Measure AI Workflow Impact with a Cumulative Flow Diagram (CFD)
A cumulative flow diagram (CFD) shows how work accumulates and moves through each stage of delivery. To see AI's true impact, tag which workflow stages are AI-assisted and plot a CFD with those stages highlighted.
Then watch three things week over week:
- The slope of "done" for throughput
- The thickness of each stage band for work in progress
- The horizontal distance between "started" and "done" for Lead Time for Changes
If an AI-assisted stage gets faster but the next stage band widens or the "done" slope does not improve, AI saves local time but shifts the bottleneck. This approach aligns with Kanban guidance that CFDs visualize WIP, cycle time, and throughput in one view, which is why they are used to manage flow at the team level Kanban Guide for Scrum Teams and Professional Scrum with Kanban glossary.
What is a cumulative flow diagram and why use it for AI work?
A CFD is a stacked area chart of work-item counts over time. Each colored band is a workflow state, the top line is total items in the system, and the horizontal distance between "started" and "done" approximates average lead time. CFDs are recommended because they let teams see WIP, throughput, and cycle time together rather than in separate charts Kanban Guide for Scrum Teams.
The reason this matters for AI is simple queueing math. Little's Law connects average WIP, average throughput, and average lead time. If you reduce time in one stage but do not increase downstream capacity, WIP piles up and end-to-end lead time does not improve MIT lecture notes on Little's Law. The same principle is used in software process flow research to expose waiting waste and architectural bottlenecks IEEE Software article referencing Little's Law via SEI.
How to instrument a CFD for AI-assisted stages
- Define workflow states clearly: for example backlog, ready, in progress, review, verify, release. Make policies explicit so movement between states is auditable Kanban Guide for Scrum Teams.
- Tag AI-assisted work: add a boolean or label on items touched by code generation, AI test authoring, AI review, or AI release tooling, such as
ai_assisted=trueorai=codegen. - Mark AI-assisted stages: add a property on the stage definition itself, such as
ai_assisted=truefor review automation or test generation. - Render the CFD with overlays: keep the classic stacked bands, then:
- outline AI-assisted bands to compare their thickness over time
- add a second series for throughput from "done" to see if the slope changes
- overlay control lines for WIP limits on AI-assisted stages
- Read the picture: if an AI-assisted band narrows and the next band widens, AI accelerated local work but moved the constraint. If the "done" slope steepens while total WIP does not grow, AI improved end-to-end flow.
This mirrors Scrum with Kanban's advice to use CFDs to inspect WIP and throughput routinely when optimizing flow Kanban Guide for Scrum Teams and Scrum.org's flow metrics guidance that relates CFD band shape to throughput and aging signals 4 key flow metrics and how to use them.
What to look for in the chart
- Throughput slope: the steeper the "done" line, the more items finished per time unit. If AI assistance is working, this slope should increase after a short learning window.
- Band thickness at AI-assisted stages: thickness is WIP. Narrowing shows faster pull-through. Widening shows a pileup.
- Lead time gap: the horizontal distance between "started" and "done." A narrowing gap signals true lead-time improvement.
- Arrival versus departure parallelism: if the top line of the CFD grows faster than "done," intake exceeds completion and the risk of backlog growth and aging increases.
- WIP limit adherence: frequent flat sections in a band show periods with no movement in that state. If other bands keep growing while this one is flat, work is likely blocked or misbalanced and you may need to adjust WIP limits or capacity.
These interpretations follow the standard CFD reading that Kanban practitioners teach and that Scrum.org documents for teams combining Scrum and Kanban Kanban Guide for Scrum Teams and 4 key flow metrics and how to use them.
Metrics to pair with your CFD
Use these lightweight measures with a weekly view. Tie them to how AI changes your system, not just a single step.
| Metric | Definition | How AI affects it | Decision cue | Related glossary |
|---|---|---|---|---|
| Stage WIP | Average items in a stage | Should fall in AI-assisted stages | Falling WIP with stable or higher departure rate is healthy. Rising WIP signals a pileup | Lead Time for Changes |
| Departure rate | Items exiting a stage per day | Should rise where AI automates work | If departure rate rises upstream but not downstream, rebalance capacity | Pipeline Run Time |
| WIP age p90 - AI-assisted stages | 90th percentile time spent in stage | Should drop | Diagnose why the AI assistance doesn’t have the expected impact | Flow Efficiency |
| WIP age p90 - non-AI-assisted stages | 90th percentile time spent in stage | Likely rises in stages after the AI-assisted stages | Rising age means hidden waiting, so adjust WIP limits, tools, and/or capacity | Review Latency |
| Flow efficiency | Active time divided by total elapsed time | AI should reduce active time. The end-to-end effect depends on waits | Improve handoffs if efficiency does not rise with AI | Lead Time for Changes |
Little's Law explains why these metrics move together. WIP, throughput, and lead time form a simple relationship, so local speed-ups only help if WIP is controlled and downstream capacity exists MIT lecture notes on Little's Law. SRE practice offers a complementary lesson: measure outcomes with clear SLIs and act when targets are missed, which is directly applicable when AI changes a delivery stage SRE workbook chapter on implementing SLOs.
Example decisions the CFD should enable
- We sped up code review with AI suggestions: the review band narrowed, the validating stage widened, and the lead time gap did not shrink. Add verification capacity or decouple slow checks.
- We introduced AI-assisted validation tools: verify narrowed, and the "done" slope steepened without total WIP growth. Keep investing and raise WIP limits one step downstream to avoid starving releases.
- We added an AI release bot: the release band narrowed yet the "done" slope is unchanged. Intake outpaces completion. Reduce arrival rate or increase downstream capacity.
These decisions reflect standard flow control with explicit WIP, CFDs, and outcome guardrails Kanban Guide for Scrum Teams and are consistent with using SLIs and SLOs to evaluate whether changes improve user-visible outcomes SRE workbook chapter on implementing SLOs.
FAQ
How should AI-assisted work be tagged in a CFD?
Tag both the workflow states and the items. Add a property on the workflow state definition and a label on any item touched by AI, so you can compare band thickness, departure rate, and WIP age for AI versus non-AI work Kanban Guide for Scrum Teams.
What shows that AI improved lead time rather than moved the bottleneck?
The "done" line gets steeper, total WIP stays flat or falls, and the horizontal distance between "started" and "done" narrows. If the next band widens, the bottleneck shifted, so use Little's Law to rebalance WIP and capacity MIT lecture notes on Little's Law.
Do WIP limits need to change when AI speeds a stage?
Often yes. If departure rate rises upstream, raise downstream WIP limits temporarily and watch WIP age and departure rate to avoid starvation or pileups Kanban Guide for Scrum Teams.
What outcome should be monitored while experimenting with AI?
Track SLIs such as deploy success and quality measures alongside CFD-based flow measures so you do not trade flow speed for reliability regressions SRE workbook chapter on implementing SLOs.