When a national mission-driven organization needed to understand what actually moves people to give — and why its most committed donors give for different reasons than its occasional ones — SmartThink set out to answer it with rigor, and to do something newer: let AI carry the heavy work of reading hundreds of pages of qualitative responses without losing the nuance that makes qualitative research worth doing.
The challenge
The organization wanted to know the drivers of donation across three populations that differed in how much they gave — heavy, moderate, and light donors. Knowing which drivers mattered, and how strongly each one mattered to each group, would let leadership tailor its message and put effort where it would move giving the most.
The design
SmartThink built a two-stage study. An online qualitative phase identified the candidate drivers and gauged the strength — the sentiment — attached to each within each donor group. An online quantitative phase then modeled the magnitude of that sentiment for every driver in every group, turning rich language into something measurable.
The qualitative phase was engineered for exactly this. Skilled moderation, backroom support, a consistent set of topic questions, exhaustive responses, and exact transcription produced clean, comparable material. The questions were deliberately written so that each approached the same behavior from a different angle — which meant the same underlying "why" could legitimately surface as an answer to several questions. The more often a driver surfaced across the fifteen questions, the stronger its pull.
The hard part: teaching AI to read
The real work — and the part worth presenting — was getting AI to identify drivers and assign sentiment reliably across the three groups. That took four attempts:
| # | How the data was fed in | Result |
| 1 | One file per segment; summarize all 15 questions for each | Under-summarized; too much detail |
| 2 | One file per question across segments; summarize the differences | Little integration across questions |
| 3 | 45 files, one per segment-and-question; summarize the differences | More detail, but terms no longer comparable across groups |
| 4 | Feed attempt 1's output back in; ask only for differences between segments | "Just right" — consistent terms across groups, distinct findings within each |
The breakthrough was a two-step move: pre-process the material topic by topic, then analyze that processed output for between-group differences — rather than asking the model to do both at once.
What we found
Ranking each driver by how often it surfaced across the fifteen questions produced a clear, segment-specific picture. The drivers that topped the list for heavy donors were not the ones that mattered most to light donors; one driver showed up everywhere, while others were unique to a single group.
| Driver | Heavy | Moderate | Light |
| "K" | 1st | — | — |
| "X" | 2nd | — | — |
| "Y" | 3rd | 3rd | 3rd |
| "P" | 4th | — | 4th |
| "L" | — | 1st (tie) | — |
| "V" | — | 1st (tie) | — |
| "Z" | — | 2nd | 2nd |
| "M" | — | — | 1st |
Drivers ranked by how often each surfaced as a "why" across the 15 questions, within each donor segment. Labels are anonymized to protect the client.
The recipe
- Data collection. Skilled moderation, backroom support, consistently delivered topic questions, exhaustive responses, and exact transcription.
- Topic selection. Make each question a different facet of the behavior under study, so the same response is a valid "why" answer to more than one question — which is what lets frequency stand in for strength.
- Analysis in two steps. First pre-process the material topic by topic; then analyze that output for differences between groups.
Why it matters, and what's next
The same recipe makes modeling behavioral drivers faster, cheaper, and easier to repeat. It also points toward something larger: AI "avatars" that reproduce how an entire target segment would respond — so a message can be tested before it ever reaches a real audience.
The value of research isn't what it reports. It's the clarity it creates and the decisions it makes possible.
Presented by Timothy Gohmann, PhD, at the MASB Winter Summit (February 2026). Client identity and driver labels are anonymized; methodology and findings are shared with permission.