In a Queue Model, a 21% Rise in PR Volume Quadruples Review Wait Time: From 2 to 8 Hours
AI assistants like Copilot speed up writing code, but review, testing and release run at the same pace as before. In a simplified queue model, an increase in volume from 14 to 17 pull requests a day (21%) raises review time from 2 hours to 8. The code takes two hours to write, then the change waits days for checks. So measure review and QA load, not just how fast code gets written.
AI-processed from Habr AI; edited by Hamidun News
A developer with Copilot or Cursor writes a function in two hours instead of eight, opens a pull request (PR, a request to merge code) and... waits. The reviewer is busy, tests are stuck in a queue, and the next release window is Thursday.
Code has become cheaper, but delivery to users has not gotten faster. Using a simplified queue model, the author of a Habr analysis shows the scale of the problem: if daily PR volume grows from 14 to 17, or 21%, the average time a PR spends in review rises from 2 to 8 hours, a fourfold increase. The takeaway for you: if you adopt AI tools and look only at how fast code gets typed, you may be speeding up the wrong part of the system.
Does AI actually speed up development?
The data is mixed. In a controlled 2023 experiment by GitHub and Microsoft, the Copilot group wrote an HTTP server in JavaScript 55.8% faster on average than the control group. But the study had 95 developers recruited from Upwork, only 35 finished the task, and the task was built from scratch.
In 2025, METR took 16 experienced open-source developers and 246 real tasks in projects they knew well. With AI allowed, tasks took 19% longer on average. Before the experiment, participants expected a speedup, and afterward they still believed AI had helped them. In February 2026, METR ran a follow-up study with a larger sample. It pointed to a possible speedup, but some developers refused to do tasks without AI, and the authors could not reliably estimate the effect. They have redesigned the experiment.
A narrow assumption is enough for the rest of the argument: if AI makes some changes cheaper, load inside an engineering organization gets redistributed.
The bottleneck moves, and the queue grows in a jump
Take a team of ten people, each producing about two PRs a week, roughly 20 in total. Review, QA and releases are tuned for that volume. AI doubles output to 40 PRs. The review queue grows, so the company adds automated review. Then the flow hits QA, which also gets automated, and changes start waiting for the release window. The bottleneck does not disappear. It moves to wherever capacity runs short.
The nonlinearity shows up in an M/M/1 model: one review channel, a random flow of tasks, and a capacity of 18 PRs a day. Average time in the stage is W = 1 / (μ − λ), where λ is the number of incoming tasks and μ is the number processed. The values are modeled, not measured:
- 9 PRs a day, 50% load: 53 minutes;
- 12 PRs, 67%: 1 hour 20 minutes;
- 14 PRs, 78%: 2 hours;
- 16 PRs, 89%: 4 hours;
- 17 PRs, 94%: 8 hours.
A real team is more complicated: there are several reviewers, tasks differ, and there are priorities and repeat checks. A forecast would need actual data and a multi-server model. The qualitative conclusion holds: a stage running near its limit slows sharply with any added load. That is why a change written in two hours can wait several days for review and tests. Individual speed has gone up, but the organization's Lead Time (the time from idea to production) has stayed the same.
DORA reports are consistent with this picture. In 2024, a 25% rise in AI adoption was associated with an estimated 1.5% drop in delivery throughput and a 7.2% drop in delivery stability. In 2025, the link to throughput turned positive, while the stability problems remained. This is a correlation and does not prove causation. But the practical conclusion is clear: AI amplifies what is already there. Teams with small changes and fast feedback will speed up, while teams with long queues and manual approvals will feel the problems more.
What to change: risk-based checks and independent verification
The author proposes rethinking the delivery architecture rather than just automating each stage. Today, a fix to an error message and a new commission-calculation algorithm go through the same pipeline. In the Risk-Adaptive Delivery model, a mandatory minimum applies to every change, and additional checks depend on what the change touches. For documentation, format and link checks are enough. A database schema change needs migrations and backward compatibility, authorization needs a security review, and financial calculations need independent verification of business invariants.
The approach has two traps. The risk classifier can be wrong: an edit labeled "safe documentation" was taken for text, but it touched an executable config, and the lightweight pipeline let it through. So the classifier needs its own checks, monitoring and conservative behavior under uncertainty. Also, a language model's answer is not proof of correctness, and critical constraints should be set by formal policies.
AI agents introduce a trust problem. If one agent writes the code and the tests and also interprets the results, it can fit the criteria to its own implementation, and without malicious intent: it misunderstood the requirement and wrote tests that confirm the error. A second AI acting as reviewer does not help if both work from the same incomplete specification. What's needed are acceptance criteria fixed in advance, business invariants, deterministic checks and a human for critical changes. The executor must not be able to change security policies, and runs should happen in a sandbox with minimal privileges and temporary credentials.
Finally, money. The full cost of a change adds up from implementation, verification, delivery, and the expected costs of failures and rework. AI review, agent-based tests and isolated environments also cost tokens and compute. A cheap check that misses a critical error costs more than a thorough one, while a full set of expensive checks on every small edit does not reduce risk proportionally.
What to do in practice: measure review and QA load, not just how fast code gets written. Keep spare capacity and do not run at 90% load. Plan for rollback: code can be rolled back, but changed data and external transactions cannot always be. The author sees the next bottleneck as understanding the consequences of a change in a system of many services, and proposes a machine-readable model of the system to address it. That is his engineering hypothesis, not a proven fact.
- Nov 19, 2026Anthropic goes public and raises $81.3 billion, beating SpaceX's $75 billion record
- Nov 19, 2026Amazon and Google's Anthropic Stakes Valued at $308 Billion in IPO at $83 a Share
- Nov 19, 2026A GitHub Script Catches App Store and Google Play Rejection Reasons Before Review: 542 Stars in a Day
- Nov 19, 2026FBI: AI Tool Artex Breaches Companies in Under Three Hours
Key Questions Answered
Does AI actually speed up development?
The data is mixed. In a controlled 2023 experiment by GitHub and Microsoft, the Copilot group wrote an HTTP server in JavaScript 55.8% faster on average than the control group. But the study had 95 developers recruited from Upwork, only 35 finished the task, and the task was built from scratch. In 2025, METR took 16 experienced open-source developers and 246 real tasks in projects they knew well. With AI allowed, tasks took 19% longer on average.
What are the key details on The bottleneck moves, and the queue grows in a jump?
Take a team of ten people, each producing about two PRs a week, roughly 20 in total. Review, QA and releases are tuned for that volume. AI doubles output to 40 PRs. The review queue grows, so the company adds automated review. Then the flow hits QA, which also gets automated, and changes start waiting for the release window. The bottleneck does not disappear. It moves to wherever capacity runs short.
What to change: risk-based checks and independent verification?
The author proposes rethinking the delivery architecture rather than just automating each stage. Today, a fix to an error message and a new commission-calculation algorithm go through the same pipeline. In the Risk-Adaptive Delivery model, a mandatory minimum applies to every change, and additional checks depend on what the change touches. For documentation, format and link checks are enough.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.