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.
People can qualify for support and still never claim it. In three Israeli benefit programs, the gap was larger than half.
National Insurance Institute research series, Prof. Daniel Gottlieb with the Hebrew University School of Social Work, published in Social Security, May 2021. Pre-COVID data.
"I just want to know what I'm entitled to without feeling like I need a law degree to use an app."
"I can build a bridge, but I can't decipher this one-page letter from the tax office."
The interface explains the document before asking the user to take action.
Known rules and deadlines
Letters, voice, and open questions
Every answer includes an official source, a confidence state, or an option to continue with an adviser.
The app confirms the transcript before the AI answers. Voice entry moved earlier after usability testing.
The scan separates routine information from deadlines and required action, without asking the user to retype the letter.
The system explains why it is escalating and passes the conversation to an adviser, so the user does not repeat the story.
This requires maintaining a second interaction pattern.
This adds more interface states.
This depends on adviser tools and governance.
I used AI to scaffold prototype states and edge cases. I set the product rules, research questions, and validation boundary.
Simcha, 76 - former chemical engineer; Tami, 72 - former kindergarten teacher; Yevgeny, 58 - musician with a mobility disability; Genia, 54 - musician whose husband is disabled.
"What does this '!' want from me? If there's nothing to submit, why is it there?" - Tami
I changedReserved "!" for required action and used a neutral state when nothing was needed.
"Ask me if I have a disability - that's the most important thing to me." - Yevgeny
I changedAdded disability as a first-class onboarding category instead of forcing a workaround.
"608 isn't a round number someone made up. It looks real." - Tami
I changedPut the official source inside the answer, not in a footnote.
I sawSearch categories missed users' language, and condition lists did not communicate "all" versus "any."
I changedMoved voice entry earlier and labelled the eligibility logic explicitly.
Four qualitative sessions are not statistically representative. Wizard-of-Oz responses tested comprehension and trust, not model or legal accuracy.
The usability tests informed four interface changes. Model accuracy, legal correctness, and governance were outside the scope of testing.
The strongest gap surfaced by accident.
A short isolated test would have exposed it.
I can explain the changes, not quantify comprehension improvement.