← Delivery system

AI made writing code faster. The bottleneck moved.

A practical diagnosis of how faster code production shifts pressure into review, testing, release, and team coordination.

20 August 2026

More output does not remove the constraint

AI coding tools can reduce the effort needed to create a first implementation. That local improvement changes the flow of work. It does not automatically increase the capacity of every step that follows.

Google Cloud’s summary of the 2025 DORA report found a positive relationship between AI adoption and delivery throughput, while stability remained a concern. The report’s explanation is practical: higher change volume can expose weak control systems downstream. Strong automated testing, mature version-control practices, and fast feedback loops matter more when work arrives faster. Read the DORA summary.

The bottleneck did not disappear. It moved to the part of the system that cannot absorb the new rate of change.

Four places to look

Review capacity

A pull request can be generated quickly and still require concentrated human attention. If change volume rises while reviewer capacity and review evidence stay the same, the queue grows.

The useful question is not “How many pull requests did AI help create?” It is “How long does a change wait for a confident decision?”

Improve the signal that arrives with the change: a clear intent, bounded scope, tests, screenshots where relevant, risk notes, and an explanation of decisions that are hard to infer from the diff.

Test confidence

More code creates more behavior to verify. If test feedback is slow, flaky, or narrow, faster implementation can simply create a larger pile of uncertain changes.

AI can help generate tests, but generated tests are not proof by themselves. The team still needs acceptance criteria, meaningful coverage, and a way to determine whether the test is checking the intended behavior rather than repeating the implementation.

Release controls

A team may merge faster and still release at the same rate because deployment, compliance, change approval, or production verification remains manual and fragile.

This is not always a problem. Some systems require deliberate release controls. The problem is leaving the control implicit, so nobody can distinguish necessary governance from inherited delay.

Coordination

Local speed creates more decisions for the team: which changes matter, what should be reviewed first, how parallel work interacts, and whether the product still moves toward a shared outcome.

If planning and coordination remain informal, additional code can increase work in progress without increasing delivered value.

Measure queues, not only activity

Coding time is easy to notice because it happens in front of the person using the tool. Queue time is less visible and often more important.

For one representative change, record:

  • time waiting for clarification before implementation;
  • time from first commit to review start;
  • time spent in review and rework;
  • time waiting for test feedback;
  • time from merge to production;
  • defects or reversals found after release.

The goal is not to create a new reporting bureaucracy. It is to see which stage now governs the pace of the whole system.

Change the workflow at the constraint

Once the constraint is visible, improve that stage before increasing upstream volume again.

  • Reduce batch size so reviewers can reason about changes.
  • Require evidence with AI-assisted work instead of accepting unexplained output.
  • Move automated feedback earlier.
  • Make ownership explicit after the AI completes a task.
  • Put human approval at the points with real accountability.
  • Feed failures back into shared prompts, skills, checks, and examples.

NIST’s DevSecOps guidance similarly recommends human validation, verifiable processes, and traceability for AI-assisted development. Read the NIST guidance.

The practical starting point

Choose one value stream and map it from request to production. Mark the queues, handoffs, evidence, and approvals. Then decide where AI should accelerate work and where the surrounding control system must change first.

A private workshop is designed to do this on the team’s real stack. The output should not be a slide deck about AI. It should be a working operating pattern the team can continue to own.

Sources

Apply it to your team

Find the operating change behind the tool change.

Describe your bottleneck