I design products in which several types of users work within a single system. Each type of user makes decisions with different consequences. For example, on a student-health platform I worked on, an admin could cancel a student’s insurance. On the same system, a clinical partner could approve or reject medical documentation. An insurance broker could run enrollment and manage eligibility. Students could upload the data on which their compliance status depends.
While all four of these audiences work on one platform, their experiences of the product are very different. In this article, I’ll walk you through how we made these processes work for each of these types of users.
Defining Roles Through Permissions and Responsibility
The platform that I designed served several universities, healthcare partners, insurance brokers, and students at once—around 25,000 users in total. The hard part was building safe workflows for all of these user groups at the same time, when each group has its own permissions and its own HIPAA and FERPA requirements. I built each role around its jobs-to-be-done. A single user interface that covers every audience at once is not the answer. Let’s consider each role in turn.
Students have a specific list of required actions they need to get through: enroll in insurance, file a waiver request, upload immunization records, check their status, respond to follow-up requests, and find the information they need quickly. They sign into the system a few times a semester and have no time to work out how it is put together. For this group, I built a simple, legible user interface to make using the product feel easier.
Admins spend entire days working in the system and handle large volumes of data, so they need access to everything at once: compliance dashboards, review queues, waiver-cycle configuration, billing workflows, audit trails, filters, bulk actions, overrides, and detailed student records. I tuned the user experience for efficiency, control, and making precise decisions fast.
Insurance brokers on the platform control enrollment, manage student eligibility, and work with policy workflows and administrative controls. Their user interface is transactional and data driven.
Clinical partners make medically grounded decisions about documentation. The cost of getting this wrong is higher than for any other role. I gave these users clear context, the history of each record, and enough information to reach an accurate judgment.
Right now the platform has two distinct user types at the permission level: student and admin. Everything beyond that, including how a single user could hold more than one role, is still a work in progress, so I won’t describe it here as if it were already solved.
Champion Advertisement
Continue Reading…
Proportional Friction: Matching Safeguards to Consequences
Within any high-volume operational environment, admins make mistakes. The UX designer’s job is to ensure that these mistakes are reversible, traceable, and cheap. In healthcare and education this is critical. How much friction an action requires depends on how serious the consequences are when it goes wrong. Rejecting an immunization record is reversible, so a clear confirmation with a mandatory reason might be enough. Canceling a student’s insurance carries much heavier consequences and can affect their coverage, billing, and access to care. An action like this requires the stronger protection of giving explicit confirmation and a choice of effective dates. An admin can apply a cancellation immediately if a student requests it or schedule it for a later date. The system notifies the student either way.
The second question is how someone recovers from a mistake. I designed high-risk actions as state changes, with a full audit trail showing who made the change, when, and what exactly changed. The record remained available in the system, so an admin could see what the previous state was and an auditor could follow the whole chain of events.
All of this describes a mistake from the admin’s perspective. But there is a second perspective—that of the student on which such decisions land. If students believe a decision in their record is wrong, they need a clear path to request a review or correction. In regulated environments, the path to challenge a decision must be designed with the same care as the admin’s working screens.
The least-privilege principle applies here, allowing only those with a matching role and permissions to perform high-risk actions.
Watching Users’ Behavior Change Outside the Interview Room
One assumption determines both the way I split the roles and the amount of friction that I built in: at the moment that users act, they behave the same way they describe their own work. There is a gap between these two things, and it shows up only when people actually use the product. The project that showed this to me most clearly was a different project—a biometric-identity platform, on which users place their hand on a physical device to confirm who they are.
The data on this platform is extremely sensitive, so trust, reliability, and the prevention errors matters greatly. I designed the onboarding, authentication, device management, recovery flows, and enterprise administration. User interviews and usability walkthroughs showed whether the workflow was clear, but left out the environment in which the person was using the device. More than 20,000 devices had been deployed around the world, tens of millions of scans had gone through them, and they were running in very different circumstances.
So we conducted field research, watching people in the places where they use the devices. We looked at what people do not say in an interview—whether they hesitate before the interaction, whether they know where to put their hand without being told, how a queue behind them affects them, and how staff invent their own ways of explaining what the device does not make obvious. None of these signals would come up during a remote session because users do not register them as problems.
What was most surprising was how people behaved in regard to trust. In one environment, they walked up to the device with open skepticism about privacy and biometric data. In another, where a corporate pilot was running, the resistance was noticeably lower. The same product and a completely different trust dynamic.
Two changes came directly out of that research. The camera and scanner positions assumed a posture that people did not naturally adopt. So after observing real user interactions, we adjusted those positions and changed the on-screen guidance along with the hardware: what the user sees during a scan and what happens when a scan fails.
The second change came from a false assumption we had made about instructions. We had assumed that users would read them before using the device, but our field observations showed the opposite: most people formed their understanding during their first interaction with the hardware and skipped instructional content entirely. Therefore, instead of presenting all guidance up front, I surfaced it progressively as users moved through the interactions. So system feedback became contextual rather than instructional.
The research showed that context and perception shape adoption barriers as much as user-interface design. It also surfaced something I would not have known to ask about in a research script: people come to a product with their own ideas of how it works, with fears and cultural assumptions that might never appear in moderated testing.
Deciding How Much Autonomy to Give AI
The same principle that I used to distribute safeguards across actions works with artificial intelligence (AI): the heavier the consequences, the less the autonomy. On the student platform, users work with sensitive data about health, insurance, and identity, so asking what can AI do is the wrong question to ask here. The right question to ask is what can AI do safely—within clear boundaries and with the user checking the result.
For example, students have contextual questions all the time: what their insurance covers, what the waiver requirements are, which immunizations they need, what to do about symptoms, how to book an appointment, and what the university policy says. Ordinary navigation and a static FAQ (Frequently Asked Questions) answer questions such as these with documents, but what the student actually needs is a direct answer. The domain is narrow and controlled, so this is the main condition: the AI interprets the user’s question and returns an answer in plain language that it has drawn from approved institutional content. Diagnosis, medical decisions, changes to records, and acting on the student’s behalf are outside that domain.
I first defined the boundaries: which precise problem the AI solves, which sources it can use, and what information is explicitly closed to it. In this case the model is limited to curated institutional content, so the open Internet is not among its sources.
Next, the user interface must set the right expectations for the user, so we designed clear AI states, contextual disclaimers, source transparency, low-confidence behavior, and an always-visible path to live support. In healthcare, this is fundamental, because users lose trust fast when a system sounds more confident than it is.
I designed verifications at several levels. A user can trace an answer back to its original source, and reviewing interactions shows where the system answers incorrectly and which materials are missing from the base. Anything that changes data or affects eligibility or compliance status stays within a workflow with human review, permissions, and audit control. In regulated products, AI is most reliable where it helps people make sense of information.
As a Senior Product Designer at Volta Health, Anna architects mission-critical, regulated healthcare and secure-identity systems.
Read Anna’s blog on MediumRead More