Key takeaways
- Start with one document and one pain point, not a giant data lake.
- Chain tools, each doing what it's good at: a CustomGPT grounded in your documents, ChatGPT as the architect and a no-code builder like Base44 for the clickable views.
- Chained tools are easier to debug, easier to explain to your team and easier for nontechnical staff to trust.
I didn't set out to build some futuristic AI system. I had a simple problem: a short, dense, jargon-filled budget memo on $644.9 million in last-second cuts to Michigan's state budget that almost no human could actually read.
If you work in nonprofits, advocacy or policy, you know the type: technically short, practically unreadable and hugely consequential. So I asked the usual question: what's a clever way to turn this into useful, clickable, plain-language information, using tools anyone in nonprofit tech could handle?
The answer was a dynamic "Follow the Money" app. Here's how you can do it too.
Step 1: Start with one memo, not a giant data lake
I didn't start with "all data about the state budget." I started with one detailed memo, the kind staffers pass around and organizers squint at. The goal: make it readable, connect it to real context from Michigan.gov and turn it into something people could explore without being budget experts. No new platform. No data warehouse.
Step 2: Build a focused CustomGPT as the memo explainer
I connected a dedicated CustomGPT to the memo and the background notes I'd gathered. It had one job: help build a detailed, human-readable report on what's in the memo. Not a public chatbot. A specialist.
Because it was grounded in my own documents and links, it could find the right sites and data, unpack acronyms, translate legalese, call out which programs were hit and how, point to more information on state websites and cite everything.
Step 3: Use ChatGPT to turn the explainer into a blueprint
Next, I used ChatGPT to design how this should work as an app. We went back and forth: if people care about X and Y, what views should they see? How do they click into a program, an agency or a theme? What fields and filters let them explore without breaking anything?
The result was a clear set of instructions for a Base44 app: what tables, what fields, how the data connects and which views people would switch between. ChatGPT moved me from "here's what this memo means" to "here's how this should be structured."
Step 4: Use Base44 as the views-and-clicks layer
With the blueprint in hand, I fed Base44 the structured data the CustomGPT created and worked with its bot to build the views. My role was editor: "This filter goes on the left." "Let people toggle between these two views." "This label needs to be plain English, not budget jargon."
By the end, people could explore what the memo really did without ever reading the memo.
The bigger pattern: chain your tools, don't worship one
- CustomGPT: deep on your documents
- ChatGPT: turns understanding into a data and app design
- Base44: turns the design into something people can click through
No single system does everything. It may feel old-fashioned in the era of "one AI to rule them all," but it's easier to debug, easier to explain to your team and much easier for nontechnical staff to trust.
Why this matters for nonprofits
To decode government and policy language, most nonprofits don't need the newest, flashiest tool. They need budget memos explained in English, impacts broken down by community or program, and tools staff and volunteers can actually use.
You don't have to be a lab or a unicorn startup to build useful AI tools. Start with one memo, one document, one pain point. Then let AI move you from dense text to a clear explainer to a structured app to usable views.
That's all I did.
Originally published on LinkedIn, Dec. 22, 2025. Lightly edited.