Social Benefits

An AI-first benefits workflow invented from nothing in two weeks, designed across both requestor and fulfiller.

A social benefits application process put its heaviest burden on the people least equipped to carry it. Applicants abandoned at verification and document collection, and agents received cases they had to reassemble before they could assess them. This was an internal concept, so there was no instance to work in, no requirements to inherit, and nothing to react to. The assignment was to show how emerging AI capability could change the shape of the workflow, and to show it convincingly in two weeks.

MY ROLE

Design Lead

TIMELINE

2 weeks

SCOPE

Open scope, capability question focus

I treated mobile intake and the fulfiller workspace as one workflow

The obvious version of this project is an applicant-facing intake experience, and that version fails. Every shortcut taken on the applicant side lands as unverified input on an agent's desk and gets paid for there. I designed the two as two sides of one workflow: conversational, AI-assisted mobile intake paired with a fulfiller workspace that receives its output. Extraction happens once, at collection, and arrives as a summarized case with its sources attached.

Constituent and case agent personas

Verification came first because nothing downstream holds without it

Identity was the first bottleneck and the one that made the rest tractable. I sequenced it ahead of everything else and designed it as a consequential moment rather than another form step, using a third-party service so the applicant clears identity once and the workflow can then trust it. Downstream validation and assessment are designed against a resolved identity, which is what lets the agent's view be a summary rather than an interrogation.

Final updated journey

I designed for the gaps, not the screens

Two weeks does not buy a complete state inventory, and pretending otherwise produces a deck that is thorough and unconvincing. The work also had to live inside an existing public-sector playbook pattern rather than proposing a new framework. I selected screens against the friction points: the moments where the capability gap was visible and where the concept changed the outcome. The journey carried the narrative and the screens proved the intervention.

Final presentation demo

What I'd revisit

The concept assumes extraction is good enough to summarize a case an agent will act on. I did not scope what happens when it is not: no confidence surfacing, no fallback path, no correction cost for the agent. I would now design the failure state alongside the success state, because in a benefits context a wrong summary is more expensive than a slow one.