014 of 4 participants
The warning icon
"!" now means action is required. When nothing is needed, the state is neutral.
Back to work
OutputMobile app and web prototype
RoleSolo product designer
Research4 moderated sessions · June 2026
Project scopeAI interaction model
Confidence and escalation
Usability testing
Hebrew RTL
ZCHUT-AI
I designed and tested a prototype that turns an Israeli benefit letter into one plain-language next step, and says when the AI is unsure.
About the project00The problem
People can qualify for support and still never claim it. In three Israeli benefit programs, more than half of those eligible did not.
53%Long-term income support
61%Unemployment benefit
57%Rent assistance
Eligible people who never claimNational Insurance Institute, Gottlieb, 2021. Pre-COVID data
01Explain the letter before asking for action
Scanning a letter and asking by voice converge on one interpretation layer. The system explains what the document means before it asks for a decision, and ends in either a specific action or a clear reason nothing is needed.
01
Guided flow
Known rules and deadlines
- Focused question
- Structured check
- Clear action
02
AI interpretation
Letters, voice and open questions
- Natural input
- Interpretation
- Clarify or escalate
Every answer carries
An official source, a confidence state, or a way to continue with an adviser.
Two interaction modelsGuided where the rules are known, AI where the input is open
- 01
Says what it sees
The scanner names the letter before the photo is taken.
- 02
A way around the camera
A photo from the gallery works too.
02Every answer says how sure it is
A confidence gate decides how the answer is framed. When it escalates, the system says why and passes the conversation to an adviser, so the user does not repeat the story. The cost: more interface states, and a dependency on adviser tools.
ConfidentAct with a source
Answer and official source
PartialVerify before acting
Answer with a verification step
EscalateContinue with a person
No answer, route to an adviser
The confidence gateThree outcomes after interpretation
03What four sessions changed
I ran four moderated sessions with Wizard-of-Oz responses. Participants used the AI to work out what to ask, not as a final authority on money, so the source, the verification step and the handoff stay visible in every response.
- 01
Count first
The headline says how many items need action.
- 02
"!" means act
The warning badge appears only where action is required.
- 03
Specific action
The button names the task: sign the form.
- 04
AI on offer
Help sits beside the work, not in front of it.
022 of 2 immigrants
Disability in onboarding
Disability became a first-class onboarding category.
034 of 4 participants
Source placement
The official source sits inside the answer, not in a footnote.
04Secondary pattern
Search and eligibility labels
Voice entry moved earlier, and eligibility lists say "all" or "any".
Before and afterFour of the five changes
04How I used AI
I used AI to scaffold prototype states and edge cases. I set the product rules, the research questions and the validation boundary.
What this does not prove
Four qualitative sessions are not statistically representative. Wizard-of-Oz responses tested comprehension and trust, not model or legal accuracy. Without a baseline, I can explain the changes but not quantify a gain in comprehension.
