At Rise of AI 2019 in Berlin, a panel with Ben Blume of Atomico, Jasper Masemann of HV Holtzbrinck Ventures and Rasmus Rothe, CTO at Merantix, takes audience questions on what actually works in AI-driven business models. Rasmus Rothe describes Merantix as a venture studio that incubates AI companies for one to three years with up to a few million euros in funding, running joint projects with industry partners such as hospitals and car manufacturers. The panel discusses how machine learning maturity varies by field, with areas like speech recognition already supported by mature libraries while others, such as large-scale simulation, remain research-led and carry both technical and commercialization risk. Audience questions cover which industries AI will disrupt most, liability and explainability for black-box models, why the panelists keep investing from Europe despite the pull of Silicon Valley and Asia, how startups can build labeled data sets through customer-facing products before adding machine learning features, and what business-outcome measures, rather than raw model metrics, investors actually use to judge an AI company's performance.
Key takeaways
- Rasmus Rothe describes Merantix as a venture studio incubating AI companies for one to three years with up to a few million euros in funding.
- The panel says companies in less mature ML fields face both technical risk and commercialization risk at the same time.
- Panelists name healthcare, autonomous driving, manufacturing, back-office automation and cybersecurity as the industries most likely to be disrupted by AI.
- The panel raises liability and certification as open problems for black-box neural networks used in high-stakes decisions.
- Investors say they judge AI companies on business-outcome metrics like faster onboarding rather than on raw scores like F1 or precision and recall.
Detailed notes
Two investing models: applied venture capital and the Merantix venture studio
Ben Blume describes Atomico as a London-based generalist venture fund that invests both in applied AI and analytics companies solving specific commercial problems and in more capital-intensive frontier technology bets, citing an investment in a company building a new semiconductor for AI workloads. Rasmus Rothe describes Merantix as a venture studio that incubates AI companies for one to three years, funding them with up to a few million euros and helping with talent and industry partnerships through its MX Labs unit, which runs joint projects with hospital partners in healthcare and car manufacturers in automotive before spinning out companies, sometimes together with a corporate partner. He notes the studio's own core team stays deliberately small, at around four people, since portfolio companies operate in different industries and share knowledge and code without competing with each other.
From research to product: how mature is machine learning really
The panel argues that maturity varies sharply by field: some areas, such as speech recognition, already have mature libraries suitable for applied products, while others, such as massive-scale simulation, remain research-led. Companies working in the less mature areas face two risks at once, technical risk in getting the underlying approach to work and commercialization risk in turning it into a product, which the panel says makes such companies more capital-intensive to build but potentially more rewarding. Medical imaging is offered as an example: training a classification algorithm on a public data set can be done in a day, but building a medical-grade product requires handling large volumes of medical data, deciding carefully what to annotate, and quantifying prediction uncertainty, work the panel says takes far longer and is often underestimated because academic research deals with smaller, more standardized data sets.
Which industries face disruption, and who is liable
Asked which business areas AI will disrupt most, the panel names healthcare as an obvious candidate given the volume of available data, along with mobility and autonomous driving, manufacturing and robotics, back-office automation, and cybersecurity. On autonomous driving and other regulated applications, the panel raises the question of liability when a large neural network, which it describes as effectively a black box, makes a mistake or shows bias, and how such systems can be certified. On personal data more broadly, the panel distinguishes between low-stakes uses such as targeted advertising and higher-stakes ones such as credit or healthcare decisions based on genetics, arguing that public willingness to trust AI-driven decisions in these more sensitive areas will take longer to develop than the technology itself.
Europe's edge in talent and data
Responding to a question about competing with the US and Asia, the panel points to strong, comparatively affordable technical talent in Europe, including migration of talent from Eastern and Western Europe into German startups, and well-regarded university education. It also argues that European data can differ meaningfully from US data, for example in accounting or healthcare systems, and repeats an unverified account that a large US company's algorithm, trained on data from Boston, did not work as well once deployed in Germany, framing this as an argument for building AI products on genuinely local data and industry experience.
Building labeled data and defining AI-specific KPIs
On data strategy, the panel recommends standardizing around use cases where the underlying process is already standardized, such as mammography imaging, and building self-service products rather than custom, project-based integrations. It cites Sitecode, which automated small-business accounting manually at first while generating and annotating data, and Odin Technologies and Streetbees, which shipped a data-visualization platform and a photo-based survey product respectively as first versions, using resulting user-generated labels to train machine learning features added later. On what counts as an AI company, the panel says the distinction matters less as machine learning becomes one component of a broader tech stack, and that they invest based on the business problem being solved rather than on AI branding, advising founders not to overstate AI use when fundraising. Asked how to evaluate an AI company's performance, the panel says they rely on business-outcome proxies, such as faster onboarding or more accurate matches, rather than raw metrics like F1 score, precision or recall, verified through their portfolio CEOs' own diligence with data scientists.
Quotes
It's a venture studio focused on incubating AI companies for one to three years. We're very flexible on how long we incubate the companies.
Rasmus Rothe
Questions answered in this talk
What areas of machine learning are mature enough for real products today?
According to the panel at Rise of AI 2019, speech recognition already has mature libraries suitable for applied products, while areas like large-scale simulation remain more research-led and carry both technical and commercialization risk for companies building on them.
Which industries do venture investors expect AI to disrupt most?
According to the panel at Rise of AI 2019, panelists named healthcare, autonomous driving and mobility, manufacturing and robotics, back-office automation, and cybersecurity as the industries where AI's impact is likely to be most disruptive.
How do venture capitalists evaluate the performance of an AI company?
According to the panel at Rise of AI 2019, investors say they rely on business-outcome measures such as faster onboarding or more accurate results rather than raw machine learning metrics like F1 score, precision or recall, checked through their portfolio companies' own data scientists.