A mixed-methods study at Homage Senior Services into why partner agencies weren't confirming their listings — synthesized with AI assistance, stress-tested against a contradicting interview, then scoped into a redesign I can actually defend.
Homage Senior Services is a nonprofit connecting older adults and people with disabilities in Snohomish County, WA to community resources that support independent living. As the resource database curator, I manage 800+ partner agency listings and run annual outreach to verify their information is still accurate.
Agencies confirmed vs. Database Curator (DC) updated, by month — all six months on record.
DC = Database Curator, the role responsible for keeping listings accurate when agencies don't respond.
| Month | Listings | Agencies confirmed | DC updated |
|---|
When agencies don't verify directly, accuracy falls back on secondary sources, like agency websites, which can be outdated or incomplete. This matters beyond operational efficiency: the database feeds directly into Washington Community Living Connections, a public portal older adults and caregivers use to find services. A wrong phone number or outdated program listing can mean a missed connection to care.
17 of 27 respondents (63%) said "No" or "Maybe" when asked if they'd ever received the verification email.
My working hypothesis going in: the email itself was the friction point — too dense, and easy to distrust. This case study is as much about testing that hypothesis honestly as it is about the redesign itself.
Surveyed partner agency contacts to understand whether the verification email was even being noticed, and if not, why.
Talked with three survey respondents to go deeper on what made the email easy to overlook or distrust — one of whom, notably, had no complaints at all. More on why that mattered under Stress-Testing.
Built three personas from behavioral patterns across the survey and interviews — covered below — to keep the redesign accountable to real behavior.
Reading two full transcripts against 27 survey responses to find real patterns is slow by hand. I used Claude to produce a first-pass synthesis — themes, supporting quotes, an insights table — then reviewed that draft against my own transcripts, line by line, and corrected it before treating any of it as finished.
I anonymized interview transcripts and survey responses before feeding them to AI (Claude) for synthesis — identifying details were stripped first, not after. Quotes, personas, and findings shown here stay anonymized throughout, to protect the people who spoke with me.
Gave Claude the research plan, both transcripts, and the raw screener data — no pre-filtering — and had it structure the material into themes tied to my original research questions.
Read the draft back against the actual transcripts rather than accepting it at face value. Two of the four things I corrected weren't the AI being careless — it read the transcripts accurately, but missed context only I had from actually running this program.
| # | What the AI draft said | What I changed, and why |
|---|---|---|
| 1 | Framed our partner network as primarily "senior services," with a dental association treated as an outlier. | Our database covers food banks, utilities, housing, disability services, and more — there's no single "typical" agency. Reframed as a structural issue, not an edge case. |
| 2 | Said a participant "wasn't sure at first why Homage was contacting her." | She actually understood the partnership — her predecessor had explained it. Her comment described how a future staffer without that context might react. Reframed around institutional-memory risk, not confusion. |
| 3 | Treated her "this has nothing to do with dentistry" comment as mainly a request for friendlier copy. | The more concrete ask underneath it was proof of impact — usage data. Pulled that out as its own opportunity instead of letting it get flattened into a copy tweak. |
| 4 | Wrote the contact-turnover finding as something this project should fix. | Solving org-wide nonprofit staff turnover isn't something an email redesign can do. Scoped it down to what's actually actionable: a forwarding line. |
AI is fast at first-pass pattern-matching across transcripts. It has no memory of this program or these people. The synthesis got faster; the accuracy still depended on knowing my own project well enough to catch a plausible-sounding read that was quietly wrong.
"The shorter the e-mail, the more chance you have of somebody completing reading the full email... if you make it too wordy, they'll get through part of it and just be like, oh, I've got to go do something else."— Interview participant, caregiver support nonprofit
Three patterns across the survey and interviews — not proof of a market segment with two data points, but a way to keep the redesign accountable to real behavior instead of an averaged-out "user."
"If you make it too wordy, they'll get through part of it and just be like, oh, I've got to go do something else."
Hesitated on the review link despite trusting Homage — phishing-awareness training makes any unsolicited link feel risky.
Called the email "fine" in interview — then rated it only "somewhat clear" on the survey and flagged a technical issue.
My third interview didn't fit the story I was building. I almost left it out of the synthesis entirely — that would have been a mistake, and it's worth naming directly rather than smoothing over.
That participant said the email was fine. No complaints. But their survey answers told a slightly different story — rating the email only "somewhat clear" and flagging a technical issue, a gap between what someone says out loud and what they click that's a finding in its own right, not noise to explain away.
More importantly, it pushed me to re-read the survey data as a whole instead of through the lens of my two most engaged interviewees. Across all qualifying responses, the most-cited barrier was lack of time, not confusion — only a small minority called the email too dense.
A better email reduces friction for agencies who already engage with it. It's unlikely to be the reason most non-responders don't respond — that's more plausibly about priority and time than clarity. The redesign below is worth shipping because it's low-cost and evidence-backed, not because I expect it to fix response rates on its own.
It's easy to blur "the content is evidence-backed" with "the design is done" into a single claim. They're not the same thing, and treating them as one risks calling something finished that's only half-checked. So: two tracks, kept separate on purpose.
A redline of the current email, not a full visual overhaul. Every change below is traceable to a specific finding. Drag the divider to compare, and hover any marker for what changed and why. One thing is deliberately left out — phishing fatigue eroding link trust needs a different deliverable (a separate pre-notification email, sent a few days ahead), not a copy change here.
Separately, I explored what an on-brand version could look like — same redlined copy at two fidelities, so structure and content are checked before any visual polish gets applied. Low-fi carries the real content in a plain skeleton; hi-fi adds Homage's actual colors and photography on top of the same words. A wide-layout variant and further polish exist, but they're layout detail, not new signal, so they're left out here.
No one has reviewed or tested this yet — no design critique, no usability pass. The design direction itself I'd call done; whether it actually works for the people using it is still an open question, not a footnote. That's the next step before it ships, not after.
Three interviews against a goal of 3–5. These findings are directional, not conclusive — worth validating with more survey respondents before finalizing anything.
The email redesign is a small lever. My actual takeaway is that partnering with data analysis — real numbers on referrals and listing outcomes, not just a better-written ask — is what would move this. That's what makes the case measurable, and what would make agencies more engaged with the database long-term.
Compressing the mechanical first pass — reading, pattern-finding, structuring — so I could spend my time on the part that needed judgment: knowing my program well enough to catch a plausible-sounding but wrong read, and knowing when to stop scoping a "fix" past what an email can actually do.