Atlas AI Part 3: How the hell did we end up here?!
January 2021 - May 2022
This period was the most painful to write about - it’s like we lost touch with reality. Huge kudos to a mentor who kept hammering on us to do things differently.
This is the third post in a series about our failed startup, Atlas AI. If you haven’t already read the previous posts, you can find them at:
Everyone’s got a Gantt Chart ‘til they get punched in the face
How the hell did we end up here?!
Context: Two years in and back to Square One
This is the third post in a series about our failed startup, Atlas AI, which we worked on for four years. If you haven’t yet read the first post, start here.
We pick up the story here after two years. It’s January 2022 and we’ve just finished the prestigious Techstars accelerator. We were much better connected among tech companies and investors. Underneath this rosy exterior, we were back to square one. After Techstars helped us see the fatal flaws in our business model, we had no product and no customers.
Six months with zero users (Jan - June 2022)
We came up with a new idea to reuse our existing technology, inspired by the running app Strava. There are hundreds of millions of people following workout videos at home, who have no similar app. Given they’re following a workout video, they’re already in front of a camera - either their laptop or phone. We repackaged our AI into a laptop app, which would help people automatically track their home workouts and share results with friends. Genius - again!
We did some “market research” in the first few months of the year, speaking to 20 people about home workouts. Among other questions, we asked “Would you like to track your progress and share it with friends?”. Many of them said they would, so we started building a laptop app for people following home workout videos. We made the app free initially because Strava, like most fitness apps, has a free tier. In the future, we’d charge for a premium tier with a monthly subscription.
Despite the app being free, and many improvements over several months, we had a grand total of zero active users. People we’d spoken with during our “market research” didn’t want to use it, despite their previous indication of interest. We struggled to think of anyone else we knew who would try it. This was demoralising, but perhaps unsurprising: we didn’t do home workouts ourselves.
Four months with hundreds of users (July - November 2022)
In July, we recognised having an app with no users was a dire place to be. We took drastic measures, like banning our entire 5-person team from writing any code. We only allowed work that could lead to more people using our app at least once. We had no idea how to get users, so we flung a bunch of stuff at the wall, desperately hoping that something would stick:
We signed up for classes at HIIT studios, then asked other customers whether they were doing home workouts too.
We reached out cold to YouTube fitness influencers offering to pay them to promote our app in their workout videos
We asked strangers following workout videos outdoors in parks to try our app
We started posting on online communities for home workouts, hoping to start a conversation and steer users to our app
Nothing stuck. Few people cared enough about tracking their progress to even try our app once. However, around the same time we learned that two strangers had been using Atlas every day for a month. Confused but very keen to learn more, we bombarded them both with emails asking for a Zoom call. Thankfully, one of them agreed!
This keen user, who we’ll call Terry, had used Atlas for nearly an hour every day for over a month. We asked if it was due to our snazzy code tracking her workout in real-time, or our dashboard of detailed statistics showing her performance after every session. But no! It turned out that Terry didn’t know what those numbers popping up on her screen - that we’d spent six months coding - even meant. Instead, she was using Atlas to see whether she matched the trainer in the video. It turned her laptop into an “AI mirror” - which was a small upgrade on the regular mirror she’d been placing behind her laptop before.
This was only a small part of the app and it was by far the easiest to build, but it was the only part providing any value to users. We took the painful decision to get rid of the rest of the app, rebranding Atlas as the “AI Mirror for YouTube workouts”. This felt like a big risk, given the months spent coding those flashy numbers, but we crossed our fingers that it would lead to more users getting something they really wanted…
In September, we started getting hundreds of users to do workouts with our rebranded app. This felt like huge progress - and a huge relief - given we couldn’t get a single person to try it just a few months earlier. We were back on track! Most users didn’t return after their first workout, but we knew that was normal for an early-stage app. After all, there were plenty of improvements still to make.
Some problems aren’t worth solving (November - December 2022)
After two months of improving the “AI Mirror”, most people were still leaving after one workout. It was tough to get leavers to speak with us, even after offering to pay them for feedback. In order to better prioritise what to work on, we started reading about product management.
A key principle we read about was the role of problem severity. Maybe our users weren’t leaving because of friction in the app, like loading times. Instead, maybe they just didn’t care that much about checking their form in the first place…
We needed to find out ASAP. In December, we introduced a paywall of $10 per month and anxiously waited to see how many users stayed with us.
None did.
They all fled.
Even Terry, who’d now been using Atlas every day for six months, wouldn’t pay. We realised that we’d made a huge mistake: some problems aren’t worth solving.
Finding a problem worth solving (January 2023 - May 2023)
We returned after the Christmas holiday, frustrated that we’d spent a year solving a problem that nobody cared about. Going forward, we resolved to find a problem worth solving. We decided to charge at least $10 for any new apps upfront: if the problem is severe enough, people should pay for the solution.
We stopped coding, again, and spent the next few months running marketing campaigns to test what people would pay for. Even after working with experienced marketing consultants, we struggled to find a compelling story for a home workout app - it was costing $200 in ads to find one person who’d pay $10 upfront. It didn’t take a business genius to realise this wasn’t going to work.
We read marketing books to see if we improve our results. While the books were quite different, they all started with broadly the same message:
Good marketing doesn’t start with ads. It starts with a story you want to spread. This story should already fit in with the customer’s worldview, so they’ll spread your story too.
Shit.
We didn’t have that story. Our poor marketing results were just a symptom of a much bigger problem - we didn’t care enough about our customers. We’d never actually helped anyone with their home workouts, we’d only ever helped people lifting weights in gyms. How the hell did two powerlifters end up selling an app for counting starjumps?!
Exasperated and embarrassed that we’d spent 18 months trying to sell a home workout app we didn’t believe in, we knew that there was only one place we could go from here: back to the drawing board for a last roll of the dice.
You can find out how this went in the fourth, and final, part.
What we learned
We mistakenly believed (again) that re-packaging our existing product for a new user was a genius move. Like our first pivot, we didn’t see that a new user would have a new set of needs. Instead, we lazily assumed that people doing home workouts would want a Strava-like app because runners used it. We’re planning to write a longer post about making better pivot decisions.
Startups must care deeply about their users. This is hard without a genuine mission that’s personally meaningful to the founders. We would start any new company with a written mission statement, and then make decisions with reference to the mission. We plan to write more about this.
Your early approach should not be easy to scale. We were acquiring users with ads, and then servicing them with a fully automated product. Users never had to speak with us, so it took us a long time to get feedback on our app. It would have been better if our early approach involved more manual work, as we’d have spent more time with our users and learned more about them. We plan to write more on this.
We hired full-time team members too early. We focused on work that was aligned with the team’s skillset, coding, rather than the work we really needed to do, acquiring users. We had to ban coding to refocus. We plan to write more about this.
Special thanks to Dr. Tishtrya Mehta for her detailed review and comments, which greatly improved this post.