AI Concierge: trust in a regulated flow

The challenge

Health insurance is confusing, and shoppers have real questions. An AI assistant can help, but in a regulated industry the cost of a wrong or overconfident answer is high, and people need a clear path to a licensed human.

Key decisions I drove

1. Scoped the real work. Clarified the difference between "add a chat button" and designing a chat experience, so the team planned for the right effort.
2. Defined the journey. How shoppers enter AI chat, and how they move from AI to a live agent without starting over.
3. Reused before reinventing. Advocated for the company's established AI chat pattern instead of a new one, for consistency and speed, and worked out how engineering could reuse the existing components.
4. Unblocked engineering. Identified design-system and implementation dependencies early and got engineers access to the design files.

Designing for trust

I pushed the team to answer the questions that decide whether people trust an AI in a high-stakes moment:

• Disclosure: does the shopper know they're talking to AI?
• Privacy and retention: what happens to what they type?
• Prompt testing: how do we know answers hold up before launch?
• Escalation: when should AI step aside for a human?

The design team followed up with research on introductory prompts and quick-start options, so shoppers who don't know what to ask still get a useful start.

Good AI design is knowing where AI belongs, and where it doesn't.