AI donor communication drafting.

What gets drafted, and what never goes out unread

AI donor communication drafting covers the high-volume writing small fundraising teams fall behind on: thank-you emails after each gift, impact updates, failed regular-giving payment notices and first drafts of appeals. Each draft is personalised from the gift record (amount, date, campaign, designation) and written only from facts the team has approved.

Relationship letters are a different matter, covered below. For sector context, see AI for nonprofits.

How the drafting loop runs

Inputs

Gift records from the donor CRM, communication preferences and consent flags, an approved impact facts library with a source and date for each fact, a tone guide, and receipt data from the finance system.

Systems

A donor CRM such as HubSpot or Salesforce (or a well-kept spreadsheet), the email platform, Xero or MYOB for receipting, and a language model that receives only allowlisted fields.

Steps

Build the batch and remove do-not-contact, deceased and unsubscribed records. Assemble allowed fields per donor. Draft from the facts library only. Check automatically that amounts match the CRM, sender details and an unsubscribe link are present, and no fact or phrase falls outside the library. Queue for review.

Human review

A fundraiser reads every draft above a gift threshold or with an in-memoriam flag, samples one in ten standard thank-yous, and approves each appeal. Finance checks the receipt batch separately.

Outputs

Approved emails and letters, a receipt batch, and an exceptions list (missing names, mismatched amounts).

Measures

Days from gift to thank-you, share of drafts edited, errors caught at review, unsubscribe and complaint rates, and second gifts from first-time donors within 12 months.

Risks

A wrong name or amount, thanking a deceased donor, impact claims the charity cannot support, emailing someone who unsubscribed, and a tone that jars on a memorial gift.

Illustrative example: one month of thank-yous

Illustrative example: an animal welfare charity in Ōtautahi Christchurch receives about 1,150 one-off gifts a month. Assumptions: the team now sends a generic receipt email plus personal letters for larger gifts, which takes about 16 staff hours a month, and thanks arrive nine days after the gift on average. All numbers are synthetic.

At one minute per sampled email and four minutes per letter or memorial note, review takes 106 + 360 = 466 minutes, close to 8 hours, and thanks can go out within two working days. A typical approved fact reads: “Your gift to the winter desexing drive helped fund 212 surgeries (programme report, August 2026).” The model may quote that; it may not round it up to “hundreds of lives saved”.

Rules that apply to donor messages in New Zealand

• Electronic messages. The Unsolicited Electronic Messages Act 2007 requires consent for commercial electronic messages (section 9), accurate sender details (section 10) and a functional unsubscribe facility (section 11). An unsubscribe takes effect five working days after use, and the sender must prove consent. The Department of Internal Affairs says a charity message that does not market goods or services is not commercial; an email selling gala tickets or merchandise does market them. The simplest control is to run consent, sender and unsubscribe checks on every message.
• Receipts. Inland Revenue lists what a donation receipt must show, including the donor’s full name, the amount of money, the date, a clear statement that it was a donation, the organisation’s IRD number and an authorised person’s signature. Gifts of property do not qualify. Donors can claim a one-third tax credit on donations over $5. Receipt fields come from finance data through a fixed template.
• Prospect research. Since 1 May 2026, IPP 3A requires an agency that collects personal information about someone from another source to take reasonable steps to tell them, unless an exception applies. The Commissioner’s guidance treats publicly available sources such as news and public registers as a possible exception, but details from a paid data provider, a mutual contact or a private social account are not. Record which exception you rely on before the research starts.
• Accuracy. IPP 8 requires reasonable steps to check personal information is accurate before use. Stale addresses and unrecorded deaths are the usual failures, so filtering runs before drafting.

What stays with the fundraiser

People own the relationships: major donor conversations, bequests, complaints, and anything after a death. They approve the facts library, set ask amounts for individual donors and judge tone. A model can suggest which lapsed donors to contact first; a fundraiser decides whether that contact is welcome. Our lesson on drafting emails with human review sets out the routine.

Where donor drafting goes wrong

• Claim inflation. Models generalise from specific facts to grand ones. Control: facts library only, a banned-phrase list, and rejection of any number not in the library.
• Merge errors. “Dear Mr and Mrs” for a single donor, or a joint gift thanked under one name. Control: salutation rules held in the CRM, not the prompt.
• Unwelcome personalisation. Mentioning a donor’s employer or likely wealth when they never told you. Control: a field allowlist limited to what donors gave you for this relationship.

Costs and a first project

Costs depend on CRM data quality, the number of segments and templates, and how the email platform connects; HubSpot and similar CRMs have APIs that suit a batch-and-review loop. Model fees for a monthly batch are small next to review hours.

Start with thank-you emails for one campaign and count how many drafts the fundraiser changes. Our AI workflow automation work covers batching, checks and the review queue. Consented stories gathered for grant evidence assembly make strong impact updates, and people who arrive through volunteer enquiry handling should only receive appeals if they have agreed to them.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.