If you can write an email, you can build your own fundraising tool

When I sat down to plan my Learning Lab at AFP ICON 2026, I set myself one rule: by the end of 20 minutes, the room should leave with a blueprint for building a working fundraising tool, with no code.

The example I used was turning a generic 50-page annual report into a personalised impact report for any donor in your database.

Here’s what I covered, why I think it matters, and how you can try it yourself.

The problem with the annual impact report

Most not-for-profits put real time into their annual report. The PDFs are well designed. The stories are compelling. The financials and impact numbers are all there.

The problem is donors rarely read them.

That’s not a design failure. The reports are built for everyone, which means they’re not built for any one donor. A monthly donor who cares about climate work has to wade through pages on programs they don’t engage with to find the bit that matters to them.

For high-volume programs that’s tolerable. For mid-value donors, who give more than the annual file but don’t have a major gift officer, it’s a missed opportunity.

What changed in AI recently

I’ll repeat the point I opened the session with: this same talk would have failed three months ago. Two things shifted.

First, context windows got much bigger. Tools like ChatGPT, Claude and Gemini can now hold a full annual report, a donor profile and brand guidelines in one conversation. Outputs get sharper because the model can see the full picture, not a fragment.

Second, models can now use tools. They can browse the web, read files, write code and produce HTML pages or PDFs. That turns the chat box from a writing assistant into something closer to a junior analyst with permission to ship.

Together, these changes make it realistic to build small fundraising tools yourself.

Four levels of AI use

In the session I mapped four levels of how fundraisers are using AI today. They build on one another.

Most of the fundraisers I spoke with at AFP lived somewhere between Level 1 and Level 2. The bigger payoff sits at Level 3 and Level 4.

A worked example: Maria's impact report

I used a hypothetical mid-value donor named Maria. Eight years of giving. $500 a year. Interested in climate justice.

Each level of AI use produced something more useful for Maria.

• A Level 1 prompt produced a generic thank-you email.

• A Level 2 prompt, with the annual report attached, produced an email that referenced specific climate programs and tied them back to her giving.

• A Level 3 prompt produced a one-page HTML report personalised to her, with quotes, stories and citations back to the full annual report.

• A Level 4 prompt produced a small web app where any colleague could upload an annual report and a list of donors, choose an output format and generate personalised reports in bulk.

The point isn't the specific output. It's that one annual report can become something each donor will actually open.

Where this fits in your fundraising program

A few themes came up in the questions I got during the session.

Start with the top 20%. Personalisation matters most where it changes a donor's experience and where you have time to review every output. Mid-value donors are an obvious place to start: too many for personal stewardship, too important for a generic file.

Keep a human in the loop. Every output should be read before it goes out. AI is good at drafting and weak at judgement. Your team makes the call on what gets sent.

Don't try to integrate everything on day one. Exporting a list to a spreadsheet and uploading it is faster than building a CRM integration. The point is to get value into donors' hands, not to ship perfect infrastructure.

Cite the source. When you ask the model to reference page numbers in the report, two things happen. Hallucinations drop, and donors get a reason to open the full document.

What to try this week

If any of this resonated, here’s a simple way to put it into practice.

1. Pick 10 mid-value donors with at least three years of giving and a clear interest area.

2. Drop your most recent annual report into ChatGPT, Claude or Gemini.

3. Ask the model to draft a personal impact report for each donor that ties their gifts to specific programs and cites page numbers.

4. Read every output. Edit anything that doesn’t sound right.

5. Send.

That’s a Level 2 workflow you can run today. The Level 3 and Level 4 versions are within reach as soon as you carve out the time.

The bigger picture

The case for building your own tools isn't novelty. It’s leverage. Small teams under permanent constraint don't need bigger systems. They need to choose better and act faster.

AI doesn't replace fundraisers. It removes the work that gets in the way of judgement. Personalised impact reports are one example. Major donor research, appeal analysis, content review and campaign QA are others.

A test I keep coming back to: if you can describe a fundraising task in plain language to a colleague, you can probably describe it to a model. If you can describe it to a model, you can probably build a tool around it.

Closing thought

The version of Maria's report I built in the room wasn't perfect. It didn't need to be. It was specific, it was hers, and it pointed her back to the work she was already paying attention to.

That's the move. One annual report, hundreds of donors, one piece of writing each of them might actually open. The tools to do this are sitting on your laptop. The judgement that makes the output worth sending is already on your team.

So start small. Pick one decision. Build something you can show a colleague next week. If it works, build the next one. That's how a 50-page PDF turns into a stewardship programme your donors look forward to, and how a fundraising team that can't be replaced gets the leverage it deserves.