Unlocking AI

The pace at which AI tools are moving is astounding. My team has fully embraced Figma Make — Figma’s AI feature built on Anthropic’s Claude — and we haven’t looked back. It lets designers and non-designers alike turn a prompt into a coded prototype. We even trained Product Managers to use it, so they can explore concepts creatively without needing a designer’s help. We’ve used it to quickly prototype ideas and user-test hypotheses on interactive experiences without needing developers at all.

We’re also partnering with our development team to connect our Figma designs to a repository of design tokens and our coded React components, so that eventually Figma Make can build using our exact components — letting developers implement new pages and elements in a fraction of the time.

There are plenty of other ways we lean on AI for design work:

• Analyzing a screenshot and providing design feedback
• Suggesting edge cases we may not have considered
• Recommending ways to better engage users
• Creating slides for presentations
• Image manipulation
• Assisting with research plans
• Analyzing research findings
• Persona development
• Content writing
• Brainstorming concepts
• And many others, I am sure

I’ve watched weaker designers become much stronger by embracing AI tools, which lets them provide more value and, in turn, increases their own value. It’s exciting to see how quickly these capabilities are improving, and how much faster they let us work. What does the future hold for design teams?

One place AI is already having a material impact is in customer support within our call centers. We recently rolled out an AI Call Screening agent that can answer every call even at our busiest — like during the Medicare Annual Enrollment Period. In past years, eHealth pulled people from across the company to screen calls before transferring customers to licensed agents, and even with the extra help, thousands of calls went unanswered. With the AI call screener we now have the capacity to answer every call, and it performs just as well — if not better — than a person. Research with customers showed they’re both surprised it’s AI and pleased with the experience; we’ve found the right balance of empathy while keeping the caller on task.

We’re now training an AI Sales Agent to field customer calls and answer the simple questions callers ask most often, transferring to a real agent whenever a question gets more complicated. Everyone’s excited to see how this goes.

Another project piloting generative AI is loading insurance plan content from our carrier partners. Every year hundreds of carriers send plan data for thousands of plans with no standard format, so up to now we’ve had to manually process and review each one to parse the data — hundreds of hours across dozens of people, done well but slowly, with the occasional mistake.

By leveraging AI, we import the plan data using two different models and use a third to check the work and flag where the data differs. Instead of hours per plan, the data is consumed and processed in minutes, with more accuracy — freeing up most of those people to work on something else.

After launching a few customer-service AI projects, we wanted to help our online users find and enroll in a health insurance plan directly. Our primary customers are shopping Medicare plans during the Annual Enrollment Period, and Medicare is an overly complicated program that puts serious cognitive load on our users. To find a plan that actually fits, they typically add six doctors and ten medications, then still have to analyze dozens of different plans — a tall order given the cognitive capacity, motor skills, eyesight, and other accessibility factors of the average Medicare beneficiary.

We already had chat for users to reach a sales support advisor, so we didn’t want to build another version of that. We wanted something that kept users on the website itself, so we built an AI Shopping Assistant.

It turns out AI is currently bad at reading websites and struggles to understand what’s going on even after scraping the HTML and reading the DOM. What AI is genuinely good at is talking to APIs, so we built an experience that lets users add their doctors, drugs, and filters entirely through the AI. Our first concept was a chat interface, and we plan to expand that to conversational voice soon — imagine simply telling an AI, “I need a plan that covers Dr. John Smith in San Francisco, and it needs to cover these medications — oh, and I really like Sutter Health, so only show me their plans.” A frictionless way to narrow down the options.

It feels like we’re only scratching the surface with AI, so it’s exciting to see where things go from here. I know it’s going to let us build some amazing digital experiences for our customers.

None of these projects were about using AI for its own sake. Each one started from a real constraint — calls we couldn’t staff for, plan data we couldn’t process fast enough, a shopping experience too complex for a form — and AI was the tool that removed it. That’s the standard I hold the whole team to: AI earns a place in the product or the workflow by solving a specific problem better than the alternative, not by being new. As the tools get better, I expect that standard to let us take on problems we couldn’t have justified solving before.