Augintel Assist
Teaching a generative AI model when to say "I don't know."
Platform
B2B SaaS - GovTech
My Role
Lead product designer
With
PM, ML, CS, eng.
Timeline
~13 months
The OPPORTUNITY
Too many notes. Too little time. Too much at stake.
Augintel is case management software for child welfare agencies. A single active case can accumulate over a hundred notes covering visits, court filings, service updates, and risk flags, often times written by different workers. New workers inheriting a case face the steepest climb: hundreds of entries and a looming deadline.
I led design on Augintel Assist, a conversational AI feature letting workers ask plain-language questions about a case, grounded only in that case's own notes. The risk was obvious from day one — this isn't a domain where a model can guess. Getting it wrong in a court report has real consequences for a family.
Discovery
I surveyed workers and interviewed six supervisors before we built anything.
Alongside early surveying, I ran an ideation workshop with stakeholders to understand where AI could create the most value, then interviewed six supervisors at a county child welfare agency to understand what would make an AI summary trustworthy enough to use before a hearing.
Survey data
What Supervisors said
What We learned
01
Excitement, with conditions
Every supervisor was enthusiastic, but tied their enthusiasm directly to editorial control. They wanted a draft to review, not a finished answer to submit.
02
Omissions are the real risk
"It can really vary. We could be putting a child back at risk." Supervisors were less worried about the AI inventing things than quietly leaving something out.
03
"Where it came from" matters as much as "what it says"
Several interviewees raised court admissibility unprompted. A generated summary needs to trace to a source or it's unusable.
Act 3 — The solution
AI had to show its work
The interface design choices were almost all in service of one goal: making it safe for a supervisor to use in a workflow that ends in court. Research gave us four clear principles to design against.
Key decisions
Each one tied directly to a supervisor research finding.
The Trade off
I chose guided, structured input over natural language
The obvious direction for a conversational AI feature is a free-text box. But early model testing revealed a consistent failure pattern: open-ended queries produced unreliable scoping. The model would answer about the wrong person, collapse timeframes, or synthesize across contexts it shouldn't combine.
The guided approach wasn't a UX preference. It was a strategy for constraining the problem space early, so the model could produce reliable answers within a bounded scope. Each structured interaction also generated clean, labeled training signal — the path toward natural language input, not a detour from it.
The Flow
Welcome state → topic chips → disambiguation → cited draft answer with feedback controls → time range control
