Under H.R.1, many Medicaid members will be subject to community-engagement requirements, but a large share will qualify for an exemption: caregivers of young children, people with a disabling condition, pregnant members, those in substance-use treatment, students, and others depending on how each state codifies the rules. The problem is that most members will not know which category they fall into. An "Am I exempt?" assistant answers one question in plain language: does this rule apply to me, and if not, why not?
Start from the decision tree, not the database
The temptation is to bolt a chatbot onto the eligibility system and let it query records. Resist that first. Begin by mapping the exemption logic as an explicit decision tree, one branch per exemption category, written from the official state rule text. Each branch should reduce to a yes or no question a person can answer about their own life: Are you currently pregnant? Do you care for a child under a defined age? Are you in a treatment program? This tree is your source of truth, reviewable by policy and legal staff before a single line of interface is built.
Keep the questions short and sequential. Asking five plain questions one at a time outperforms one dense form. At each step, show the member where they are and what happens next.
Three design rules that build trust
First, never give a false definitive. If the assistant cannot confirm an exemption with confidence, it should say so and route the member to a human or to the formal attestation path, not guess. A wrong "you are exempt" is far more dangerous than an honest "let us verify this."
Second, always explain the result. "You appear to qualify for the caregiver exemption because you reported caring for a child under six" teaches the member the rule. A bare yes or no teaches nothing and invites doubt. The explanation also doubles as a record the member can screenshot and reference.
Third, hand off cleanly to action. Determining exemption is not the goal; staying covered is. If the member is exempt, tell them exactly what to do to record it. If they are not, tell them what reporting they owe and by when. The Arkansas experience showed that confusion, not ineligibility, drove the bulk of the roughly 18,000 coverage losses. An assistant that ends in clarity rather than a dead end is the entire point.
Finally, measure the assistant like any other intervention: completion rate, the share of sessions that end in a confirmed next step, and downstream retention among users versus non-users. Build it before the June 30 to August 31, 2026 notice window so members have a place to turn the moment the first notices land.