Product Talk How to Avoid Being Another Failed AI Project: AI Architect & Strategy Lead
9 snips
Feb 23, 2026 
Greg Nudelman, Product and UX leader with 16+ years in AI and creator of the Snowball Sprint, shares frameworks for reducing AI project risk. He covers why many AI initiatives fail, how to frame the right use cases, build fast thin prototypes with real data, and restructure teams and processes to escape POC purgatory.
AI Snips
Chapters
Books
Transcript
Episode notes
AI Requires Probabilistic Product Processes
- AI is probabilistic, not deterministic, so traditional deterministic product processes (hand-offs, specs, long waterfalls) often fail for AI projects.
- Greg Nudelman warns teams must 'surf the wave' with iterative, adaptive workflows rather than laying fixed train tracks to avoid multi-month, multi-million failures.
Define Use Case Data Metrics And Process First
- Do explicitly define the AI use case, data readiness, success metrics, and process before building to avoid common failure modes.
- Greg lists the four 'horsemen': wrong use case, unready data, undefined success (ROI/acceptable error), and using deterministic processes for probabilistic systems.
Frame The Problem With Storyboard Digital Twin And Value Matrix
- Do frame the problem before coding: storyboard the agent flow, build a digital twin of handoffs, and run a value matrix to quantify false positive/negative costs.
- Greg says these three exercises take a few hours and reveal whether you have the right question, data, and ROI constraints before building.
