How to Avoid Being Another Failed AI Project: AI Architect & Strategy Lead

Why do 85% of AI projects fail, and how can product leaders beat the odds? In this episode of the Product Talk podcast, host Denise Hemke sits down with Greg Nudelman, product and UX leader and creator of the Snowball Sprint, to share how to avoid being another failed AI project. Drawing on real-world examples from cybersecurity, enterprise AI, and IoT, Greg shares practical frameworks for framing the right use cases, thin-slicing real data, escaping POC purgatory, and redefining success beyond accuracy. A must-listen for product managers and leaders navigating AI-driven product strategy.

Join us for new conversations with leading product executives every week. Roll through the highlights of this week’s podcast below, then head on over to our Events page to see which product leaders will be joining us next week.

Show Notes:

1. Most AI projects fail due to process, not technology. The often-cited 85% failure rate is largely driven by flawed product thinking and execution patterns, not by weak models or immature tools.
2. AI is probabilistic, not deterministic. Traditional product development assumes predictable outputs; AI systems behave differently and require experimentation rather than rigid planning.
3. Teams rush to solutions without nailing the exact use case. Broad problem statements aren’t enough—AI requires a sharply defined, narrow application to succeed.
4. Data readiness is widely overestimated. Many teams believe they “have the data,” but it’s often unclean, unstructured, or misaligned with the specific question the AI must answer.
5. Success metrics are rarely defined upfront. Teams focus on “accuracy” without clarifying acceptable error rates, business impact, or the cost of false positives and false negatives.
6. Accuracy alone is the wrong KPI. In real-world AI systems, optimizing for accuracy can reduce business value. ROI and outcome impact matter more than model purity.
7. The cost of being wrong must be quantified. AI systems will fail at some point; leaders must understand the economic and operational implications of those failures.
8. Traditional Agile handoffs break down in AI projects. Writing specs and passing them to developers assumes predictability, which doesn’t align with AI’s iterative, discovery-driven nature.
9. Time is the most expensive failure. Six to nine months spent scaling the wrong solution is more damaging than the financial investment itself.
10. Product-market fit must precede scaling. Teams often overinvest in infrastructure before validating whether customers truly value the AI capability.
11. Thin-slice validation is critical. Start with a narrowly scoped dataset or subset (e.g., 50 rules instead of 1,000) to prove value before expanding.
12. Frame the problem before writing code. The Snowball Sprint begins with three exercises: storyboarding the agent flow, mapping the digital twin workflow, and building a value matrix.
13. Storyboarding clarifies agent-human interactions. Teams must explicitly define what the AI does, what inputs it receives, and where handoffs occur.
14. Digital twin mapping forces operational clarity. Documenting every data format and handoff reveals gaps that abstract planning misses.
15. The value matrix replaces the confusion matrix. Instead of focusing on model metrics, calculate real-world impact by multiplying outcome frequency by cost.
16. Build the real thing—not a POC with synthetic data. Proofs of concept often rely on artificial inputs that mask real constraints and create false confidence.
17. Rapid iteration should happen against live AI with real data. Feedback loops should occur in days—not months—with actual outputs customers can react to.
18. UI polish is secondary in early validation. If the AI delivers meaningful value, customers will tolerate rough interfaces; content and outcome matter more.
19. Team roles are becoming more fluid. AI lowers technical barriers, enabling product, UX, and engineering to collaborate more directly rather than operate in silos.
20. Land and expand creates a learning system. Once deployed in a thin slice, the AI begins generating real feedback data, enabling refinement, scaling, and long-term defensibility.