Susa Ventures’ Misha Gordon-Rowe: Vertical AI Applications
Thesis BriefsMisha Gordon-RoweSusa VenturesMaddoxAI/MLSaaSVeevaViz.aiParsylAlife Health

Terra Nova Insights Team,

Susa Ventures’ Misha Gordon-Rowe: Vertical AI Applications



Industry or function specific AI applications are emerging with 10x better tools, predictive capabilities, and personalized user experiences compared to traditional software. These companies will turn out to be more durable businesses than traditional SaaS, if they are able to get high-quality and differentiated datasets for their training models. Misha Gordon-Rowe, principal at Susa Ventures, assesses how vertical AI companies can find long-term defensible strategies. The key ingredients will be seamless integration into existing industry workflows, and a focus on user-facing design.

Why are vertical AI applications such an important category moving forward?

  • Vertical AI companies provide more business value than traditional B2B workflow SaaS. These new AI-driven companies can augment human decision-making, provide personalized experiences for every user, and embed predictive capabilities into workflows.
  • “In the future, moats will come from differentiated datasets that improve through product use,” says Gordon-Rowe. Underlying models are more and more accessible, and software is easier to build than before.

How can companies kick-start collecting differentiated datasets?

  • Build tools for non-technical employees to label their own data. "One example is Maddox AI," says Gordon Rowe. "Maddox sells visual inspection systems to factories so they can automatically identify product defects and quality issues. When a new customer is onboarded, instead of sending images to third-party labeling operations, Maddox has created an intuitive labeling tool tailored for supervisors on the factory floor. By leveraging the domain expertise and experience of factory-floor workers, Maddox is able to get data labeled more accurately and cheaply, whereas an outsourced labeler might not have the knowledge to accurately identify or classify defects. This is an example of user-friendly UX being used to kickstart a proprietary data flywheel."
  • Acquire data for the core AI product, "Product Number 2," by delivering value with "Product Number 1," according to Gordon-Rowe.
  • Parsyl is an insurance company for physical goods that require cold-chain logistics. Parsyl initially goes to market by offering proprietary sensors and cloud-based insights for shippers to track their shipments across the cold chain." At the same, this sensor data allows Parsyl to help shippers mitigate risk and reduce losses," says Gordon-Rowe. "And ultimately offer insurance more cheaply than other insurance companies who don’t have this data advantage."
  • Another example, Alife Health provides a set of algorithms across the IVF cycle that augment fertility doctors’ decision-making and increase the chance of successful pregnancies via IVF. "Alife started out by building trust and partnerships with the nation’s leading fertility doctors and clinics," says Gordon-Rowe. "They built one of the most robust and diverse IVF datasets from the top clinics globally, allowing them to deliver more personalized care with a higher chance of success to each patient."

What are some of the potential roadblocks?

  • Companies with tunnel vision on model optimization can fall behind competitors that prioritize the end-user experience. “The best vertical AI teams spike in product, design, and user empathy, not AI research,” says Gordon-Rowe.


saidbyblock~ via Zoom.
Misha Gordon-Rowe

Major paradigm shifts like the move to cloud have happened substantially in many industries over the past ten years. Now, vertical AI applications that serve large industries will outcompete traditional vertical software with 10x better tools, predictive capabilities and personalized user experiences. And these vertical AI companies will ultimately prove to be more durable businesses if they nail their data strategy.

As AI models become more powerful and commodify, high quality data is becoming the biggest constraint on building better vertical applications. Unique or hard to get data, and a data flywheel, are the keys to building a defensible business here. For the dataset to be truly differentiated in the long run, it generally has to be generated through the usage of the product — or at least be meaningfully difficult for competitors to get. Data generated in other forms is fine, but the mechanism generally can be copied.


Q: What do other market participants or observers misunderstand about these categories?

A: First, for the AI to be useful, it needs to be embedded seamlessly in the existing day-to-day workflow of an industry.  For example, is a healthcare company that uses AI and computer vision to detect strokes on CT scans, and then coordinates care across physicians and hospitals to reduce time to treatment, a critical factor in treating stroke. While Viz’s AI algorithm for stroke diagnostics was a breakthrough (it was the first FDA-cleared AI triage software), the real big breakthrough was building a platform that allowed the AI to augment a doctor’s decision-making and streamline care coordination all within the context of a hospital’s existing workflows. Astutely, Viz neither went to market with a pure diagnostic algorithm nor did they try to rearrange a hospital's stroke handling workflow entirely. Products that require a lot of extra work from users, or force large behavior changes, are much less likely to succeed in most verticals

Secondly, market participants should remember that value created by new AI tools – however big – doesn’t guarantee value capture. For example, over the next year there will be a proliferation of industry vertical specific tools powered by Large Language Models (LLMs) that will be an order of magnitude better than the status quo, but many will fail to capture long run value as they’re built on open source or open API models. Value will only be captured by those companies that have a compounding data advantage - generally through product usage – and are customer-obsessed in their UX design. The better and more intuitive your product is, and the more it aligns with customer needs and expectations, the more data you'll be able to collect that makes your product even better.

Finally, the most critical skills on vertical AI application teams are generally product and design, as well as back-end infrastructure – not AI researchers.

Q: What can you say about the time horizon for broad business/corporate or consumer adoption of the technologies or behaviors that underpin your Thesis?

A: Large enterprises won’t always be able to rely on new vertical AI applications due to security and data protection risks, cost and other reasons. As a result, new developer platforms will emerge for enterprises in certain scenarios to build and deploy their own self-hosted, enterprise-grade ML and LLM apps.


Ultimately, some vertical AI applications will vertically integrate. Over the past 10 years, we saw that some companies that started by building software for industries like real estate brokerage, wealth management, and insurance brokerage ultimately leveraged their software platform and entered those markets themselves to compete with incumbents. We may see this happen in instances here, with AI companies vertically integrating and competing head to head with the industry they were set up to serve. In the end, they may be able to provide more personalized services and rethink these workflows with AI-augmented humans at the center. For example you could imagine an AI software company for insurance claims management building their own Third Party Administrator (TPA) that is more accurate and automated than the rest of the fractured TPA market. That said, vertical AI application companies may be able to build an advantage by guaranteeing to customers that they will never compete.

In addition to vertical AI applications for specific end markets, industry-focused foundation models will emerge in industries like healthcare, insurance and manufacturing. These companies can be the AI infrastructure layer for the vertical. Other companies can then train domain-specific models for vertical AI applications using these foundation models (this seems to be the position of Sam Altman at OpenAI). A SaaS analogy here is how Veeva used Salesforce as a starting point to build the CRM for life sciences.


Alife Health, Maddox, Parsyl , Veeva,

The 2022 EVC List honors the top 50 rising starts in venture capital. Terra Nova’s Thesis Brief series showcases each investor’s insights and category expertise.

Related Posts

Costanoa's John Cowgill: Overlooked Bets On AI in the Enterprise

Costanoa's John Cowgill: Overlooked Bets On AI in the Enterprise

John Cowgill, Partner at Costanoa Ventures, points to some of the more overlooked and contrarian business models within this closely watched ecosystem. As the space matures, competition between emerging companies and quick-to-adopt incumbents will determine the winners...