AI can save money and time. Or it can generate a $20,000 bill overnight. What makes the difference?

Mandip Kanjiya has shipped AI-powered products in HR, insurance, and identity verification. Some worked. Some didn’t. Each one taught clear lessons for leaders who want AI to help, not hurt. In this conversation, Mandip shows why an “advisory layer” helps teams avoid leaking sensitive data, how on-device processing cuts costs and compliance exposure, and where using proven models beats trying to build your own.

Daniel and Mandip get specific: claims triage that moved faster once the workflow was redesigned, HR document generation that stayed private by templating inputs, and KYC checks that extracted only the fields needed on the user’s device. They also talk about where things went wrong and got expensive, including the real-world example of a cloud bill that spiked to $20K in a day and a routing project that failed because bad data looked clean.

🔑 What You’ll Learn in This Episode

• ✅ How to design an AI “advisory layer” that protects company and client data
• ✅ When edge processing reduces latency, spend, and compliance risk
• ✅ The difference between building AI and using AI to build solutions
• ✅ Why enabling policies beat fear-based restrictions for employee adoption
• ✅ How data quality and validation guard against confident, wrong outputs

🔗 Resources & Links

• 🌐 Communitech: https://www.communitech.ca/
• 📩 Subscribe to the Artificial Insights newsletter for key takeaways: https://manary.haus/podcast/#haus
• 👉 Have a guest in mind? Reach out to Daniel at [email protected]

💬 Thinking about bills? Share this with a colleague who signs the cloud bills and start a conversation about your edge strategy.

The hardest part of AI isn’t the technology. It’s knowing when not to use it.

Jake Walker has led products at Toyota, Otto Motors, DarwinAI, and Sportsnet. Across each role, he has seen that success does not come from using AI everywhere, it comes from matching the right technology to the problem. Sometimes that means deep learning, but often the solution is much simpler.

In this episode, Daniel and Jake explore the difference between capability, feature, and technology. Jake explains why voice and vision need AI, but why many manufacturing and streaming challenges can be solved without it. He shares how product managers should think of themselves as “value managers,” how to weigh ROI against timelines, and why being second to market can still succeed if you build distinct competencies.

🔑 What You’ll Learn in This Episode

• ✅ Why AI is a tool, not a strategy
• ✅ How to choose between AI and simpler technologies
• ✅ What makes a product manager a “value manager”
• ✅ How timelines and willingness to pay shape product decisions
• ✅ Why competitors’ strengths create urgency in your roadmap

🔗 Resources & Links

• 🤝 Connect with Jake on LinkedIn: https://www.linkedin.com/in/jakewalker123/
• 📩 Subscribe to the Artificial Insights newsletter for key takeaways: https://manary.haus/podcast/#haus
• 👉 Have a guest in mind? Reach out to Daniel at [email protected]

💬 Learn something new? Share this episode with a teammate who thinks “AI everywhere” is the answer.

What if the smartest move in AI isn’t chasing the shiniest model, but keeping things simple enough to work in production?

Anwar Jeffrey has lived that lesson. As co-founder and CTO of Glowstick, he built a revenue operations assistant that helps post-sales teams detect expansion opportunities, churn risks, and customer needs hidden in everyday conversations. But instead of leaning on expensive neural nets or large language models, his team focused on what he calls “dumb AI”: intent detection, sentiment, and similarity signals combined in pragmatic ways that delivered real value.

In this conversation, Daniel and Anwar unpack what it takes to bring an AI product from idea to adoption. They talk about why data is always messier than you expect, how to earn trust by meeting users inside the tools they already use, and why simplicity often beats sophistication when it comes to scaling.

🔑 What You’ll Learn in This Episode

• ✅ Why Glowstick chose simple, explainable AI over LLMs
• ✅ How messy customer data shapes product architecture
• ✅ What makes generalizability and scalability different challenges
• ✅ Why user trust depends on where insights live, not just how accurate they are
• ✅ How “dumb AI” can reach production faster and survive long enough to matter

🔗 Resources & Links

• 🤝 Connect with Anwar on LinkedIn: https://www.linkedin.com/in/anthonyajeffrey/
• 💻 Explore Azimuth (open-source tool Anwar’s team used): https://github.com/Servicenow/azimuth
• 💻 Explore Baal (Bayesian Active Learning): https://github.com/baal-org/baal
• 📩 Subscribe to the Artificial Insights newsletter for key takeaways: https://manary.haus/podcast/#haus
• 👉 Have a guest in mind? Reach out to Daniel at [email protected]

💬 Learned something here? Share this episode with someone building AI who might need permission to keep it simple.

Host Daniel Manary delves into the essential aspects of creating AI products that meets people's needs. Daniel explores the motivation behind building impactful AI solutions, then questions one of the most groundbreaking technologies in our lifetime about what makes it so impactful.