Product Innovation
Can you build a conversational interface in a week?
How we learned to build a conversational interface for the Freeletics AI Coach
Chatbots are everywhere.
Facebook Messenger. Intercom. Navigator. Hugging Face.
Until this summer, I was mostly uninterested in bots, mainly because they are relegated to customer support and sales tasks. Essential tasks for any business, but not the most interesting for me.
What changed?
I attended F8 and drank the Messenger kool-aid. A lot of it. It also didn’t hurt that I met some cool guys from Spectrm, Berlin’s premier conversational marketing platform.
The engineers at Facebook are building great tools for businesses to connect with their customers in a more human way. One new feature is Personas, which enables an exceptional experience for customers with multiple bots serving the conversation as needed.
And I wanted to explore this.
I found the tools to build my own bot and pitched the idea at our summer Hackathon. Thankfully, a few people were interested, and we kicked off the project later that day.
In addition to myself, the team now had two engineers, a data analyst, a training/nutrition scientist, and a PR specialist.
Goals
We kicked off the first day by setting a clear purpose for the week by asking: what did we want to get out of this week, and what should we accomplish? It took a little debate, but we quickly aligned on building a conversational interface for our existing customers using the AI Coach.
We tossed out onboarding and getting started, which are arguably just as valuable. We did this because we aligned on the most significant user problem: I don’t know what to do here.
We believed if we could solve this at one stage of the customer journey, our system could easily address this question at other stages of the product experience.
Principles
As we debated, I captured the outcomes and turned those into our project principles. Doing this saves the team time in the future and codifies what is essential. If you are unfamiliar with this, I highly suggest the book Principles by Ray Dalio. Here are a few from the week:
- Humans First — we optimize for how people naturally communicate, not how machines prefer input.
- Personal — when possible, we personalize the communication based on what people assume is reasonable (a Coach knows your preferences and history).
- Clear — we make the capabilities clear because people do not know what to expect from a conversational interface.
- Escape Hatch — don’t let people get lost, always provide a way out.
Ideation
Throughout the afternoon, we ideated on the possible questions a person would want to ask their coach. After a few minutes, patterns emerged: actions, context, topics, and more.
We noticed that our brainstorm of potential requests mapped across two dimensions — general to personal, information to action. For example, a general action could be “what are the new journeys,” while a personal information request could be “what’s a healthy snack today.”
Discovery
We wrapped the afternoon ideation with some clear next steps: we all needed to understand conversational interfaces better. From what it took to build one to what people expect when using them.
The engineers landed on Google’s Diagflow as the foundation for the project. Of course, there are several other platforms out there: AWS Lex, Wit.ai, and others.
Prototype Test
While engineers read through platform documentation, I started interviewing people. My goal was to understand consumer expectations and unmet needs, as well as validate the overall user experience we intended to build. I quickly made a prototype with Marvel and a test on UserTesting.com.
A few hours later, I had a mountain of learnings and a few surprises. Generally, I learned:
- Training, nutrition, and mindfulness are the most important topics for these participants
- How to build a schedule around fitness, nutrition, or sleep/rest is almost as important as the topics themselves
- The participants don’t know what they could ask an AI Coach and relied on helper UI to introduce the potential options
- Participants were delighted by personalization and going a step beyond with content recommendations from our blog
- Some participants struggle with motivation and reaching their goals (no surprise here, we all do) — but want to solve that first before having a conversation with the AI Coach
- Some participants wanted to treat the AI Coach like a real person and have “meetups” to discuss their progress and next steps
Tuesday — Thursday
There is no recipe to the rest of the week. It was just a lot of hard work.
Every. Single. Day.
Other than a morning standup, we spent the week building and collaborating. No formula, no meetings, no framework.
Friday
This week is a reminder of something I believe in: you can build anything in a short amount of time.
Assemble a team and bring the right mindset. Have a bias for action. Trust each other to deliver. And work hard.
On Monday, we all knew almost nothing about conversational interfaces, the problem we were going to solve, or how we would work together.
By Friday afternoon, we had it working in our Android app and were experts inside our company when it comes to building conversational interfaces. Here’s to what happens next…
Keep calm and ship it.
