Helping companies build AI that actually works.

🙋 Ways you can help - intros to AI teams 🙋
We’d like to meet people with the following profile:

1/ Startup with an LLM-based product in the market
2/ Technical leader e.g. CTO, Head of AI
3/ The startup is post-revenue, but has fewer than 100 employees

Please get in touch if you know anyone who fits the bill!

📉 Progress in November - ruling out FMCG operations 📈
Our main goal for November was to commit to a narrower product, because initial discovery calls suggested that a horizontal solution for process automation was too broad. We weren’t able to find a compelling opportunity for a narrower product after considering the following:

1/ Order processing - there was some market pull for this, but it wasn’t a slam-dunk. Only some FMCG ops teams do this manually, and only for a fraction of their orders. There are also other AI order processing solutions. Overall, we felt there was a viable business opportunity, but it would’ve been an uphill struggle and not a strong fit for our team.

2/ Demand planning - this came up repeatedly on customer calls as a major pain point. However, it’s a longstanding problem that mostly stems from poor data availability. We therefore felt like there was no compelling “why now” for how we could solve it as a team.

A bright spot was great results from cold calling, as we improved our lead selection and how we asked questions on the call. This got a lot of interest in the last update, so we decided to share the metrics again:

- 264 dials (across ~100 prospects)
- converted to 49 answered
- converted to 15 discovery calls

Ruling out a strategy after speaking to customers isn’t a bad thing - it’s arguably the name of the startup game. But it does lead to a difficult question: what’s next?

🤔 How well do you need to understand your customers? 🤔
We didn’t feel in a strong position when speaking to FMCG operational leaders. While we felt credible when talking about AI, we didn’t have expertise when broader operational topics came up. This is tolerable if you’re willing to put in the time to build expertise. We’ve written before on how well you need to understand your customer, and realised we weren’t following our own advice!

We were also concerned that we were becoming “AI in search of a problem”. We wanted to deal with customers for whom AI is the problem, rather than (potentially) the solution. AI is most likely to be the problem for AI teams.

We decided our next move needs to take us back to a customer segment we felt stronger with - teams building AI products. Our expertise is AI, and that’s where we want to keep building.

🐌 The hard thing about going back to square one 🐌
Pivoting customer segment is a major decision. It takes us back to being not sure what we need to build, or even exactly what problems we’re looking to solve. The team we’d hired in March, initially to scale a Palantir-type approach, was no longer the right fit for the new direction. As a result, our team has shrunk from five to three.

If you’re looking for an AI engineer, then get in touch or reach out to Andrew directly. He's excelled at our customer-facing work, getting rapidly up to speed with existing customers and building solutions for new customers. He brings a rare combination of technical depth and stakeholder-engagement/communication skills.

🏹 Goal for December - validate accuracy monitoring 🏹
Monitoring accuracy in production is a problem for AI teams. It’s why observability and tracing tools get a lot of attention: AI teams hope that it will allow them to monitor accuracy after deployment. However, they’re missing a big piece. While they can monitor inputs and outputs, they can’t verify whether the output was correct. This would require knowledge of precisely what the model should have done, i.e. the “ground truth”.

We have a hunch that there’s a gap in the market for a monitoring platform that enables robust accuracy measurement. Our main goal is to validate three key hypotheses by speaking with AI teams:

1/ There’s a strong link between their AI accuracy and revenue
2/ The main bottleneck to better accuracy is a lack of robust measurement (or “evals”)
3/ Existing solutions can measure inputs and outputs, but not accuracy

🙏 Shout-outs 🙏
Special thanks for November go to:

Srecko D - for being a great “sparring partner” (again)
Conrad L - for always looking to help
Lorenzo S - for good advice, over Vietnamese
Henry M - for being the most helpful £1k angel in the UK
Guillaume B - for not one but two informative discovery calls
Paul F - for the intro to Guillaume!
Mo N - for repeating our advice back to us
Oliver W - for the detailed look at where accuracy matters
Amil K - for the shout on exploring compliance use cases
Zahid M - for LLM-as-judge in prod insights
Stefano G - for insights on how accuracy affects buyer trust
Maurice B - for the intro to JJ, and plugging our Substack!

En Taro Tassadar,
Artanis Team