QUALITATIVE RESEARCH · AI-ASSISTED METHODOLOGY
Should I Say Yes to This?
A qualitative study of the moment a vet recommends something optional, and how pet owners decide whether to say yes. Three semi-structured interviews plus an AI-simulated pilot, with the reasoning behind every research decision made visible.
00 — At a glance
One line.
A study of the exact moment a pet owner has to say yes or no to something extra at the vet, and whether an AI assistant would help.
What this page proves.
Not a validated concept. A study reasoned in the open: a research question narrowed to a specific decision moment, a proto-persona built specifically enough to be falsified, an AI pilot that improved the guide before it went live, and three interviews whose findings rewrote the concept direction when the original hypothesis didn't hold.
01 — The Assignment
A class assignment on an AI-infused concept.
The brief was to run a small qualitative study around an AI-infused concept: two or three interviews, plus an executive deck. I chose vet visits because the "optional service" moment felt like an under-examined decision that AI might help with. The concept I described was a phone assistant that would explain a recommendation in plain English, give typical cost, describe what happens if skipped, and suggest questions to ask the vet.
A methodology that mirrored the concept.
The class also asked me to pressure-test the interview guide with an AI-simulated participant before fielding. So the study used an AI on both sides: one helping pet owners make better vet decisions, one helping me make a better interview.
02 — Narrowing to a Moment
The research statement started too big.
I began at "vaccinations and wellness," which is a topic, not a research question. A topic that broad gives an interviewer nothing to steer against. Every participant would arrive with a different definition of what the study was about, and I'd end up with three loose collections of anecdotes instead of three focused stories.
The pattern I actually wanted to study was smaller.
I've watched friends and family agree to something at the vet, walk to the car, and immediately second-guess themselves. Dental cleanings, senior bloodwork, non-core vaccines, supplements. The clinical validity of any given recommendation isn't the interesting question. The decision experience is.
Two decisions inside the statement.
First: I'm studying the decision, not the medicine. Second: I'm looking specifically for breakdown points, which means the interview needs to probe emotional and financial aftermath, not just the exam-room moment. Both of those choices shaped every question that followed.
Typeset · research statement, as fielded
"How do pet owners decide whether to accept the optional services their vet recommends, and where does that decision break down enough that an AI assistant could help?"
03 — Maya Reyes, v2
Proto-persona v1 didn't survive its own critique.
The first version was demographic plus attitude: a stressed millennial dog owner who cares about her pet and worries about cost. Everyone I know fits that. A critique pass called it "a composite of attitudes, not a person," which is exactly what it was. A persona at that level of abstraction confirms whatever the interviewer already believed.
V2 became a specific fictional person.
Maya Reyes, 38, assistant project manager at an architecture firm in Cleveland. Single, rents a 1BR in Tremont. Currently paying off a $2,200 emergency vet bill on Affirm, two months left. Her dog Charlie is a 7-year-old shepherd mix, adopted from a county shelter. She's seen three different vets across five visits at the same clinic, so there's no relationship carrying the recommendations. Her most recent visit produced $1,200 in pending charges. She said yes in the room, opened the estimate in the parking lot, and cancelled the dental two days later at a $50 fee.
The specifics were doing design work.
Charlie's age determined which optional services were even on the table. The lack of vet continuity meant Maya couldn't lean on trust. The $2,200 Affirm bill was a concrete financial weight, not "has a budget." The decision pattern named three different design surfaces: the exam-room moment, the parking-lot moment, and the 11pm-Google moment. Each is a different product. The persona forced a choice about which one.
The point of a proto-persona is to be wrong specifically.
Being wrong in specific ways is more useful than being vaguely right, because you can see where reality diverges. Maya predicted parking-lot regret. None of the three real participants had it. That was a useful finding, and I could only get it because Maya was specific enough to be falsified.
Proto-persona v2 · Maya Reyes
- Role
- Assistant project manager at an architecture firm in Cleveland.
- Household
- Single, rents a 1BR in Tremont.
- Financial
- Paying off a $2,200 emergency vet bill on Affirm. Two months left.
- Pet
- Charlie, 7-year-old shepherd mix, 55 lbs, adopted at age 2 from a county shelter. Mostly healthy. Mild seasonal allergies, graying muzzle, a little slower on walks.
- Vet
- Tremont Veterinary Care, 18 months in. Three different vets across five visits. No relationship with any one of them.
- The recommendation
- March wellness visit. Dental cleaning ($800 to $1,100), senior bloodwork ($175), joint supplement (about $45 a month). $1,200 in pending charges.
- Decision pattern
- Said yes in the room. Opened the estimate in the parking lot. Cancelled the dental two days later at a $50 fee. Did the bloodwork. The supplement is on the counter, unopened.
- Where the friction lives
- The parking lot. Or 11pm on her phone. Not the exam room, because by then she has already said yes.
- Deliberately left out
- Partner negotiation, multi-pet coordination, her tech relationship with the clinic. A proto-persona carries what bears on this decision, not everything that could exist about a person.
04 — Preparing the Moderator
Two moderator traps, named in the guide before fielding.
This was my second qualitative study ever. The first taught me that in the moment, I default to reassurance. When a participant says something vulnerable ("I'm probably overthinking it," "I felt like a bad pet owner"), the instinct is to comfort. Reassurance closes the door. Probing keeps it open. I wrote a reminder into the top of the guide: when you hear vulnerability, say "tell me more" instead of "that makes sense." The other trap was closed paraphrasing. "It sounds like you're saying X, right?" is a yes/no question with the answer inside it. The reminder: open it back up.
Claude as the pilot participant.
The class asked for an AI pressure test. I gave Claude the Maya persona in full detail and ran the interview as if she were a real participant. The instruction was to behave like a real person: vague answers, going off-topic, missing the question I actually asked. Then break character and debrief.
The debrief surfaced patterns I couldn't have seen from inside my own guide.
Claude flagged two moments where I'd reassured instead of probed, and one where I'd paraphrased a participant into agreement instead of asking her to expand. Neither was in the traps I'd already written down. Both were reasons the participant had given a short answer when a long one was available. The guide got a third reminder before the first live interview. Whether the concept itself would survive contact with users was what the rest of the study existed to find out; the methodology had already justified itself.
Interview guide · Moderator reminders, pre-pilot
- Reassurance
- When the participant says something vulnerable ("not great," "bad pet owner," "I'm probably overthinking it"), the instinct is to comfort them. Resist it. Probe instead: "Say more about that." Reassurance closes the door; probing keeps it open. The vulnerability is the research.
- Closed paraphrasing
- If you find yourself saying "it sounds like you're saying X, right?" that is a yes/no with the answer already inside it. Open it back up: "Tell me more about what you mean."
- On the guide itself
- The guide is a map, not a script. If a participant hands you a thread, a phrase, a moment, a small detail, pull it before moving on. The threads they don't expand on themselves are usually where the deck quotes live.
05 — When the Plan Met Three People
Michaela, dog Lila and cat Misty.
Michaela's current vet at West Park Animal Hospital is thorough. When they recommend an optional vaccine, they explain which lifestyles it fits and let her decide. Her most recent optional-service moment was three non-core vaccines for dog parks and daycare. She declined the two she didn't need, took the one she did, and never thought about it again. This was the interview where I first suspected the concept might not have a use case for anyone with a good current vet. The interview also almost failed. In the last five minutes, when I asked if there was anything I hadn't covered, Michaela mentioned Misty's dental surgery at her previous vet: they hadn't fully explained why it was needed, she felt pushed into it, and she went through with it anyway. That's the exact moment the concept was designed for. It arrived in the wrap-up because I hadn't asked the right way to surface it earlier.
Amy, dog Freya.
Amy takes Freya to Banfield. The most recent optional-service moment was a flea/tick/heartworm alternative to Simparica Trio. Amy declined because the alternative required more frequent dosing and she preferred the once-monthly convenience. Frictionless in the room. What made her interview valuable was her position on trust. She doesn't trust the internet, and when she wants to check something, she looks for real people with lived experience or a journal article geared to vets. Brand-published content is bias by default. Amy was the first participant to name what a trustworthy source looked like, and her answer wasn't the vet.
Sean, dogs Pearl and Poncho.
Sean is a designer and self-described AI power user. The optional-service moment was leptospirosis in Colorado: declined the first time based on the dogs' apartment lifestyle, revisited and accepted when the family moved to a property where the dogs were outside more. Mature and lifestyle-calibrated. Sean gave the sharpest trust answer of the three, categorizing pet health as a "your money or your life" domain and describing exactly what he'd want to see: citations at the individual claim level, credentials of sources, review by actual vets over time. He also raised a design constraint that neither previous participant had touched: using the app in the exam room would be socially awkward.
What n=3 can and can't tell you.
Three interviews are enough to see contrast between participants, not to generalize about pet owners. What n=3 shows is where a hypothesis holds across obviously different people and where it doesn't. All three had good current vets. All three said "before" for timing. Two of three wanted source-transparent trust; one wanted the AI to match the vet. The signal is worth acting on directionally. It's not a claim about the population.
06 — Three Patterns
The concept's value depends on the vet.
All three participants described their current vet as thorough enough that the concept felt redundant against them. The value case appeared only when they thought about a previous vet, or a vet they didn't trust. The concept is a corrective for bad vets, not a general assistant for the veterinary experience. If it launches without saying so, most users will find it useless most of the time against the good vet they already have.
Typeset · Michaela on her old vet
"I think that would have been very helpful with our old vet. They were just kind of tacking things on because they could. And I was uneducated."
Before is the only viable surface.
All three said they'd use the assistant before the appointment. Nobody wanted it after. Sean explicitly ruled out in-room use as awkward, and the socially specific reason he gave (openly checking your phone against the professional in front of you) generalizes past his particular case. The proto-persona was built around parking-lot regret and 11pm-Google. Zero of three had that pattern. Whatever this product becomes, it belongs before the appointment.
Typeset · Sean on using the app in the room
"If I was in the room, I probably wouldn't use it. That would probably be awkward for a lot of people. If it says something, you're like, let me just check my app. I'm gonna need to double-check you on that. You stand right there."
Trust wants source transparency, not answer authority.
Amy and Sean both located their trust in provenance. Amy in real people with lived experience and journal articles for vets. Sean in cited claims, named institutions, credentialed reviewers. Neither wanted the app to tell them what to do; they wanted the app to show them what it was drawing on, so they could decide. Michaela's position was the minority: she wanted the app's information to match the vet's, which effectively cancels the value case, since the concept only matters when the vet is wrong or incomplete. Two of three point toward a design direction that resolves both problems at once: a source-transparent question-generator rather than an answer-giver, with visible provenance behind every question.
Typeset · Sean on what trust would look like in the UX
"I would want to know where an individual line that's making a claim, where that source is actually coming from and how those types of things are curated. Is it coming from the National Vet Institute, or is it coming from Joe Schmo's blog about his pet one time?"
07 — What It Changed
Maya needs rebuilding, or a second persona.
The parking-lot pattern that structured Maya's proto-persona didn't appear once in three interviews. Vet quality, which Maya's clinic-hopping was meant to background out of the picture, turned out to be the primary variable determining whether the concept has any value at all. A next-round persona anchors on vet quality directly, or the study needs a second persona anchored to a trusted vet, to test the concept's range against both cases.
The concept moves from answer-giver to source-transparent question-generator.
The original description was a phone assistant that explains recommendations in plain English and gives you the answer. Three interviews reframed that. Two participants want to see the sourcing, not the conclusion. Positioning the product as a preparation tool that generates questions worth asking, with visible sourcing behind each one, resolves the trust paradox and the awkwardness objection in a single move. Sean also raised a bigger reframe worth naming: an async model where routine questions are answered by the app and only novel or high-stakes ones escalate to a vet. That's a different product. It deserves its own study.
What I got better at.
Between interview one and interview three, I started catching my own patterns in the moment. In the second, I called out mid-interview that I was struggling to keep questions open-ended and worked through the rephrasing out loud. In the third, I recognized leading questions faster and backed off them. The skill isn't asking perfect questions; it's noticing what I'm doing while I'm doing it.