Intermediate

AI Project Execution

Running AI projects requires adapting traditional agile practices to accommodate experimentation and uncertainty. Learn how to structure sprints, manage experiments, and track progress effectively.

Adapting Agile for AI

Standard agile works well for software but needs modifications for AI work:

• Experiment-driven sprints: Instead of feature stories, plan experiments with hypotheses and expected outcomes
• Longer sprint cycles: Consider 2-3 week sprints instead of 1 week, since experiments need time to run and analyze
• Dual-track development: Run data/model work and software engineering work in parallel tracks
• Accept negative results: An experiment that disproves a hypothesis is still progress. Track what you learned, not just what you shipped.

Experiment Management

Every AI experiment should be documented with:

Element	Description	Example
Hypothesis	What you believe will happen	"Adding more training data will improve accuracy by 5%"
Method	How you will test it	"Train model with 2x data, evaluate on held-out test set"
Metrics	What you will measure	"Accuracy, precision, recall, inference latency"
Results	What actually happened	"Accuracy improved 3.2%, latency increased 15ms"
Decision	What to do next	"Accept: accuracy gain outweighs latency increase"

Progress Tracking

Traditional velocity metrics do not capture AI project progress well. Instead, track:

• Model performance over time: Plot key metrics across experiments to show the improvement trajectory
• Experiments completed: Number of hypotheses tested (not just successful ones)
• Data quality metrics: Label accuracy, data coverage, edge case identification
• Technical readiness: Infrastructure, pipeline, and deployment readiness as percentage complete
• Risk reduction: Number of unknowns resolved and risks mitigated

Stakeholder Communication

Communicate progress in terms stakeholders understand:

• Translate technical metrics into business impact: "The model now correctly handles 92% of customer queries, up from 85%"
• Show the trajectory, not just the current state: "We are improving at 2% per sprint and need to reach 95%"
• Be honest about uncertainty: "We have a 70% confidence we will hit the target by the deadline"
• Present options: "We can reach 90% accuracy in 2 weeks or 95% in 6 weeks. Which matters more?"

Common Execution Pitfalls