Artanis #9: Record revenue! And some growing pains...
We're celebrating our £50k monthly revenue milestone!
Artanis: Helping companies build AI that they can control. Previous updates at https://artanis.ai/
🙋♂️ Ways you can help: people building AI products 🙋♂️
We're looking to speak with people at companies meeting the following criteria:
1. An LLM is a core component in their main product
2. Their product is already in the market and has some traction
3. They have a modelling team of at least 5 people and one of their team has a background in AI/ML rather than software engineering
Please get in touch if you know anyone meeting these criteria.
📉 Progress in December - learning about the labelling market 📈
In December, we started discovery conversations among companies selling data labelling services. We met our initial target of confirming similar pain points across at least three companies in the segment. We booked eight calls and have written about the conclusions in more detail below. We’ve also proceeded to demo with two prospects.
Pain points were mostly in line with our priors. However, 3 out of 4 companies said they wouldn't buy third-party labelling tooling because it's too close to their core IP. We've therefore decided to alter our focus to speaking with companies building AI products and who are labelling data in-house, rather than selling labelled data.
Separately, our consulting metrics were:
Primary metric: 4 customers (+1)
Secondary metric - revenue: £50.0k (+£14.5k)
Secondary metric - team size: 3 (this is a proxy for cost)
We’re very proud to hit £50k monthly revenue within 7 months of starting Artanis 🎉 While not our primary metric, it’s proven we can go-to-market in the B2B AI space. (And it pays the bills, of course)
🔬 What we learned about the labelling market 🔬
We learned some super interesting stuff about the cutting-edge players in data labelling:
Most of their business is now labelling difficult tasks (e.g. answering complex financial questions) rather than simple ones (e.g. is this image a cat or a dog). Volumes have shrunk, but they can pay up to $1,000 per labelled query!
Mislabelling is very costly, as it leads to models being trained on bad data, and they're spending heavily on QA to reduce labelling mistakes.
They draft very detailed labelling guidelines to reduce ambiguity and errors. However, it's hard for their labellers to remember all the guidelines, so they still make mistakes. The tradeoff between the length of guidelines and ambiguity is an open problem.
Sourcing labellers with the right domain expertise for these difficult tasks is very expensive. One company told us they needed to source 200 lawyers within a fortnight!
🧐 Challenges - competing business ideas 🧐
We have multiple compelling go-to-market strategies and we’re finding it hard to choose the best one. We’re keen to hear from others who’ve been in founding teams with competing ideas - how do you think co-founders should resolve strategy disputes?
🏹 Goal for January - validate at least one customer segment 🏹
This is pretty similar to last month's goal. We want to confirm pain points across at least 3 companies building ML models that are labelling data in-house, and that they're willing to buy third-party labelling tooling. If we can do this, we'll progress from discovery to sales.
📣 Shout-outs 📣
Thanks for December go to:
Bertie V - our intro MVP for 2024!
Sivesh S - spreading the gospel about us
Anthony F - for connecting us to Josh
Nandu A - for sending curious investors our way
Karun S - for the chat about labelling data at Arsenal FC!
Cheers,
Yousef, Jerome & Sam