Conversation meets structured action.
After chat-only failed the hard tasks, I turned Surfaces into the working model: persistent conversation, structured UI, source evidence, and existing product capabilities in one experience.
The system used the simplest form that could support the task.
I studied emerging AI products, including Mindtrip, a trip planner that paired chat with a side panel. That reference helped define three levels of interaction.
Markdown and inline UI
Definitions, concise explanations, suggestions, citations, and low-complexity choices remain in the transcript.
Evidence and artifacts
Documents, benefit IDs, account status, and relatively static rendered content open beside chat.
States and widgets
Forms, tables, adjustments, comparison, and multi-step workflows use deterministic interface patterns.
Users should not need to understand which Nayya app completes the work.
Unlike the platform approach, where users had to choose into multiple apps for different features, that routing was no longer needed. Titan could coordinate capabilities and data already distributed across the ecosystem. It could make API calls and render assets in the moment, without additional reroutes or exits.
The strategic question was how much of the underlying app architecture should remain visible to the employee, when it should be visible, and when it should simply be expressed in chat.
Chat remained persistent
I defined a responsive panel system for navigation, conversation, and Surfaces. Panels could expand, shrink, split, or occupy most of the workspace while keeping stable dimensions between minimized and expanded states. With AI chat as the consistent piece across every view, users stayed grounded in the same product regardless of which capability was in play, and could always ask a question.
They could inspect a document or complete a task without losing the conversational context that brought them there, and shrink chat when they needed the Surface larger for the document or the work itself.
Prioritizing features
Rebuilding the whole platform with Leave, Choose, Use, Retire, and Claims at once was not the right move. We still had a lot to learn and test, and fully recreating the ecosystem's flows would have taken far too long. As a team, we prioritized what made sense to prove first.
Because the first sprint could not integrate every app, we started with benefits shopping and recommendation through Choose, the most-used feature in the ecosystem. Traffic was seasonal, but that also meant more potential to pilot, sell, and reach users in our existing client pool. Choose was the clearest use case for dynamic Surfaces: guided enrollment, profile and family data, recommendations, plan comparison, and enrollment summary all needed structured UI beside persistent chat.
Trust was the product problem, not a disclaimer.
Trust is the largest problem AI has to solve in any product. In a B2B2C benefits platform, accuracy reaches past Nayya: it affects the credibility of HR and the employer. Nayya’s moat has always been accurate relevance. Recommendations draw on employer, user, and third-party data about conditions and medical history. Titan could also learn from conversations and interactions across the ecosystem. That context could increase accuracy, and it raised the bar for how trust had to be designed.
Trust still had to be earned. We approached it in three ways.
Employer source of truth
Work with employers so the documents and data Titan treats as truth are approved and current, not improvised guidance.
Say what AI cannot answer
Be transparent when the system does not know or should not answer, and redirect people to support instead of guessing.
Calculations and treatment
Anything involving calculation or treatment states two limits: this is not medical advice and a clinician should confirm care decisions; and cost estimates can be wrong because of delayed claims or payments, network coverage, or the procedure codes a provider uses.
Titan tested better than Jupiter because it could help people do more than read.
Luckily we had a client willing to run this as a pilot. It was a small client, but I would take that over another round of random usability tests at this point in the project. Thankfully the feedback pointed in the right direction. Clients saw a substantial improvement in interactivity, task support, and collateral sharing. The City of Bullhead pilot added evidence from administrators and employees using the product in context.
Chat plus UI improved utility
Clients viewed Titan as a meaningful step forward in interactivity, task completion, and collateral sharing.
AI products need history
Users brought expectations from mainstream AI tools, including conversation history and easy return to prior material.
Evidence must be easy to inspect
Pilot feedback reinforced source validation and a more recognizable AI presence.
Capability implies action
Once users saw the product as intelligent, they expected it to help with claims, caregiver support, and other actions.
The strongest outcome was a product direction the organization continued to invest in.
From chat feature to ecosystem
Titan reached production pilot and established a model for combining conversation, evidence, artifacts, and deterministic workflows.
A shared AI-native strategy
The work influenced positioning, follow-on Choose Native investment, a lighter concept for legacy clients, and distribution exploration.
A durable interaction principle
The interface should become more structured as consequence, information density, and task state increase.
Full commercial rollout paused while the team added core Choose and benefits-shopping capability and strengthened the platform architecture. Recruiting more pilots was difficult because of the implementation effort required on the client side. Many wanted to stay on the older infrastructure for an open enrollment that was already close, which meant no extra work on their end. That was a time and workload constraint more than a concern about employee readiness, privacy, trust, or safety. Priorities shifted to getting the AI model we had built into the legacy product for this OE, fast, while planning to onboard clients onto the new AI ecosystem the following year. The result is directional validation and organizational commitment, not quantified adoption or recurring engagement.
AI should not erase the interface. It should know when the interface is the clearest form of help.