I wanted to understand how people come to trust a voice AI, so I built one to find out. It did three jobs at once: it helped participants navigate, it interviewed them afterward, and it coded every session in the background. Six people used it for real errands over a few weeks. Trust didn’t fade slowly. It dropped the instant the assistant ignored something someone had already asked for.
TIMELINE
Spring 2026
ROLE
UX Researcher & Ethnographer
TEAM
Merlyn Koonamparampath, Eric Lopez, Conor Mack
FOCUS
UX Research
Cities are quietly rebuilding mobility around AI. New York alone plans to fold public transit into a single digital platform. But when we talk about accessibility, we still tend to mean curb cuts and crosswalks, not whether the people most exposed on the street will actually adopt the systems being built around them. Adults over 65 are about 15% of New York’s population and close to 45% of its pedestrian fatalities. Woven wanted to know what really happens when a mixed-age, mixed-mobility group plans real trips through a voice AI, while there was still time to change it before shipping.
The diary study instrument, built as a custom GPT with three simultaneous roles.
I ran it as a diary study, but not the usual kind. Normally you ask people to log an experience and then reflect on it later, from memory. I wanted to close that gap, so I built the reflection into the tool itself.
Six participants, ages 26 to 72, planned 24 real errands out loud: date nights, a contractor search, a family cruise, a dog walk. The same custom GPT that helped them plan would then turn around and interview them about how it went, inside the same conversation. When someone said “submit entry,” a Zapier automation exported everything to a master Google Sheet, where the assistant had already done a first pass of coding on the session. That first pass included six reflection ratings plus interpretive notes on emotion, confusion, and mobility friction. Personal details were generalized the moment they were captured, so a real address became “home base” in the structured record, though the instrument’s own rule was to never remove what someone actually said out loud, so a volunteered detail could still surface verbatim in a transcript. A manual QA pass caught that gap before analysis. Privacy here was layered, not absolute.
Which meant I never had to touch a raw, un-coded, un-anonymized transcript. The instrument did the first pass on itself.
A participant’s live navigation session inside the instrument.
Trust worked like a series of pass/fail checkpoints, not a running total. People handled real complexity fine. Subway closures, an unfamiliar city, none of it shook them much. What did was the assistant ignoring something they had already told it. One participant asked, more than once, to avoid a particular café and to just be given a location instead. The assistant kept routing her there anyway. Her reliance rating for that session dropped to 3/5, even though she rated the rest of it well. One dropped checkpoint was enough to undo several good exchanges. And she wasn’t the outlier. Four of the six participants hit some version of this. Jarrod, 49, learned the Home Depot the assistant first recommended wasn’t actually the closest one (“okay, why didn’t you tell me that the first time?”). Another participant’s single entry ended before the assistant ever delivered directions at all.
“I told you repeatedly that I didn’t want to go to Blanchet Coffee, and I had to also tell you repeatedly that I wanted a location, but you didn’t drop directions.” Hridaya, 26
Age split how people handled privacy, not how comfortable they were with the tech. The 26-year-old wouldn’t hand over her address at all, and navigated by landmark instead. The 72-year-old volunteered his full home address without being asked, and rated the session 5/5 straight through the assistant’s small mistakes. The same prompt, “what’s your address,” landed as a privacy risk for one person and as ordinary helpfulness for the other. Three segments came out of the data:
Confident Delegators: older, high baseline trust, happy to disclose. They want the assistant to just decide well.
Active Verifiers: younger, test and correct everything, protective of their data. They want fast, concrete answers.
Life-Domain Concierges: navigation is almost incidental. They hand the assistant a whole plan, from contractors to classes to travel, and expect it to keep up.
No single privacy default can serve all three.
People handed it a purpose, not a destination. Almost nobody opened with “how do I get to X.” They opened with a goal: a romantic dinner where the wife is vegetarian, a seven-day cruise routed through Venice and Rome, a trip to catch the cherry blossoms at peak bloom. The route came last, and sometimes not at all. A navigation-first design, where you type in an address and get directions back, would miss most of what people actually reached for this tool to do, which was to decide what to do in the first place.
The friction came from the interface, not the city. In one session the assistant misheard a participant’s age four separate times. It invented a café, “The Flatiron Café,” and presented it as a real place. It filled processing delays with rambling. For most people the errand itself was easy. The hard part was the assistant’s own behavior.
“You were just rambling. I would have liked a quick response.” Hridaya, 26
A participant correcting the assistant mid-session.
If trust fails at single checkpoints, then every unacknowledged error is the moment a user quietly decides to stop delegating. That’s the business case for what I handed over: a short set of recommendations for Woven’s research and product teams, ordered by impact against effort. The cheapest ones are retention features.
Say the correction out loud. A quick “got it, correcting that now” before quietly fixing a mistake. Trust drops at the moment an error goes unacknowledged, not at the error itself. (High impact, low effort)
Add a concise mode and a “still working” status. Cut the filler, and show progress on any lookup that runs longer than about five seconds. Several people asked for shorter answers when they were in a hurry. (High impact, low effort)
Make landmarks the default and addresses opt-in. The younger participant’s own workaround, navigating from “Southern Spice” instead of her front door, kept her private without costing her anything. Build that in as the default and it quietly protects the guarded segment for free. (High impact, medium effort)
I presented these in May 2026, at the end of the engagement. What Woven has shipped since isn’t something I can see from here, so I won’t claim adoption I can’t show.
A few honest caveats. These were errands people happened to be running, not scenarios I assigned, which makes them realistic but uneven. Entry counts ranged from 1 to 8 per person, and several reflection fields came back empty, so I read the ratings as directional rather than as clean numbers. And the privacy split I built the segments around could be about age, or it could be about digital literacy. This sample can’t tell the two apart.
The reason I still trust the pattern is consistency within each person. Whichever way someone handled the first privacy prompt or the first error, they handled every one after it the same way, across all of their entries. And the design implication holds at any sample size: the same address question read as a privacy risk to one person and as ordinary helpfulness to another, so a single default was already failing someone at six participants, and it would keep failing someone at six thousand.
The thread I want to pull next is trust recovery. After an identical mistake, does an Active Verifier’s skepticism reset within the session, or is it broken for the rest of it? The answer changes what you build. It’s the difference between designing to earn trust back once you’ve spent it, and designing so you never spend it in the first place.
This page is looked after by the Archivist, an agent I built that works from my primary research and can’t publish without me. Here’s how that works.
I wanted to understand how people come to trust a voice AI, so I built one to find out. It did three jobs at once: it helped participants navigate, it interviewed them afterward, and it coded every session in the background. Six people used it for real errands over a few weeks. Trust didn’t fade slowly. It dropped the instant the assistant ignored something someone had already asked for.
TIMELINE
Spring 2026
ROLE
UX Researcher & Ethnographer
TEAM
Merlyn Koonamparampath, Eric Lopez, Conor Mack
FOCUS
UX Research
Cities are quietly rebuilding mobility around AI. New York alone plans to fold public transit into a single digital platform. But when we talk about accessibility, we still tend to mean curb cuts and crosswalks, not whether the people most exposed on the street will actually adopt the systems being built around them. Adults over 65 are about 15% of New York’s population and close to 45% of its pedestrian fatalities. Woven wanted to know what really happens when a mixed-age, mixed-mobility group plans real trips through a voice AI, while there was still time to change it before shipping.
The diary study instrument, built as a custom GPT with three simultaneous roles.
I ran it as a diary study, but not the usual kind. Normally you ask people to log an experience and then reflect on it later, from memory. I wanted to close that gap, so I built the reflection into the tool itself.
Six participants, ages 26 to 72, planned 24 real errands out loud: date nights, a contractor search, a family cruise, a dog walk. The same custom GPT that helped them plan would then turn around and interview them about how it went, inside the same conversation. When someone said “submit entry,” a Zapier automation exported everything to a master Google Sheet, where the assistant had already done a first pass of coding on the session. That first pass included six reflection ratings plus interpretive notes on emotion, confusion, and mobility friction. Personal details were generalized the moment they were captured, so a real address became “home base” in the structured record, though the instrument’s own rule was to never remove what someone actually said out loud, so a volunteered detail could still surface verbatim in a transcript. A manual QA pass caught that gap before analysis. Privacy here was layered, not absolute.
Which meant I never had to touch a raw, un-coded, un-anonymized transcript. The instrument did the first pass on itself.
A participant’s live navigation session inside the instrument.
Trust worked like a series of pass/fail checkpoints, not a running total. People handled real complexity fine. Subway closures, an unfamiliar city, none of it shook them much. What did was the assistant ignoring something they had already told it. One participant asked, more than once, to avoid a particular café and to just be given a location instead. The assistant kept routing her there anyway. Her reliance rating for that session dropped to 3/5, even though she rated the rest of it well. One dropped checkpoint was enough to undo several good exchanges. And she wasn’t the outlier. Four of the six participants hit some version of this. Jarrod, 49, learned the Home Depot the assistant first recommended wasn’t actually the closest one (“okay, why didn’t you tell me that the first time?”). Another participant’s single entry ended before the assistant ever delivered directions at all.
“I told you repeatedly that I didn’t want to go to Blanchet Coffee, and I had to also tell you repeatedly that I wanted a location, but you didn’t drop directions.” Hridaya, 26
Age split how people handled privacy, not how comfortable they were with the tech. The 26-year-old wouldn’t hand over her address at all, and navigated by landmark instead. The 72-year-old volunteered his full home address without being asked, and rated the session 5/5 straight through the assistant’s small mistakes. The same prompt, “what’s your address,” landed as a privacy risk for one person and as ordinary helpfulness for the other. Three segments came out of the data:
Confident Delegators: older, high baseline trust, happy to disclose. They want the assistant to just decide well.
Active Verifiers: younger, test and correct everything, protective of their data. They want fast, concrete answers.
Life-Domain Concierges: navigation is almost incidental. They hand the assistant a whole plan, from contractors to classes to travel, and expect it to keep up.
No single privacy default can serve all three.
People handed it a purpose, not a destination. Almost nobody opened with “how do I get to X.” They opened with a goal: a romantic dinner where the wife is vegetarian, a seven-day cruise routed through Venice and Rome, a trip to catch the cherry blossoms at peak bloom. The route came last, and sometimes not at all. A navigation-first design, where you type in an address and get directions back, would miss most of what people actually reached for this tool to do, which was to decide what to do in the first place.
The friction came from the interface, not the city. In one session the assistant misheard a participant’s age four separate times. It invented a café, “The Flatiron Café,” and presented it as a real place. It filled processing delays with rambling. For most people the errand itself was easy. The hard part was the assistant’s own behavior.
“You were just rambling. I would have liked a quick response.” Hridaya, 26
A participant correcting the assistant mid-session.
If trust fails at single checkpoints, then every unacknowledged error is the moment a user quietly decides to stop delegating. That’s the business case for what I handed over: a short set of recommendations for Woven’s research and product teams, ordered by impact against effort. The cheapest ones are retention features.
Say the correction out loud. A quick “got it, correcting that now” before quietly fixing a mistake. Trust drops at the moment an error goes unacknowledged, not at the error itself. (High impact, low effort)
Add a concise mode and a “still working” status. Cut the filler, and show progress on any lookup that runs longer than about five seconds. Several people asked for shorter answers when they were in a hurry. (High impact, low effort)
Make landmarks the default and addresses opt-in. The younger participant’s own workaround, navigating from “Southern Spice” instead of her front door, kept her private without costing her anything. Build that in as the default and it quietly protects the guarded segment for free. (High impact, medium effort)
I presented these in May 2026, at the end of the engagement. What Woven has shipped since isn’t something I can see from here, so I won’t claim adoption I can’t show.
A few honest caveats. These were errands people happened to be running, not scenarios I assigned, which makes them realistic but uneven. Entry counts ranged from 1 to 8 per person, and several reflection fields came back empty, so I read the ratings as directional rather than as clean numbers. And the privacy split I built the segments around could be about age, or it could be about digital literacy. This sample can’t tell the two apart.
The reason I still trust the pattern is consistency within each person. Whichever way someone handled the first privacy prompt or the first error, they handled every one after it the same way, across all of their entries. And the design implication holds at any sample size: the same address question read as a privacy risk to one person and as ordinary helpfulness to another, so a single default was already failing someone at six participants, and it would keep failing someone at six thousand.
The thread I want to pull next is trust recovery. After an identical mistake, does an Active Verifier’s skepticism reset within the session, or is it broken for the rest of it? The answer changes what you build. It’s the difference between designing to earn trust back once you’ve spent it, and designing so you never spend it in the first place.
This page is looked after by the Archivist, an agent I built that works from my primary research and can’t publish without me. Here’s how that works.
I wanted to understand how people come to trust a voice AI, so I built one to find out. It did three jobs at once: it helped participants navigate, it interviewed them afterward, and it coded every session in the background. Six people used it for real errands over a few weeks. Trust didn’t fade slowly. It dropped the instant the assistant ignored something someone had already asked for.
TIMELINE
Spring 2026
ROLE
UX Researcher & Ethnographer
TEAM
Merlyn Koonamparampath, Eric Lopez, Conor Mack
FOCUS
UX Research
Cities are quietly rebuilding mobility around AI. New York alone plans to fold public transit into a single digital platform. But when we talk about accessibility, we still tend to mean curb cuts and crosswalks, not whether the people most exposed on the street will actually adopt the systems being built around them. Adults over 65 are about 15% of New York’s population and close to 45% of its pedestrian fatalities. Woven wanted to know what really happens when a mixed-age, mixed-mobility group plans real trips through a voice AI, while there was still time to change it before shipping.
The diary study instrument, built as a custom GPT with three simultaneous roles.
I ran it as a diary study, but not the usual kind. Normally you ask people to log an experience and then reflect on it later, from memory. I wanted to close that gap, so I built the reflection into the tool itself.
Six participants, ages 26 to 72, planned 24 real errands out loud: date nights, a contractor search, a family cruise, a dog walk. The same custom GPT that helped them plan would then turn around and interview them about how it went, inside the same conversation. When someone said “submit entry,” a Zapier automation exported everything to a master Google Sheet, where the assistant had already done a first pass of coding on the session. That first pass included six reflection ratings plus interpretive notes on emotion, confusion, and mobility friction. Personal details were generalized the moment they were captured, so a real address became “home base” in the structured record, though the instrument’s own rule was to never remove what someone actually said out loud, so a volunteered detail could still surface verbatim in a transcript. A manual QA pass caught that gap before analysis. Privacy here was layered, not absolute.
Which meant I never had to touch a raw, un-coded, un-anonymized transcript. The instrument did the first pass on itself.
A participant’s live navigation session inside the instrument.
Trust worked like a series of pass/fail checkpoints, not a running total. People handled real complexity fine. Subway closures, an unfamiliar city, none of it shook them much. What did was the assistant ignoring something they had already told it. One participant asked, more than once, to avoid a particular café and to just be given a location instead. The assistant kept routing her there anyway. Her reliance rating for that session dropped to 3/5, even though she rated the rest of it well. One dropped checkpoint was enough to undo several good exchanges. And she wasn’t the outlier. Four of the six participants hit some version of this. Jarrod, 49, learned the Home Depot the assistant first recommended wasn’t actually the closest one (“okay, why didn’t you tell me that the first time?”). Another participant’s single entry ended before the assistant ever delivered directions at all.
“I told you repeatedly that I didn’t want to go to Blanchet Coffee, and I had to also tell you repeatedly that I wanted a location, but you didn’t drop directions.” Hridaya, 26
Age split how people handled privacy, not how comfortable they were with the tech. The 26-year-old wouldn’t hand over her address at all, and navigated by landmark instead. The 72-year-old volunteered his full home address without being asked, and rated the session 5/5 straight through the assistant’s small mistakes. The same prompt, “what’s your address,” landed as a privacy risk for one person and as ordinary helpfulness for the other. Three segments came out of the data:
Confident Delegators: older, high baseline trust, happy to disclose. They want the assistant to just decide well.
Active Verifiers: younger, test and correct everything, protective of their data. They want fast, concrete answers.
Life-Domain Concierges: navigation is almost incidental. They hand the assistant a whole plan, from contractors to classes to travel, and expect it to keep up.
No single privacy default can serve all three.
People handed it a purpose, not a destination. Almost nobody opened with “how do I get to X.” They opened with a goal: a romantic dinner where the wife is vegetarian, a seven-day cruise routed through Venice and Rome, a trip to catch the cherry blossoms at peak bloom. The route came last, and sometimes not at all. A navigation-first design, where you type in an address and get directions back, would miss most of what people actually reached for this tool to do, which was to decide what to do in the first place.
The friction came from the interface, not the city. In one session the assistant misheard a participant’s age four separate times. It invented a café, “The Flatiron Café,” and presented it as a real place. It filled processing delays with rambling. For most people the errand itself was easy. The hard part was the assistant’s own behavior.
“You were just rambling. I would have liked a quick response.” Hridaya, 26
A participant correcting the assistant mid-session.
If trust fails at single checkpoints, then every unacknowledged error is the moment a user quietly decides to stop delegating. That’s the business case for what I handed over: a short set of recommendations for Woven’s research and product teams, ordered by impact against effort. The cheapest ones are retention features.
Say the correction out loud. A quick “got it, correcting that now” before quietly fixing a mistake. Trust drops at the moment an error goes unacknowledged, not at the error itself. (High impact, low effort)
Add a concise mode and a “still working” status. Cut the filler, and show progress on any lookup that runs longer than about five seconds. Several people asked for shorter answers when they were in a hurry. (High impact, low effort)
Make landmarks the default and addresses opt-in. The younger participant’s own workaround, navigating from “Southern Spice” instead of her front door, kept her private without costing her anything. Build that in as the default and it quietly protects the guarded segment for free. (High impact, medium effort)
I presented these in May 2026, at the end of the engagement. What Woven has shipped since isn’t something I can see from here, so I won’t claim adoption I can’t show.
A few honest caveats. These were errands people happened to be running, not scenarios I assigned, which makes them realistic but uneven. Entry counts ranged from 1 to 8 per person, and several reflection fields came back empty, so I read the ratings as directional rather than as clean numbers. And the privacy split I built the segments around could be about age, or it could be about digital literacy. This sample can’t tell the two apart.
The reason I still trust the pattern is consistency within each person. Whichever way someone handled the first privacy prompt or the first error, they handled every one after it the same way, across all of their entries. And the design implication holds at any sample size: the same address question read as a privacy risk to one person and as ordinary helpfulness to another, so a single default was already failing someone at six participants, and it would keep failing someone at six thousand.
The thread I want to pull next is trust recovery. After an identical mistake, does an Active Verifier’s skepticism reset within the session, or is it broken for the rest of it? The answer changes what you build. It’s the difference between designing to earn trust back once you’ve spent it, and designing so you never spend it in the first place.
This page is looked after by the Archivist, an agent I built that works from my primary research and can’t publish without me. Here’s how that works.