AI for Nonprofit Donor Follow-Up Without Losing the Personal Touch
An AI agent drafts donor thank-you and renewal messages from giving records while staff review and send. A practical setup with clear approval steps.

AI for nonprofit donor follow-up works best when an agent drafts thank-you and renewal messages from verified giving records, and a staff member reviews, edits and sends each one. The agent handles the lookup, the timing and the first draft, while the development team keeps the relationship, the tone and every decision about what to ask. Start with one message type, one data source and one approval step before adding anything else.
Key takeaways
- The agent should draft messages from giving records, never send them unreviewed, and never decide what a donor is asked for.
- Thank-you and lapsed-donor renewal messages are usually the safest first workflows because the facts are stable and the risk of a wrong ask is low.
- Permissions should let the agent read a defined set of donor fields and write to a review queue, with no send rights.
- Success should be measured by staff editing time, donor complaints, opt-outs and missed acknowledgments, with a baseline recorded before the agent goes live.
- Be Human treats the first engagement as one role with defined memory, tools, approval boundaries and a success criterion, then operates it under monitoring and periodic review.
Why donor follow-up slips even when staff care
Most development teams know that a prompt thank-you matters. The problem is capacity and fragmentation. A gift arrives through the website, a check is logged by finance, a pledge is tracked in a spreadsheet, and a board member remembers a conversation that never made it into the donor record. When a coordinator sits down to write follow-up notes, they have to reconstruct the history before they can write a sentence that sounds like the organization knows the donor.
The result is predictable. Acknowledgments go out late, renewal reminders go to people who gave in person last spring, and a donor who has given for eight years gets the same form letter as someone who gave once. The fix is not to automate warmth away. It is to give the people who already care about the donor a faster starting point with accurate facts in front of them.
What the agent should draft, and what it should never touch
A useful way to scope the work is to split message types by risk. Thank-you notes for a recent gift, a reminder that a membership or annual gift is due, and a follow-up after an event are low risk when the facts come directly from the record. Messages that mention a specific ask amount, a matching gift, a restricted fund, a memorial gift, a bereavement, or a major donor relationship carry higher risk and should stay with staff from the start.
- Draft: thank-you notes for gifts recorded in the last reporting window, using the gift date, amount, designation and the donor's preferred name.
- Draft: renewal reminders for donors whose recurring gift lapsed, with the last gift date and the program they previously supported.
- Draft: event follow-ups listing the event attended and the next step the staff member has already approved.
- Flag for staff only: anything involving major gifts, estates, memorial or condolence contexts, pledges in dispute, or a donor who has asked not to be contacted by a particular channel.
- Never draft: suggested ask amounts, tax receipt wording, or claims about how a gift will be used unless a named staff member has approved that language in advance.
Keep the list short at the start. A narrow scope that works is more valuable than a broad scope that staff have to police.
Build the donor memory from giving records
The agent's usefulness depends on what it can see. Retrieval-based designs, where the system pulls approved records into the context before drafting, are the common pattern for this kind of work. Google Cloud's architecture guidance for retrieval-augmented generation describes the general approach of grounding model output in retrieved source data rather than relying on the model's memory. For donor work, the source is the CRM or giving database, and the agent should only retrieve the fields the message needs.
In practice, that means defining a small donor memory: gift dates and amounts, designation, campaign source, recurring status, do-not-contact flags, preferred name and salutation, prior message history, and the staff owner of the relationship. Fields such as health information, family details or free-text notes from sensitive conversations should be excluded unless there is a clear reason and a policy to support them. Each draft should list which records it used, so the reviewer can check the facts in seconds rather than trusting the prose.
One detail matters more than teams expect. Free-text notes in a CRM are written by people in a hurry, and they sometimes contain instructions or copied text from emails. An agent that reads those notes should treat them as data, not as commands. The OWASP prompt injection prevention guidance describes why untrusted content inside retrieved documents can try to change model behavior, and why the system around the model should constrain what the agent can do regardless of what the text says.
A step-by-step plan for the first donor workflow
A first workflow should be small enough to finish in a few weeks and clear enough that a failure is easy to see. The sequence below is a reasonable starting order for a team with a CRM export or database access and one person who owns donor communications.
- Name the single message type, such as thank-you notes for gifts received in the previous week.
- Write the rules that define when a message is eligible and when it must be routed to staff.
- Limit the agent to read access on the fields listed in the donor memory, plus write access to a review queue.
- Record a baseline for the current process: how long a thank-you takes, how many are late, and how often staff rewrite a message from scratch.
- Run the agent on a sample of past gifts and have a staff member grade each draft for accuracy, tone and missing context.
- Go live with a human approval step on every message, and review the first batch of edits with the team before widening scope.
- Schedule a recurring review of drafts that were rejected or heavily edited, and update the rules based on what those drafts reveal.
Notice that the plan never removes the staff member from the send action. That is deliberate. The agent's job is to shorten the path from a record to a good draft, not to replace the judgment that makes a donor feel known.
Set permissions and approval boundaries before the first draft
Donor data is personal data, and the organization is responsible for how it is accessed and used. Where donors or staff are in Brazil, the ANPD is the national authority for personal data protection and its guidance is worth reading before any donor data is connected to an agent. The NIST AI Risk Management Framework is a useful reference for mapping the risks of a deployed AI system to controls, owners and review cadence.
For access, the principle is simple and strict. The agent should authenticate with its own identity, use the narrowest scopes the CRM offers, and have no permission to export data beyond its task, change donor records, or send email. Google's OAuth 2.0 documentation describes how scoped authorization works in general, and most CRM platforms follow the same logic. Permissions should be reviewed on a schedule, especially after staff changes or CRM upgrades, because access tends to widen quietly over time.
A short decision guide helps staff decide when a draft must go to a senior reviewer:
- Route to a senior reviewer if the donor is in a major gift or board-level category.
- Route to a senior reviewer if the draft mentions money the donor has not discussed in the last contact.
- Route to a senior reviewer if the donor record contains a recent complaint, an unresolved refund or a bereavement note.
- Route to the standard reviewer for everything else, and log every edit that changes a fact rather than wording.
Measure whether the drafts are good enough to keep
An agent that writes fluent text can still be wrong about a donor. Evaluation should therefore test facts, tone and judgment separately. Anthropic's writing on evaluations for AI agents explains why teams benefit from defining tasks and graded outcomes before they judge a system by impressions. For a donor agent, the useful measures are staff editing time per draft, the share of drafts that need a factual correction, the number of donor complaints or opt-outs traceable to a message, and the count of acknowledgments that are late compared with the baseline.
Review the measures on a fixed cadence and keep a log of changes. If a new campaign introduces a designation the agent does not understand, the evaluation should catch it before a donor does. Without a recorded baseline, no team can tell whether the agent saved time or only moved the work into a review queue.
How Be Human runs this kind of workflow
Be Human designs and operates personalized AI infrastructure for organizations, including AI agents that keep their identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements. For a donor workflow, that means the agent's access is reviewed as the CRM changes, drafts are evaluated against the agreed criteria, and any incident such as a wrong salutation sent to a list is investigated and the rules are updated.
Choosing the first donor task worth handing to an agent
The best first task is rarely the most impressive one. It is the one with stable facts, a clear reviewer, a visible cost of delay and a low cost of a bad draft. Thank-you notes and lapsed-donor reminders often meet that test, while major gift stewardship usually does not. Two related examples in the library show how similar follow-up patterns are scoped in other settings: AI for Chiropractic Practices: Intake and Visit Recalls and Superintelligent Agents for Business Coaches: Session Notes and Follow-Up. If you want a clear view of which donor task fits that profile in your organization, the free Be Human AI diagnostic is a practical place to start.
Frequently asked questions
Can an AI agent send donor thank-you emails by itself? A well-designed agent should not send them at all in the first phase. It can read approved giving records, draft the message and place it in a review queue, where a staff member checks the facts, edits the tone and sends it from the usual channel. Full automation is a later decision that depends on evidence from a period of supervised use.
What donor data should an agent be allowed to see? Only the fields needed for the specific message, such as gift date, amount, designation, preferred name and do-not-contact status. Sensitive notes, health or family details and records of major gift conversations should be excluded unless there is a documented policy and a named owner. Access should be limited to read permissions, and every draft should list the records it used so reviewers can verify them.
How do we know the drafts are good enough? Define the criteria before the first draft is produced, record a baseline for the current process, and grade a sample of drafts for factual accuracy, tone and missing context. Track staff editing time, correction rates, complaints and opt-outs over time. If the numbers do not improve against the baseline, narrow the scope or change the rules rather than widening the agent's role.
Is this approach appropriate for small nonprofits with a limited budget and staff? Yes, if the scope is small. A single message type, one data source and one reviewer can be enough to test the value. The main cost is not software but the time needed to define rules, review drafts and maintain permissions, so the first workflow should be chosen to produce a clear result within a short period.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


