With the rise of open-source AI models, the commoditization of this groundbreaking technology is upon us. It’s easy to fall into the trap of aiming a newly-released model at a desirable tech demographic and hoping it catches on.
Creating a moat when so many models are easily accessible creates a dilemma for early-stage AI startups, but leveraging deep relationships with customers in your domain is a simple, yet effective tactic.
The real moat is a combination of AI models trained on proprietary data, as well as a deep understanding of how an expert goes about their daily tasks to solve nuanced workflow problems.
In highly-regulated industries where outcomes have real-world implications, data storage must pass a high bar of compliance checks. Typically, customers prefer companies with prior track records over startups, which promotes an industry of fragmented datasets where no single player has access to all the data. Today, we have a multi-modal reality in which players of all sizes hold datasets behind highly compliant walled-garden servers.
This creates an opportunity for startups with existing relationships to approach potential customers who would typically outsource their technology to launch a test pilot with their software to solve specific customer problems. These relationships could arise through co-founders, investors, advisors, or even prior professional networks.
The real moat is a combination of AI models trained on proprietary data, as well as a deep understanding of how an expert goes about their daily tasks to solve nuanced workflow problems.
Showing customers tangential credentials is an effective way to build trust: positive indicators include team members from a university known for AI experts, a strong demo where the prototype enables prospective customers to visualize outcomes, or a clear business case analysis of how your solution will help them save or make money.
One mistake founders commonly make at this stage is to assume that building models of client data is sufficient for product-market-fit and differentiation. In reality, finding PMF is much more complex: just throwing AI at a problem creates issues regarding accuracy and customer acceptance.
Clearing the high bar of augmenting experienced experts in highly-regulated industries who have an intricate knowledge of day-to-day changes typically turns out to be a tall order. Even AI models that are trained well on data can lack the accuracy and nuance of expert domain knowledge, or even more importantly, any connection to reality.
A risk-detection system trained on a decade of data may have no idea about industry expert conversations or recent news that could render a formerly-considered “risky” widget completely harmless. Another example could be a coding assistant suggesting code completion of a prior version of a front-end framework which has separately benefitted from a succession of high-frequency breaking feature releases.
In these types of situations, it’s better for startups to rely on the pattern of launching and iterating, even with pilots.
There are three key tactics in managing pilots:
