Module 6 — Metrics That Matter · Lesson 6.3
Work Metrics
Cycle time, estimate accuracy, KVN completeness, rework — the throughput layer
~11 min
What you'll learn
- Read cycle time at task and project granularity
- Use estimate accuracy as a calibration signal, not a performance one
- Track KVN completeness against the high-stakes baseline
- Distinguish rework from iteration
Work metrics are the throughput layer. They tell you whether the system is moving, whether the estimator is calibrated, whether the KVN is doing its job, and whether the work is shipping at the quality you intended. They are the metrics most teams overinvest in — but the right subset, well chosen, is the dashboard's center of gravity.
Cycle time — task and project
Cycle time is the time between a task entering 'in progress' and reaching 'done.' It is the most basic throughput metric and the most useful first signal.
Kavanah computes cycle time at task and project granularity. Task-level cycle time, distributed by skill tag, tells you where the team is fast and where it is slow. Project-level cycle time, computed as the duration between project kickoff and project close, tells you whether the team's project shape is sustainable.
The healthy signal is stable median cycle time with a thin right tail. Stable means the team's process is predictable; thin right tail means the long-running tasks are rare. A long right tail — many tasks that take 5x median to ship — usually indicates either poor breakdown (the task should have been multiple tasks) or unclear acceptance (the doer doesn't know when to stop).
The action when cycle time drifts: look at the right tail. The fix is almost never 'work faster'; it is 'break smaller' or 'sharpen acceptance.'
Estimate accuracy — calibration, not performance
Estimate accuracy is the rolling MAPE between estimated and actual time on completed tasks. It is the calibration signal for the estimator.
The accuracy should sit under 30% for a calibrated workspace and trend down as the team's history accumulates. The unhealthy signal is rising MAPE — the system was calibrated, and something changed (new project type, new contributor, shift in description style) and the estimator hasn't caught up.
The critical discipline is to read estimate accuracy as a calibration metric, not a performance metric. A team where one member's estimates consistently underrun is not a team with one bad estimator; it is a team that has discovered something about that member's domain that the model hasn't yet — usually that their skill is genuinely higher than declared. The action is to retag, not to retrain the member.
Kavanah publishes estimate accuracy at workspace and project levels, with per-skill breakdowns. It is a workspace-health metric, not a per-member ranking.
KVN completeness — overall and by stakes
Two metrics share this name and they should be read together.
Overall KVN completeness is the fraction of created tasks with all three task-level KVN axes populated. The right number for this is variable — depends on the mix of routine vs. high-stakes work — and a healthy workspace usually sits around 20–40%.
KVN completeness on high-stakes tasks (the agent flags these based on cost, complexity, cross-functional touchpoints) should be above 80%. If it is lower, the discipline is failing exactly where it matters most.
Reading these together protects against both failure modes: KVN theater (high overall completeness, low information value) and KVN avoidance (low completeness on high-stakes work).
The action when high-stakes completeness drops: the team has stopped writing KVN where it matters. Surface this in the next retro; usually the cause is that the team's recent run of work has been familiar enough that KVN seemed unnecessary — and a 'we don't need it' attitude on familiar work carries forward into 'we don't need it' on the next high-stakes task.
Rework rate — when work has to be redone vs. iterated on
Rework is the situation where a task that was previously closed has to be reopened and redone — not iterated upon, but redone. Iteration is healthy and expected. Rework is a quality signal.
The rework rate is the fraction of recently-closed tasks that were reopened within two weeks. A healthy workspace runs at 5–10% — there will always be some surprises. Above 15% sustained is a quality signal worth investigating.
The critical distinction is rework vs. iteration. Iteration is when the work shipped, was reviewed by the next-step stakeholder, and the feedback produced a follow-up task. The original task is closed and the follow-up is a new task; this is healthy. Rework is when the task itself comes back open because what shipped did not actually meet the stated acceptance — usually because the acceptance was unclear or the V was vague.
Kavanah's task lifecycle records the reopen events. The Reports surface distinguishes reopens-with-edit (someone reopened the existing task) from new-task follow-ups. Track the former; the latter is just iteration.
Dismissal rate — the back side of capture
Dismissal rate is the fraction of candidate tasks that get dismissed at triage. It is the counterweight to capture coverage.
A healthy workspace sits at 15–35% dismissal. Below 15% means triage is too permissive — the team is accepting work that should not have been work. Above 35% means the agent is over-proposing (probably surfacing recurring operations as candidates), or the team is too restrictive at triage.
The dismissal-reason distribution (Module 2.4) is the diagnostic. If a single reason dominates — usually a Negation — that is healthy and the boundary is doing its job. If 'no checkable outcome' dominates, the team is rephrasing topics as candidates and the workspace V probably needs sharpening.
Set up the throughput dashboard
- 1
Confirm cycle time, estimate accuracy, KVN completeness, rework rate, and dismissal rate are all displayed.
- 2
Look at cycle time's right tail
Pull the top five longest-running recent tasks. The pattern in their delays is your action item.
- 3
Check KVN completeness on high-stakes vs. all
The two-axis view is the protection against KVN theater and KVN avoidance.
- 4
Plot rework rate over time
Trending up is a quality signal worth a retro session. Distinguish rework from iteration explicitly.
Work metrics
- Cycle time (median, p90)
- Time between 'in progress' and 'done' for completed tasks.
- Healthy signal: Stable median, thin right tail. p90 within 3x median.
- Estimate accuracy (MAPE)
- Mean absolute percentage error of estimates vs. actuals on completed tasks.
- Healthy signal: Under 30%, trending down as history accumulates.
- KVN completeness — overall
- Fraction of created tasks with all three task-level KVN fields.
- Healthy signal: 20–40% for typical workspace mix; the absolute level is less informative than the breakdown.
- KVN completeness — high-stakes
- Same metric restricted to high-stakes tasks.
- Healthy signal: Above 80%.
- Rework rate
- Fraction of closed tasks reopened-with-edit within two weeks.
- Healthy signal: 5–10%.
- Dismissal rate
- Fraction of candidate tasks dismissed at triage.
- Healthy signal: 15–35%. Outside that band, investigate the dismissal-reason distribution.
Key takeaways
- ·Cycle time tells you about flow; the right tail is more informative than the median.
- ·Estimate accuracy is a calibration metric, never a per-member ranking.
- ·KVN completeness must be read as two numbers — overall and high-stakes — to avoid both failure modes.
- ·Rework rate distinguishes quality issues from healthy iteration.
Work metrics describe what is shipping. The last lesson covers the operating layer — capture latency, AI Employee contribution, and the customer-facing outcomes that everything is ultimately in service of.