Google Product Lead on Building AI Products That Actually Work

How do you build an AI product that’s powerful, reliable, and grounded in real user needs? In this podcast hosted by Qventus Product Director Mark Bailes, Google Product Lead Alankar Agnihotri discusses building AI products that actually work. He shares the hidden complexities behind model behavior, the shift from deterministic to probabilistic design, and what product managers must do to deliver AI experiences that scale. You’ll hear firsthand insights from someone shaping Gemini in Android Automotive and building the future of in-car intelligence.

Subscribe to the Product Talk podcast on Spotify and Apple Podcasts and catch every conversation with leading product executives.

---

---

LISTEN ON ITUNES

---

LISTEN ON SPOTIFY

---

---

Show Notes

1. People often overestimate what AI can do and underestimate what it takes to build robust AI products.
2. Large language models aren’t magic; they require data quality, tuning, and real infrastructure to perform well.
3. True AI product success still begins with user needs, not with the technology itself.
4. Many teams unintentionally start with the model and work backwards—leading to misaligned solutions.
5. Industries like autonomous driving and medical diagnosis show how long it can take for AI promises to materialize.
6. Building AI responsibly requires early collaboration across UX, engineering, data science, and product.
7. AI ecosystems must be evaluated holistically—internal tools, external partners, platform compatibility, and long-term strategy all matter.
8. Compute costs for AI can be significant; PMs must understand model expenses and ROI.
9. Regulatory, compliance, latency, and accuracy requirements differ widely across industries.
10. Data pipelines, data health, and ongoing evaluation (evals) now sit at the core of modern product management.
11. AI outputs feed back into the product, creating continuous cycles of model improvement.
12. Collaboration changes because LLM responses are probabilistic, not deterministic—teams can’t predict exact outputs.
13. Designers, PMs, and engineers must align expectations, since a “valid” LLM response may still differ from the designed behavior.
14. UX flows can no longer be planned as rigid, linear steps; conversation-based outputs branch dramatically.
15. PMs developing AI skills should learn how models work, strengthen data literacy, and experiment directly with AI tools.
16. Using multiple chatbots side by side reveals how differently models behave—and why this matters for product choices.
17. B2B AI products prioritize reliability and deep integration into an organization’s workflow.
18. B2C AI products require speed, experimentation, and a high tolerance for failure.
19. B2B2C blends both worlds—demanding stability for the business and delight for the end user.
20. AI is a once-in-a-generation technological shift, and today’s PMs have the rare opportunity to shape how it transforms the world.