Insights from Founders on Building Reliable AI Products
At the recent Deploy 2026 conference in San Francisco, a panel of founders discussed the challenges and strategies involved in developing AI products that users can trust. The conversation highlighted that while access to advanced AI models is easier than ever through APIs, the real differentiator lies in how these models are integrated and maintained within applications. The panel featured Angela Hoover of Andi AI, Alex Mashrabov of Higgsfield AI, Hovsep Seraydarian of LawVo, and Peter Elias of Probably.
The Importance of Team Composition
One key takeaway from the discussion was the significance of having a diverse team that includes individuals with creative backgrounds. Alex Mashrabov shared how Higgsfield AI initially struggled to gain traction until they integrated non-technical team members who understood creative processes. By placing creative professionals alongside engineers, the company was able to develop a product that resonated more with its target audience.
“We started to see success when we got non-technical people on the team, like ex-creative directors, who now work daily with engineers to wrap this powerful technology and make it accessible for creatives,” Mashrabov explained.
The Need for Human Oversight
The founders agreed that human oversight remains crucial, especially in high-stakes environments like law. Hovsep Seraydarian noted that while LawVo’s AI agents could provide legal guidance, they required constant validation from human lawyers. Contrary to expectations that automation would reduce human involvement, Seraydarian indicated that as their agents became more sophisticated, the need for human judgment actually increased.
“We need human lawyers to verify the data and test these agents every day,” he stated. This sentiment was echoed by Angela Hoover, who emphasized the necessity of monitoring AI agents to ensure they operate with high-quality data.
Establishing Measurement Infrastructure
The discussion also touched on the importance of creating robust measurement infrastructure to evaluate AI performance. Peter Elias highlighted that establishing analytics systems early on is essential for understanding whether an agent is functioning effectively. His company, Probably, developed an internal system capable of monitoring agent behavior and suggesting improvements based on performance metrics.
“As long as we record everything the AI is doing, this particular AI can now actually aid us in improving its own performance,” Elias explained. This proactive approach allows teams to refine their products continuously based on real-world usage data.
Navigating Model Selection
With a plethora of available models flooding the market—from frontier releases by companies like OpenAI to open-source alternatives—selecting the right model has become increasingly complex. Peter Elias outlined four critical factors to consider: cost, latency (the time it takes for a model to respond), intelligence (the model’s capability), and capacity (how much data it can handle). He advised aiming for the simplest model that meets product requirements without sacrificing performance.
“You want to get to the dumbest model you can get to before you actually go below the product performance that you need,” he noted. This approach helps balance speed and efficiency against user expectations for responsiveness.
Understanding ‘Agentic’ vs. Autonomous
A significant point raised during the panel was clarifying what “agentic” means in relation to AI systems. None of the founders considered their agents as fully autonomous; instead, they viewed them as tools requiring continuous human guidance. Hovsep Seraydarian pointed out regulatory constraints within legal practices that prevent full autonomy from being feasible or acceptable.
“Regulations won’t allow you to go fully autonomous. You literally get shut down if you do that in this space,” he said. Meanwhile, Peter Elias emphasized that true agency involves spontaneous action without external input—a characteristic LLMs (Large Language Models) do not possess.
What This Means
The insights shared by these founders underscore several critical aspects for anyone looking to develop reliable AI products: building a multidisciplinary team enhances creativity; maintaining human oversight is crucial in high-stakes applications; establishing measurement infrastructure is essential for continuous improvement; navigating model selection requires careful balancing; and understanding what constitutes agency in AI systems is vital for responsible development.
The landscape of AI continues to evolve rapidly, making it imperative for companies not only to leverage cutting-edge technology but also to focus on practical implementation strategies that prioritize user trust and reliability.
For more information, read the original report here.
