In October 2026, Harvard researchers Fiona Chen and James Stratton published a working paper with a title that should have been read by every engineering leader in America: Artificial Intelligence in the Firm: Bottlenecks in Software Production.
The paper analyzed data from more than 700 firms and over 700,000 workers between 2021 and March 2026. It found that AI coding agents increased the amount of code produced by 30 percent. Lines of code rose by 30 percent, code commits by 20 percent, and pull requests by 23 percent.
The increase in completed software was statistically indistinguishable from zero.
The code had to be reviewed by humans — and the review process could not absorb the additional volume. The time required to review and merge a pull request increased by 49 percent, from an average of 7.03 days to 10.48 days. The number of reviewer comments per pull request rose 35 percent. The share of employees doing code reviews rose 14 percent, as organizations pulled more people into the bottleneck rather than relieving it.
"The productivity gains are absorbed by downstream constraints in the production process," the authors wrote. "Code review becomes the bottleneck."
The Verification Tax
The Harvard finding confirms a pattern that has been building across the industry for more than a year.
Faros AI's telemetry analysis, covering 4,000 teams and 22,000 developers over two years, found that while task completion per developer rose 34 percent and code activity surged 210 percent, the costs were equally dramatic. The code rewrite rate increased 861 percent. The ratio of production incidents per pull request rose 242.7 percent. The average pull request review time extended by 441.5 percent. The proportion of pull requests merged without any review at all increased 31.3 percent.
The pattern is consistent across datasets and methodologies. AI writes more code. Humans review it. The review process cannot scale at the same rate. The result is a bottleneck that shifts the constraint from writing to verifying, without reducing the total time required to ship software.
A survey of 1,528 developers and technology buyers across six countries found that 85 percent agreed AI had shifted the bottleneck from writing code to reviewing and validating it.
The constraint lies less in what the model can generate than in what the organization can verify.
The 80 Percent Adoption Paradox
The Harvard paper included a detail that makes the bottleneck more difficult to dismiss: by March 2026, approximately 80 percent of the firms studied were using some form of AI-assisted code review.
The tools were already deployed. The bottleneck persisted anyway.
That outcome is not surprising when you look at what the tools actually do. Most AI code review products — CodeRabbit, Greptile, Qodo, Cursor Bugbot, Anthropic's Code Review — are designed to find bugs faster than humans can. They are not designed to replace the human judgment required to decide whether a change should be merged. They accelerate one step in the process while leaving the final approval bottleneck intact.
Anthropic's Code Review, launched in March 2026, dispatches a team of agents to inspect pull requests in parallel and rank findings by severity. Its internal testing showed that on large pull requests of more than 1,000 lines, 84 percent received findings, averaging 7.5 issues per review. The tool is priced at $15 to $25 per pull request.
The tool is useful. It does not eliminate the bottleneck. The Harvard data suggests that adding AI to the review process changes the shape of the bottleneck rather than removing it.
The Code Volume Problem
The bottleneck is not only about review capacity. It is about the nature of the code being reviewed.
AI-generated code is 2.6 times larger per pull request than human-written code, according to LinearB's 2026 benchmarks, which analyzed more than 8 million pull requests from 4,800 teams across 42 countries. The time required to review it is 5.3 times longer before a reviewer even picks it up. The acceptance rate is 32.7 percent, compared with 84.4 percent for human-written code.
A separate analysis found that AI-generated pull requests contain 1.7 times more issues per review — 10.83 issues versus 6.45. The logic and correctness defects in AI code rose 75 percent compared with human code.
The review bottleneck is also a quality problem. The code arriving at the reviewer's desk is larger, buggier, and more likely to be rejected. The reviewer has to work harder to evaluate a change that is less likely to be accepted.
The Harvard data captures the consequence. When code volume rises 30 percent and review time rises 49 percent, the review queue grows faster than the codebase. The engineering organization accumulates a backlog of unmerged changes, unreviewed diffs, and deferred decisions.
The productive capacity of the organization does not increase by 30 percent. It increases by the amount of code that can be reviewed and merged — which, in the Harvard data, is approximately zero.
What the Bottleneck Actually Costs
The financial implications are not theoretical.
A mid-sized engineering organization with 500 developers spends roughly $75 million annually on engineering salaries, based on industry averages. If AI increases code output by 30 percent but the review bottleneck absorbs the gain, the organization has paid for the AI tools, the additional compute, and the integration work without capturing the productivity improvement.
The cost of the bottleneck is not just the AI spend. It is the opportunity cost of the features that were not shipped, the bugs that were not fixed, and the products that were not launched because the review queue was too long.
The Harvard paper notes that the bottleneck may also be affecting hiring decisions. The researchers found that the share of employees doing code reviews rose 14 percent, but total employment did not change significantly. Organizations are not hiring more engineers to review AI-generated code. They are pulling existing engineers into the review process, taking them away from writing code, designing systems, and building features.
The bottleneck is reallocating senior engineering time from creation to verification. That is the most expensive resource in the organization, and it is being consumed by a task that AI was supposed to eliminate.
The Two Responses
The industry has split into two approaches to the bottleneck, and the split mirrors a broader divergence in how American and Chinese engineering organizations are responding to AI-assisted development.
The American approach has focused on buying tools. CodeRabbit, Greptile, Qodo, Cursor Bugbot, and Anthropic's Code Review are all products designed to accelerate the review process. The market for AI code review tools has grown rapidly, with pricing ranging from $1.00 to $1.50 per run for Bugbot to $24 per developer per month for CodeRabbit, and $15 to $25 per pull request for Anthropic's offering.
The American model assumes that the bottleneck is a throughput problem. If reviewers had better tools, they could process more code faster. The tools are designed to find bugs, rank severity, and reduce the time a human spends on each review.
The Chinese approach has focused on changing the pipeline. Chinese engineering organizations are embedding review into the CI/CD pipeline as a mandatory gate. If the code fails a quality or security threshold, it does not merge.
Alibaba's open-source Open Code Review is a CLI tool that combines deterministic pipelines for file selection, bundling, and rule matching with an LLM agent for code understanding. It is designed to be embedded in the pipeline, not used as a separate product. It had 20,000 monthly active users inside Alibaba before it was open-sourced, with an adoption rate above 30 percent and a false positive rate below 5 percent.
Tencent's WorkBuddy executes review and test generation automatically when code is submitted, and blocks non-compliant changes through a merge gate. The company's guidance is explicit: review, testing, and gating should be embedded into the pipeline, so that quality checks become a pre-merge requirement rather than a post-hoc human process.
Kitta, a code review agent deployed in the OpenAnolis community, has reviewed more than 70,000 patches, with a gate accuracy rate of 97 percent and a merge rate of 62 percent. The model is not just flagging issues. It is deciding whether the code is allowed to merge.
The Chinese model assumes that the bottleneck is a process problem. If review is mandatory and automated, the organization cannot accumulate a backlog of unreviewed code. The gate prevents the queue from forming.
Which Approach Works
Neither approach has solved the bottleneck entirely. The American tools accelerate review but do not eliminate the human decision point. The Chinese gates enforce review but do not necessarily reduce the time required to pass the gate.
The data suggests that the Chinese approach may be more effective at preventing the queue from forming. If review is a mandatory gate embedded in the pipeline, there is no opportunity for unreviewed code to accumulate. The code either passes or it does not. The queue is not a backlog of pending reviews. It is a backlog of code that failed the gate and needs to be fixed.
The American approach may be more effective at improving the quality of individual reviews. A tool that finds bugs faster and ranks them by severity allows a human reviewer to focus on the highest-risk changes. But the tool does not prevent the queue from growing.
The Harvard data suggests that neither approach has fully addressed the underlying constraint. The review bottleneck is a mismatch between the rate at which AI can generate code and the rate at which humans can understand it.
What the Bottleneck Reveals
The code review bottleneck is a structural feature of AI-assisted software development.
AI can generate code faster than humans can review it. That gap is widening as models improve and generate more code per prompt. The organizations that manage the bottleneck are not the ones with the best review tools. They are the ones that have redesigned their pipeline to prevent the bottleneck from forming.
The Harvard paper's conclusion is blunt: "The productivity gains are absorbed by downstream constraints in the production process." The constraint is not the AI. It is the organization's capacity to verify what the AI produces.
The companies that solve this problem will treat review as a pipeline requirement, not a product purchase. The companies that do not will keep buying tools, hiring reviewers, and wondering why 30 percent more code has not produced 30 percent more software.
Sources: Harvard working paper "Artificial Intelligence in the Firm: Bottlenecks in Software Production" by Fiona Chen and James Stratton via Yahoo Tech (October 11, 2026); Ars Technica (October 9, 2026); Faros AI telemetry via Alibaba Cloud Developer Community (August 14, 2026); LinearB 2026 Software Engineering Benchmarks Report via IT Brief Australia (June 11, 2026); CodeRabbit "State of AI vs Human Code Generation" Report (December 2025); Anthropic Code Review announcement via C114 (March 10, 2026); Alibaba Open Code Review via InfoQ (September 20, 2026); Tencent WorkBuddy CI/CD documentation (August 2026); Kitta code review agent via Alibaba Cloud Developer Community (October 16, 2026); survey of 1,528 developers via genedai.me (July 2026).
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
This analysis is based on published research and reporting available as of the article date; working papers may be revised before publication.
Statistics cited (code volume, review time, merge times) are point-in-time estimates from third-party samples; they may not generalize to every organization.
The Harvard working paper is a working paper, not a peer-reviewed publication; its findings should be read as directional.
Company pricing and product features change frequently; figures reflect the date of the cited source.
Comparative claims about national engineering practices rely on public reporting, not operational data.
Reporting standards differ across jurisdictions; figures may not be directly comparable.
This sector is changing quickly; claims can become outdated soon.
Sources
- Harvard working paper "Artificial Intelligence in the Firm: Bottlenecks in Software Production" by Fiona Chen and James Stratton via Yahoo Tech (October 11, 2026)
- Ars Technica (October 9, 2026)
- Faros AI telemetry via Alibaba Cloud Developer Community (August 14, 2026)
- LinearB 2026 Software Engineering Benchmarks Report via IT Brief Australia (June 11, 2026)
- CodeRabbit "State of AI vs Human Code Generation" Report (December 2025)
- Anthropic Code Review announcement via C114 (March 10, 2026)
- Alibaba Open Code Review via InfoQ (September 20, 2026)
- Tencent WorkBuddy CI/CD documentation (August 2026)
- Kitta code review agent via Alibaba Cloud Developer Community (October 16, 2026)
- survey of 1,528 developers via genedai.me (July 2026).
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: This analysis is based on published research and reporting available as of the article date; working papers may be revised before publication.; Statistics cited (code volume, review time, merge times) are point-in-time estimates from third-party samples; they may not generalize to every organization.; The Harvard working paper is a working paper, not a peer-reviewed publication; its findings should be read as directional.; Company pricing and product features change frequently; figures reflect the date of the cited source.; Comparative claims about national engineering practices rely on public reporting, not operational data.; Reporting standards differ across jurisdictions; figures may not be directly comparable.; This sector is changing quickly; claims can become outdated soon.