We’ve all been trained to love the predictability of a well-groomed agile roadmap. Story points, velocity, sprint commitments – it’s the bedrock of modern product development. But what happens when the product you’re building is powered by AI?
I’m seeing a major clash between traditional agile practices and the realities of machine learning development. A standard roadmap promises deterministic features: “As a user, I want to click a button to export my data.” It’s a solvable, predictable engineering task.
AI development, on the other hand, is probabilistic. The goal isn’t just to build a feature, but to achieve a target level of performance that may or may not be possible with the current data and approach. You can’t just schedule a research breakthrough for Q3. Committing to shipping a model with “95% accuracy” in a specific sprint is setting your team and your stakeholders up for failure. It treats a scientific experiment like a feature factory.
This friction leads to missed deadlines, frustrated engineers, and a loss of stakeholder trust. We need to shift our thinking from roadmapping features to roadmapping outcomes and capabilities. Instead of timelines for solutions, we should be creating timelines for experiments designed to answer key questions and de-risk our model development.
How are you adapting your roadmapping and sprint planning processes to accommodate the uncertainty of building AI-powered products?