Viv: an AI health coach for Virta

The problem

New members struggle most in their first weeks on the protocol, when questions about food and devices come up constantly. The question they asked coaches most often was "can I eat this?" Human coaches can't be there around the clock, and some moments need a human, not a bot. The real question wasn't whether we could build a chatbot. It was how to give members instant, clinically safe help without weakening their relationship with their coach.

The design approach

Viv gives clinically safe answers, and hands the moment to a human coach when a member needs a person rather than a bot.

Help, don't nag. I decided Viv would not try to coax adherence. An AI hasn't earned the trust to push or motivate someone, so pressure from it reads as nagging. Viv answers, teaches, and celebrates. Motivation stays with the human coach.
Progress over perfection. With the head of the clinical program, I set a rule: aim for small improvements rather than pushing members to hit the clinical ketosis threshold of 0.5. A coach that fixates on a number makes people feel like they're failing.

I explored practical flows in low fidelity first.

Low-fidelity Food Checker intro with an avocado illustration, offering a photo of a restaurant menu, a photo of food, or a barcode

Low-fidelity Food Checker chat where the member types that they are going to have chicken tacos from a restaurant, with the keyboard open

Low-fidelity loading screen that says Serving your food insights soon

Low-fidelity Chicken Tacos result showing total carbs, protein, and low-carb substitutions such as a lettuce wrap or cheese shell

Low-fidelity follow-up chat about what keto taco shells are made from and whether they are low carb

An AI coach lives or dies on its dialogue, so I built a working prototype early and tested real conversations instead of polishing static screens. I set up Virta's user testing from scratch, got leadership to adopt UserTesting.com, and coached junior designers in the method.

The flows

Then I added photo-based convenience to food checker. It's a practical AI answer to the question members asked coaches most often: can I eat this?

Virta home screen greeting Darcy, with a cyclist in the blue header, a tap on Food Checker, tracking rings, and a full Articles row

Viv chat asking if a food is Virta-friendly, with options to photograph a restaurant menu, photograph food, or scan a barcode

The same Viv chat with a Select a Photo sheet offering Take Photo or Choose from library

Camera framing an Applebee's menu, asking to fit the full menu inside the frame while Viv captures it

Viv recommending the House Sirloin with substitutions and estimated carbs and protein

Coach handoff: when a member feels like they're failing, Viv reassures them and brings in their human coach rather than trying to motivate them itself.

Virta home screen greeting Darcy, with a cyclist in the blue header and a tap on the Check your ketones to-do

Viv chat where the member feels like they're failing after ketones dropped, and Viv reassures them and sends the conversation to coach Audrey

Sync confirmation that one meter reading synced, showing ketones of 0.2 today and 0.7 yesterday

What I learned

50% of new members used Viv in their first week.
Answers to common questions went from hours to seconds.
More members set their diet start date within 30 days.

"I really like using an AI assistant in here. Yep... excellent." (UserTesting participant)

"I never knew that before. Thanks Viv!" (Virta member, beta)

Conversation proved extremely useful for education and for answering the long tail of members' very different questions. Proactive conversational push notifications did not work. They got about a 5% response rate, against about 12% for our standard reminders. To avoid burning out the notification channel, we stopped them. The lesson: members come to Viv when they need an answer, and pulling them in works better than pushing.