At a Rise of AI Virtual Chat in 2020, Erik Pfannmöller, CEO and founder of Solvemate, discusses how to start an AI-first company with host Fabian Westerheide. Pfannmöller describes moving from a background as a competitive athlete and founder of an e-commerce company to spending a full year in 2014-15 self-teaching machine learning, from reading Peter Norvig's textbook to attending meetups, before an idea emerged from helping his father with an email problem. He explains Solvemate's path from an Excel-based concept to a viral Facebook Messenger animal-guessing game that proved the underlying algorithm worked, calling this step "algorithmic fit," which he argues should precede product-market fit. He describes it taking one and a half years to launch a first product and two years to reach a first paying customer, and advises raising three times the funding founders think they need. He argues the label "AI-first" is largely irrelevant compared with whether a company solves a real business problem, and describes Solvemate's own approach of combining rule-based dynamic decision trees with limited natural language processing.
Key takeaways
- Pfannmöller spent about a year self-teaching machine learning, including reading Peter Norvig's textbook, before founding Solvemate.
- He calls proving an algorithm works before building a product "algorithmic fit," distinct from product-market fit.
- It took Solvemate one and a half years to launch a first product and two years to land a first paying customer.
- He advises founders to raise three times the funding they initially think an AI product will require.
- He estimates AI-driven customer service automation had penetrated only about 0.5 percent of its potential market.
Detailed notes
From competitive sports and e-commerce to AI
Pfannmöller describes a background as a twelve-time German and world champion athlete before studying business and building an e-commerce shopping-club company that reached over 20 million euros in funding, 70 employees and 20 million in revenue before he sold it. He describes his path through what he calls "quantum leaps," from athlete to entrepreneur, driven by an underlying interest in mathematics, statistics and IT. After selling the e-commerce company, he decided any new venture had to combine IT, mathematics and business, which led him to consider either Bitcoin or machine learning around 2014, ultimately choosing machine learning.
A year of self-taught research before founding Solvemate
Pfannmöller says he spent about a year deep-diving into machine learning before starting a company, reading a thousand-page math textbook by Peter Norvig, attending every relevant meetup, and meeting people for lunch to understand which data feeds which algorithms, from NLP to self-driving cars. He argues this research phase was necessary to build enough confidence to found an AI company, comparing it to needing to understand animals before running a zoo. The idea for Solvemate emerged when his father called saying he could not send emails; while diagnosing the problem, Pfannmöller recognized he was applying decision-tree logic he had just read about in a scientific paper.
Algorithmic fit before product-market fit
Pfannmöller describes a concept he calls "algorithmic fit," arguing that before a SaaS company can pursue product-market fit, it needs proof the underlying algorithm works, which Solvemate first modeled in Excel. In 2016, when Facebook opened its Messenger platform to chatbots, Solvemate built a viral guessing-animals game to test their intent-recognition technology, reaching over 85 percent accuracy and letting users train new animals into the system. He calls this the moment they had proven their algorithm before building a product on top of it.
What it actually takes to start an AI company
Pfannmöller says he immediately brought on a technical co-founder as CTO rather than attempting to be both CEO and CTO, and argues founders need at least a fair understanding of the underlying technology even if not coding it themselves. He says it took one and a half years from starting research to launching a first product, and two years total to land a first paying customer, warning there was no existing framework like TensorFlow to shortcut the process for their specific use case. His central funding advice is to raise three times the amount founders initially estimate they will need, since AI development always takes longer and costs more than expected.
Is "AI-first" a meaningful label?
Pfannmöller argues that whether a company is called "AI-first" is largely irrelevant, comparing it to asking in 2003 whether a company "uses the internet"; what matters is whether removing the underlying mathematical models breaks the product. He describes Solvemate's approach as combining rule-based dynamic decision trees with natural language processing added back only in small parts, a pragmatic choice he says let them launch faster than competitors betting fully on NLP, which he calls "still tricky." On competition, he estimates AI-driven customer service automation has penetrated only about 0.5 percent of the market, so the biggest competitive threat is companies not acting at all, not other AI vendors.
Q&A highlights
Asked whether a DeepMind-style AI research company could be built in Germany, Pfannmöller says research companies are effectively nonprofit cost centers funded by grants or state and corporate sponsorship rather than standalone businesses. Asked which industries benefit from the Corona crisis, he cites e-commerce, SaaS and automation, and predicts virtual conferences will keep thriving after the crisis. Asked about his biggest challenges, he cites generating enough personal energy as a founder and wishing Solvemate had moved faster with more strategic foresight. He advises technically trained founders to study business, fundraising, HR and company strategy, and describes Solvemate's "marble rule," where late arrivals drop a red marble and on-time attendees a green one at team meetings, as an example of using simple processes to change team behavior.
Quotes
You need to have more money than you think it takes.
Erik Pfannmöller
It's not about the AI, it's about… a technology that enables companies to thrive.
Erik Pfannmöller
The biggest fear is not the biggest competitor, it's people are not acting.
Erik Pfannmöller
Questions answered in this talk
What does Erik Pfannmöller mean by "algorithmic fit"?
According to Erik Pfannmöller at a Rise of AI Virtual Chat in 2020, algorithmic fit means proving an underlying algorithm actually works, for example through a simple prototype or viral test, before pursuing product-market fit for an AI product.
How much funding does an AI startup need compared to initial estimates?
Erik Pfannmöller advises raising about three times the amount founders initially believe they will need, since AI development consistently takes longer and costs more than expected.
Can a DeepMind-style AI research company be built in Germany?
Erik Pfannmöller argues such research organizations are effectively nonprofit cost centers funded by state grants, universities or corporate sponsorship rather than standalone profitable businesses, making the chances of building one as a company very low.