Building an AI app with 100k+ users

Back in January, we were lost in the idea maze. We had spent the last 6 weeks in pure customer discovery mode and it felt like we were at a deadend.

We decided to stop midcurving things. Instead of validating our hypotheses with potential customers, we decided to build a GPT-wrapper over a weekend: Sage. We had a clear distribution strategy (we were one of the first 100 miniapps to be built on WorldApp) and a clear hypothesis we wanted to test: personalisation makes AI apps more sticky.

Within 2 months, Sage had 175k users and 15% DAU/MAU.

We learned a lot more about the problems and opportunities of building AI apps by doing it ourselves than we would have speaking to hundreds of AI builders.

These lessons led us to build Mio, a context tool for AI apps.

If you're building in this space, sign up to our waitlist.

Distribution > product

In the short term, distribution is all you need. How you get users should come as a precursor to what you build.

Set up analytics asap

Sage received millions of messages in dozens of languages. We found no simple and affordable tool to analyse and cluster conversations. We ended up implementing Kura. We should have setup analytics earlier.

Watch your users interact with your product

Most of our users were in Latam (because of World). A month after launch, we spent 10 days there. Seeing taxi drivers use Sage taught us more than a dashboard ever could.

You can keep costs under control

Remember when your Uber rides across town cost 8€? We're at this stage in the AI buildout. We used groq for latency, built a load-balancer to loop through models to stay on the free tier and used 4o mini as fallback. There was a month where we didn't pay for inference whilst having tens of thousands of users.

Scientific approach to product building

Start with a simple app, a hypothesis and a metric to validate it. We started by validating the distribution, then we tested personalisation. Doing so at the same time wouldn't have given us the base to compare against.

Build for your mom, not for yourself

There is a huge delta between the X timeline and real world adoption. On X, the onus is on the flashiest launch video, the newest MCP server, the latest agent framework. The reality is that 3 years after the launch of chatGPT, AI hasn't significantly affected GDP. Rather than focus on beating the latest benchmarks, AI builders should focus on solving user problems. Agents can do more than book a trip or push code. The companies we find most exciting are the ones augmenting humans, not automating them.

Global by default

LLMs generate outputs in whatever language they are asked. You need minimal adaptation to reach audiences on the other side of the world.

Quality of users > quantity of users

Building on World had a lot of advantages: early mover advantage, built in identity and payments, anti sybil, clear path to scale etc. For a while, we were the largest non speculative third-party miniapp on the WorldApp. It was a great way to validate our hypothesis and learn about the problem of AI app builders but in the end, the World user isn't the user we want to serve (it currently isn't the user our customers care about).

Retention is brutal

You've built a product and have found a legit distribution channel, now you need your users to stick. AI apps face a conundrum:

• models keep getting better, why would users stick to your app rather than use GPT.

• a flood of new apps launch every week, how do you stay top of mind.

• with the pace at which AI moves, it's hard for users to form new habits.

Personalization is the moat

The key to retention is personalisation. It's the case across the internet: in social media, entertainment, productivity etc. No one wants to spend minutes giving an LLM the perfect context, user want it to just know. We tried different approaches to personalisation, the one that was most successful was deriving preferences from existing datasources. With these learnings in mind, we built Mio.

With the tools at your disposal, anyone can build an app in a matter of days. You have no excuse.

PS: We're recruiting. If you're looking for a place to do your life's work, DM me.