A nonprofit considered AI-personalized thank-you letters. They reached Create but with strict guardrails: AI drafts only, human signs every letter, sender knows the donor.
Built a draft-only AI assistant. Human relationship-officer signs every letter; AI never sends.
Our development team — 4 people writing thank-yous to 11,000 annual donors. They send template letters most of the time because they can't personalize at scale. Donors feel like ATMs. Senior donors who used to get handwritten notes now get generic ones. Staff feel guilty about it.
Christ would see that gratitude has become a workflow. He'd see donors who give sacrificially and receive a form letter that thanks 'you' but knows nothing about them. He'd see staff who entered this work to build relationships now reduced to mail-merge operators.
Gratitude is the easiest thing to industrialize and the easiest thing to ruin by industrializing. The temptation here is to use AI to scale the appearance of personal attention — which is the precise opposite of personal attention. If you proceed, the only honest path is to use AI as a drafting aid for a human who actually knows the donor, not as a replacement for knowing.
An AI assistant that drafts personalized thank-you letters using donor history, prior correspondence, and giving context. The relationship officer reviews, edits, signs by hand, and mails.
Our development director Priya manages 2,800 donor relationships. She can write 30 personal letters a week. The other 2,770 donors get a templated thank-you. We have ~11,000 active donors total. Real people: donors like Margaret, 78, who has given monthly for 19 years and gets the same letter as a first-time giver.
Partly. Hiring two more relationship officers would help but not solve scale. Gratitude at this scale genuinely benefits from drafting assistance — the question is what kind.
Real need: every donor should feel known by a specific human. The tool should support that, not simulate it.
Donors who discover their 'personal' letter was AI-written. Trust collapses. The risk is highest with our most loyal donors — exactly the people we'd most want to honor.
Hire more officers (in progress). Segment donors so top tier always gets handwritten. Send fewer letters total but make each one matter. Pursued in parallel.
Bloomerang, DonorPerfect, and Virtuous all have AI letter-drafting features. Salesforce NPSP has Einstein. Several Christian fundraising platforms (Pushpay, Subsplash) offer similar.
Yes. Most are oriented toward acquisition and conversion, not relational stewardship. The defaults push toward generated-and-sent, not drafted-and-reviewed. Theological alignment is weak — donors are 'segments,' not neighbors.
High dependency, vendor lock-in on our donor data, and a generated-by-default UX that would silently erode our practice over time. We don't trust ourselves to resist the defaults.
Build a thin layer over an existing LLM where the relationship officer enters 3 facts they actually know about the donor before any draft is generated. The AI cannot draft without those 3 facts.
Remove the 'send' button entirely. The tool produces a draft. Nothing sends from the tool. The officer must print, sign, and mail. Friction is the feature.
Yes — strictly advisory. The AI never decides what to say; it suggests phrasing for what the human already knows.
Three hard guardrails: (1) Officer must enter 3 specific facts about the donor before any generation. (2) Tool produces drafts only; no send capability. (3) Every letter is signed by hand by a named relationship officer. Disclosed to donors in our annual transparency report.
Misuse: an officer rubber-stamps the draft without reading. Detection: monthly audit, random sample of 50 letters, officer interviewed about what they remember of the donor. Misuse: fact-fields filled with junk to bypass the gate. Detection: validation on the fact fields plus quarterly review.
Priya Ramanathan, Director of Development. Named in the rollout doc. Named in this discernment.
If misuse audits show >10% rubber-stamping, the tool is paused for a quarter and we revisit. Hard kill-switch in the admin panel. Full rollback to template-letter workflow in under 24 hours.
The smallest version: a single text box that takes 3 facts and returns one paragraph. No CRM integration in v1. No batch mode. One donor at a time. One officer at a time. We can grow later if it's working; we cannot shrink later if it isn't.
This team earned Create. They worked through Reject (no, gratitude can't be fully automated), Receive (no off-the-shelf tool fit their donor base), and Reimagine (mail-merge was too cold, fully generative too dishonest) before arriving at a constrained Create. The guardrails are not decoration — they are the substance: AI drafts, human signs, sender knows the donor, no AI letter ever leaves without 90 seconds of human attention. The smallest faithful version is what was shipped.
Start from this example as a draft. Every field will be pre-filled — edit freely. Your own context will surface as you go.
These examples are illustrative. Real discernments will be more complex, more painful, and more specific to your context. The library helps you see the shape of the work.
Examples paraphrased from common patterns observed in faith-based development AI projects. Not based on any single real organization.