AI News
21 Mar 2026
Read 8 min
AI-assisted coding productivity study proves output doubles
AI-assisted coding productivity study shows teams ship twice the code with minimal quality loss now.
A new AI-assisted coding productivity study across 700+ companies finds output per engineer nearly doubles while code quality holds steady. With 63% median AI adoption and most firms now shipping AI-generated code, top teams merge 2.2 pull requests a week versus 1.12 for low adopters, as autonomous agents start to rise.
Software teams are shipping faster. A fresh benchmark from Jellyfish, which examined 200,000 engineers and 20 million pull requests, connects broad AI tool adoption with a sharp jump in throughput. The data also shows only a tiny uptick in code reversions, suggesting speed gains are not crushing quality.
What the AI-assisted coding productivity study shows
Adoption is now mainstream
- 63% median adoption of AI coding tools across companies studied
- 64% of companies generate most of their code with AI help
- Weekly usage is rising, with more engineers using AI several days each week
Throughput nearly doubles
- High-adoption teams (75%–100% of engineers using AI 3+ days/week) merge 2.2 PRs per engineer weekly
- Low-adoption teams average 1.12 PRs per engineer weekly
- Result: almost 2x output tied to consistent AI use
Why output is surging
AI accelerates routine work
- Drafting functions, tests, and docs is faster with suggestions and refactors
- Context windows help AI read large codebases and propose targeted edits
- Engineers spend less time on boilerplate and more time on product logic
The 2025 inflection
- Late 2025 model upgrades made assistants far more capable
- Developers embraced tools like Anthropic’s Claude Code, GitHub Copilot, Cursor, and OpenAI-derived models
- Holiday experimentation led to durable, daily workflows
Code quality is holding steady
Reverts barely change
- Revert rates tick from 0.61% at low-adoption firms to 0.65% at top adopters
- That small rise suggests review, testing, and CI/CD are absorbing faster change
- The real bottleneck is now validating AI output at scale, not writing it
What to watch next
- As output grows, review queues can swell and slow down merges
- Test coverage and runtime checks must keep pace with change volume
- Security and license scanning should run earlier in the pipeline
From copilots to agents: the new gap
Autonomous PRs are small but growing
- AI agents are beginning to open and commit PRs on their own
- High-adoption orgs test agents on safe, low-risk tasks first
- These firms are pulling ahead, widening the performance gap
Where agents make sense now
- Dependency bumps, config updates, and small refactor passes
- Fixes suggested by static analysis or security scanners
- Docs sync, comment fixes, and repetitive code style changes
How to capture the gains without breaking things
Guardrails for copilots and agents
- Set PR size limits and require automated tests for every change
- Mandate human review for risky files and critical services
- Enable pre-commit hooks, linters, and type checks by default
- Track revert rate, lead time, and escaped defects weekly
- Restrict agent permissions with fine-grained tokens and scopes
- Log every AI-sourced change; label PRs for easy auditing
Team practices that scale
- Write clear tickets with acceptance tests and edge cases
- Invest in fast feedback: sub-10-minute CI pipelines
- Keep architecture docs current so AI can read the “map”
- Rotate reviewers to spread context and reduce bottlenecks
What this means for leaders and developers
Leaders
- Use the AI-assisted coding productivity study benchmarks to set targets for PR throughput and quality
- Budget for platform work: testing, observability, and security shift-left
- Measure ROI: compare cycle time and incident rates before and after adoption
Developers
- Pair with AI for drafts, then focus your time on hard design choices
- Keep a high bar for tests and readability; speed means little without safety
- Learn prompt patterns, codebase context tricks, and review tactics for AI-sourced code
(Source: https://www.businessinsider.com/ai-coding-boom-more-software-shipped-no-hit-quality-2026-3)
For more news: Click Here
FAQ
Q: What did the AI-assisted coding productivity study find about software output and quality?
A: The AI-assisted coding productivity study across more than 700 companies found that output per engineer nearly doubled while code quality held steady. Median AI adoption was 63% and top teams merged about 2.2 pull requests per engineer per week versus 1.12 at low-adoption firms.
Q: How common is AI tool adoption in the companies analyzed?
A: The study reports a 63% median AI tool adoption and that 64% of companies now generate a majority of their code with AI assistance. Weekly usage has been rising, with more engineers using AI tools multiple days per week.
Q: How much faster are high-adoption teams compared with low-adoption teams?
A: High-adoption teams (75%–100% of engineers using AI three or more days per week) merged an average of 2.2 pull requests per engineer weekly, nearly double the 1.12 at low-adoption teams. The study ties much of the gain to consistent AI use for routine tasks like drafting functions, tests, and refactors.
Q: Did faster output lead to a large drop in code quality?
A: Revert rates rose only modestly from 0.61% at low-adoption firms to 0.65% at top adopters, indicating a small uptick. The study suggests review, testing, and CI/CD processes are absorbing much of the faster change and that quality is generally holding up.
Q: What role are autonomous agents currently playing in coding work?
A: Autonomous agent activity—pull requests opened or committed by AI agents—remains a small share of overall work but is climbing rapidly, particularly among top adopters. High-adoption organizations are testing agents on safe, low-risk tasks such as dependency bumps, config updates, and small refactor passes.
Q: What guardrails does the AI-assisted coding productivity study recommend to maintain quality?
A: The AI-assisted coding productivity study recommends guardrails like PR size limits, automated tests for every change, mandatory human review for risky files, pre-commit hooks, linters, type checks, and labeling or logging AI-sourced changes for auditing. It also advises restricting agent permissions and tracking metrics such as revert rate, lead time, and escaped defects while running security and license scans earlier in the pipeline.
Q: What should leaders measure and invest in after adopting AI coding tools?
A: Leaders should use the AI-assisted coding productivity study benchmarks to set targets for PR throughput and quality and budget for platform work such as testing, observability, and shifting security left. They should also measure ROI by comparing cycle time and incident rates before and after adoption.
Q: How can developers get the most benefit from AI tools without sacrificing safety?
A: Developers should pair with AI for drafts and focus their time on hard design choices while maintaining a high bar for tests and readability. The study also recommends learning prompt patterns, keeping architecture docs current, investing in fast feedback like sub-10-minute CI pipelines, and rotating reviewers to spread context.
Contents
