The Setup
A founder asked me how to build a better form. The real answer had nothing to do with forms at all.
Guide to Good (G2G) helps nonprofits build clear, compelling profiles that attract serious donors, something many organisations struggle to do alone. What started as a request to help design a client intake form turned into a discovery process spanning two of G2G's founding months, because the real bottleneck was never the questions on a page. It was the conversation happening in the founder's head every time she spoke with a client.
My role
As UX Consultant and Strategist, I pushed back on the original brief, designed and facilitated the discovery process, and translated a founder's tacit decision-making into material her CTO could use to train an AI agent.
Impact
The conversational logic and decision states from this research now power G2G's AI agent, currently in internal testing. Since then, G2G has raised $250,000 in funding, with a further $970,000 pledged by donors after presenting at the Skoll World Forum at Oxford University in April 2026.
My Research Flow
Research Intake
My initial chat with the PM on a new piece of work.
Planning & Strategy
Hashing out the brief, meeting stakeholders, signing off on objectives, and recruiting.
Research Execution
Running interviews, testing sessions, and other data collection.
Analysis & Synthesis
Coding, triangulating, and turning raw data into insights.
Knowledge Sharing
This is the decision-making phase, when deliverables and action items are shared with stakeholders.
The Problem Behind the Problem
The founder came to me with six sharp questions about form design. Something told me every one of them was aimed at the wrong target.
G2G's founder had spent months running one-to-one sessions with nonprofit clients herself, manually building their profiles by hand. It worked, but it didn't scale. As her waitlist grew, she and her CTO started building a digital form that an AI agent would read, rewrite, and turn into a client's profile.
She came to me with practical questions: how to phrase them, how to order them, what to do when someone couldn't answer, when a form gets too long, whether it should change shape for different types of nonprofits. They were all good questions. They were also all questions about a solution nobody had tested yet. So instead of answering them, I asked one of my own.
The Investigation
That last question was clearly pointing somewhere: a mapping exercise. Some people call it as-is journey mapping. I call it cognitive mapping, because what we were really mapping was a conversation, and all the invisible decisions happening inside it.
(Yes, this meant delaying the form-design questions she'd originally asked me. That felt like a risk this early in a consulting relationship. It also felt like the only way to avoid building the wrong thing beautifully.)
Research questions
- Are there "types" of nonprofits? What are they?
- How did she run client sessions, before, during and after?
- How did she ask, order and clarify her questions?
- Did sessions ever run too long, or clients get frustrated?
- What worked, what didn't, and where are the opportunities?
Who was in the room
- Founder (CEO), the domain expert who'd run hundreds of client sessions
- Co-Founder (CTO), who would build and train the AI agent from what we learned
A successful outcome would...
...surface the founder's decision-making clearly enough that her CTO could translate it into training material for an AI agent, replacing guesswork with evidence before a single line of the product got built.
The Method: Cognitive Mapping
I call this method cognitive mapping rather than as-is journey mapping, because of what's actually being mapped: a conversation between founder and client, where dozens of small decisions were happening entirely inside the founder's head. We reviewed recordings, documentation and messages from her past client sessions together, then mapped them out step by step, across two 1.5-hour remote workshops on Miro, with a possible third session held in reserve.
Why Bother, If a Form Might Work?
I've had this pushback before. Why run a whole discovery process if the answer might just be a form? Two reasons. First, this was G2G's core product, not a side feature, and nobody should wing their core product without de-risking it with evidence first. Second, I already suspected a long form wasn't the right shape for this kind of interaction, but a suspicion isn't proof. I needed to show the founder why, not just tell her.
Lucky for me, the founder loves a discovery process as much as I do, and trusted me enough to follow it.
Tools
In the Room
I facilitated the sessions using a mapping template I'd built. We started broad: a general conversation about how she ran client sessions, which immediately showed that structure was driven by the type of organisation she was speaking with.
From there, we identified four organisation archetypes: single stream and focused, multi stream and focused, diamond in the rough, and unfocused. For two to three real client examples per archetype, we mapped exactly what she did from the moment she became aware of an organisation to the moment she finished, or chose not to finish, their profile. That was how we learned her micro-decision points, and the logic driving them.
We kept going until the patterns repeated themselves. The founder noticed it before I had to point it out, which is always the best possible sign.
She was following the same strategic logic in session after session, and she could finally see it laid out in front of her.
Connecting the Dots
The form was never really the product. The conversation was.
Across every session, the founder was doing far more than asking questions in order. She was running communication strategies, before and during each call, to sequence questions, form and test hypotheses, decide when to pivot, flag an answer as not good enough, and judge when an organisation simply wasn't ready yet. Capturing that accurately meant the interaction had to stay flexible and dynamic. In other words, it had to stay a conversation.
That's when the direction became clear: the right solution was a simple form for uploading documents, paired with a conversational agent to draw out the real narrative behind each client's organisation.
The hard part
Conversational AI tooling was shifting almost weekly at the time, and the CTO's own tools and practices were evolving right along with it. The open question became: what's the best way to hand off everything we'd learned so he could start training an AI agent immediately, and how do you turn something this cognitive and human into technical material an LLM can actually learn from?
Neither of us had a clean answer ready-made, so I condensed our maps into a working document for each archetype, then used Claude to help shape them into:
- A classification system and weighting for client document types
- Descriptions for each nonprofit archetype
- Profile-building stages and key steps within each archetype
- Defined cognitive states: hypothesis, fact, observation, success or failure, pivot
- The final documents themselves, ready for the CTO to use in AI training
What Came Out of It
What emerged wasn't just a conversation script. It was a decision tree, complete with conditional logic, states and strategies the founder had refined for so long they'd become entirely subconscious to her. Surfacing them gave her CTO everything he needed to start training the AI agent for real.
Synthesised insights and recommendations
From maps to material
What the mapping sessions directly produced:
- Four nonprofit archetypes, each with two to three real examples
- Step-by-step profile-building stages per archetype
- Defined cognitive states: hypothesis, fact, observation, success or failure, pivot
- Document classification and weighting logic
What the CTO built from it
How the CTO operationalised the maps, in his own words to me:
- "State Cards", one per stage of a nonprofit's profile
- Each defining what the agent must do, must never do, and what unlocks the next state
- Micro-states like "emotional touchpoint" and "standards not met" built directly in
- A format I'd now love to reuse with other teams training AI models
Looking Back
This project was a reminder that the best answer to "how do we build this form" is sometimes "let's find out if a form is even the right question." Pushing back on a well-meaning brief felt risky in the moment, especially early in a consulting relationship. It paid off because the founder was just as curious about the real answer as I was.
It also reinforced how much expertise lives in someone's head without them realising it. The founder had been running the same sophisticated decision logic for months. She needed a structured process, not just a conversation, to see it clearly enough to hand it off to someone else.
How I'd approach this today
I'd keep the same evidence-first instinct, but formalise the archetype-to-state-card pipeline earlier. Now that I've seen how well cognitive maps translate into an LLM's training material, I'd design the mapping workshop itself with that translation in mind from the very first session, rather than figuring out the handoff format after the fact.