comes from the self-paced course and ongoing lunch-and-learns.

Common pitfalls

• Skipping the live demo. Reading about hallucination lands far weaker than watching the tool confidently invent a citation in real time. Always demo.
• Policy paralysis. Teams stall waiting for a "complete" policy. Ship the one-pager; iterate via the learning loop.
• One champion, not a network. A single AI person is a turnover risk. Insist on 2–4 across functions, each with a named backup.
• Mandating use. Pressure breeds quiet non-adoption. Lead with invitation and demonstrated value.

Discussion prompts: Where in your week could AI save you the most time? What's the one task you'd never want AI deciding? What's your biggest worry about AI — and is it about the tool, or about how it might be used here?

Tailoring by audience: A1/A2 leadership — emphasize modeling the learning curve and giving champions air cover. A3 program staff — lead with the do-not-use list and beneficiary bias (bridge to Module 7). A4 development — use the fundraising prompt examples (bridge to Module 6). A5 ops — emphasize repeatable workflows and the prompt library. A6 IT — reinforce the cardinal data rule and approved tools (Module 3). A7 champions — run the separate ~2-hour track on the charter, recruitment, and learning loop.

Addressing staff fear/resistance (a real non-profit barrier): name it directly. About a third of nonprofits cite staff resistance or ethics as a barrier. Co-create the rules with the people who'll follow them. Frame AI as cutting admin burden so staff get more time with the people they serve. Make psychological safety explicit: it's okay to not know, to fail experimenting, and to thoughtfully decline a tool. Convert skeptics into safeguard-authors. Celebrate early wins publicly and by name.

Outcome Scorecard

Tie to Kirkpatrick (weight L2–L4) and the plan's "ad hoc → embedded" target.

Further Resources & Sources

Foundational courses (free / low-cost):

• Anthropic + GivingTuesday — AI Fluency for Nonprofits (Anthropic Academy): https://anthropic.skilljar.com/ai-fluency-for-nonprofits
• AI Fluency: Framework & Foundations (the underlying 4D course): https://anthropic.skilljar.com/ai-fluency-framework-foundations
• NetHope + Microsoft — Unlocking AI for Nonprofits: https://nethope.org/programs/unlocking-ai-for-nonprofits-enroll-in-our-new-ai-skills-course-for-nonprofits/
• Microsoft Learn — Introduction to AI Skills for Nonprofits: https://learn.microsoft.com/en-us/training/paths/introduction-to-ai-skills-for-nonprofits/
• data.org — AI Skills for Nonprofits: https://data.org/resources/ai-skills-for-nonprofits/
• NTEN — AI for Nonprofits Certificate & Nonprofit Tech Readiness cohort: https://www.nten.org/learn/professional-certificates/ai-for-nonprofits · https://www.nten.org/learn/nonprofit-tech-readiness/ai-for-nonprofits

4D framework detail:

• The 4D AI Fluency framework (Dakan & Feller, with Anthropic) — public framework site: https://aifluencyframework.org/
• Practical 4D sub-component breakdown (Ringling LibGuide): https://ringling.libguides.com/ai/framework
• Disco — applying the 4D framework in an org: https://www.disco.co/blog/anthropics-ai-fluency-course-how-to-upskill-your-org-with-the-4d-framework

Change management / AI Champions:

• Bosio Digital — AI Change Management: From Skeptics to Champions Without Mandates: https://bosio.digital/articles/organizational-change-in-ai-adoption
• OpenAI Academy — The AI Champion role: https://academy.openai.com/public/clubs/champions-ecqup/resources/the-ai-champion-role

Adoption baselines (mark directional):

• 2026 Nonprofit AI Adoption Report (Virtuous + Fundraising.AI): https://virtuous.org/resource/the-2026-nonprofit-ai-adoption-report-download/ — ~92% touch AI · ~7% major impact · ~81% ad hoc/individual · ~4% repeatable workflows. Coverage: https://www.nonprofitpro.com/article/nonprofit-ai-adoption-hits-92-but-only-7-see-major-impact/
• TechSoup — State of AI in Nonprofits 2025 (1,300+ respondents): https://page.techsoup.org/ai-benchmark-report-2025 — ~85.6% exploring AI, only ~24% with a formal strategy; large-vs-small adoption gap (~66% vs ~34%).