Kanban Systems Improvement: Reading Flow and Evolving the System
Kanban Systems Improvement tests whether you can use the kanban system as an improvement engine: read flow metrics, diagnose system problems, run feedback loops, and evolve policies without destabilizing delivery.
Improve the system, not just the board
Manage flow end to end. Track lead time distribution, throughput, WIP, flow efficiency, due-date performance, and blocker frequency. Use a cumulative flow diagram to see where work accumulates, and a control chart to separate normal variation from special-cause signals. Define service level expectations (SLEs) from historical lead-time percentiles, and treat them as forecasts, not deadlines. Improvement is evolutionary: change one policy, run an experiment, check the metric, and keep or revert.
Feedback loops and cadences
Cadences connect strategy to delivery: daily standup for flow and blockers; replenishment for intake; service delivery review for customer-facing performance; operations review for system capability and cross-team flow; risk review for classes of service, dependencies, and compliance; strategy review for purpose and portfolio. The Kanban Maturity Model can guide which practices and behaviors to strengthen next, but it is not a scorecard to chase.
Worked example
A team has 12 items in progress and throughput of 4 per week, so average lead time is 12 ÷ 4 = 3 weeks. Over four weeks, WIP rises to 16 while throughput stays at 4. The CFD shows the in-progress band thickening and the done line unchanged. Little's Law predicts lead time will rise to 16 ÷ 4 = 4 weeks. The team should not add people first; it should enforce the downstream WIP limit, swarm on the oldest items, and remove the blocker causing the test queue to grow. After WIP returns to 12, expected lead time returns to 3 weeks, assuming throughput holds.
Traps
Celebrating higher throughput while lead time worsens; using averages instead of distributions; treating SLE as a contractual SLA; adding WIP to go faster; optimizing one column; running cadences as status meetings; changing many policies at once; and ignoring variability and blockers.