Individual Speedup Is Not Organizational Speedup
AI makes coding faster but delivery does not follow: the median drops while P85 stays put. The cycle inflates in waiting — govern organizational friction in four steps.
After AI coding lands, individual engineer output visibly rises, but the gains do not reach the end-to-end delivery cycle: the median cycle keeps improving while P85 stubbornly sits above three two-week sprints. That is not "AI does not work" — the bottleneck moved. AI accelerates value creation; organizational friction sits untouched, and the friction is the bulk of the cycle.
After reading, you should be able to answer:
- Why does watching only the median misread organizational health?
- What does each of the four governance steps (visible, trusted, located, mined) solve?
- How do AI and the PM divide the work in friction governance?
Coding got faster; why did delivery not
Write the formula down first: delivery cycle = value creation time + organizational friction time, and organizational efficiency is the former over their sum tencent-pm-friction. Friction — waiting for scheduling, resource coordination, cross-team sync, integration queues, bug rework — does not concentrate in one stage; it accumulates quietly as a requirement moves between stages. Four real delivery cases share one pattern: the actual coding time is short, and the bulk of the cycle is spent on coordination.
So watching only the median misjudges. An improving median says "smooth tasks got smoother"; a flat P85 says "stuck tasks stay stuck." The target of governance is the tail, and the tail is where organizational friction hides — this is the delivery-side mirror of measuring the right metrics: watch the tail, not only the median.
Four steps: visible, trusted, located, mined
Governance does not start with "more meetings and more nudging"; it starts with making the process observable tencent-pm-friction:
| Step | What it does | Why it is the precondition |
|---|---|---|
| 1. Visible data | Standardize the delivery process into a model: unified stages, responsible roles, delivery conditions, split to "one task, one owner" | If the organization's operation is unobservable, governance has no object |
| 2. Trusted data | Shift state tracking from manual reporting to system-event inference: task state changes, review completion, test transitions, release checkpoints drive it automatically | AI analyzing wrong data produces nothing but noise |
| 3. Anomaly location | AI patrols continuously: stage-wait timeouts, requirements stuck in place, abnormal load, single-point dependency risk | People go from digging through logs after the fact to receiving proactive alerts |
| 4. Bottleneck mining | Biweekly or monthly analysis of lead time, waiting duration, quality trends, risk distribution | From solving single problems to optimizing the delivery system |
The order cannot be skipped. Anomaly location without visibility sends the AI patrolling thin air; bottleneck mining without trusted data makes decisions on wrong calibers.
AI finds the problem; the PM decides the cure
One sentence for the division of labor: AI owns situational awareness and problem exposure; the PM owns judging causes, deciding reallocation, and driving resolution tencent-pm-friction. AI's strength is continuously watching the waiting and flow of hundreds of tasks and putting "this requirement has been stuck 12 days, at this stage" in front of you; judging why it is stuck, whether to jump the queue, and who to call remains human work.
The PM's role therefore shifts from process management (is anyone on the task, is any milestone late) to efficiency governance (continuously finding and removing process friction) — from pushing single tasks in the queue to governing the queue itself. A few PM skills can crystallize alongside: workflow and automation-rule governance, a morning snapshot plus evening review health patrol, and scheduling assistance that presents structure without replacing decisions.
What it looks like after governance
Tencent Health's four-month measured numbers tencent-pm-friction: median delivery cycle 19 days to 9; P85 from 52 days to 23; predictability (the gap between P85 and the median) narrowed from 33 days to 15; the share of the 30-day-plus tail from 35% to 7%, 60-day-plus from 11% to 0; problem detection moved from one or two weeks after the fact to 1–3 days after a risk appears.
Note the structure of those numbers: the median halved, and the tail was nearly cut off. Individual speedup — faster coding — usually improves only the median; what pushed down P85 and the tail is friction governance itself. Do both, but do not present one as evidence of the other.