The AI Delivery Gap: Why More Code Is Creating More Risk
AI coding tools made code cheaper to write. They did not make it cheaper to prove a change is safe to ship. That difference is the AI delivery gap: teams now produce more changes than they can verify with the same people, tests and process.
pull requests merged per engineer with AI tools
Faros AIreview time in the same data
Faros AIissues per AI co-authored pull request vs human-only
CodeRabbitdelivery instability alongside AI adoption
DORA 2025of developers report high trust in AI output
Stack Overflow 2025The answer is not to slow AI down. It is to give every change the evidence it needs before merge, in proportion to its risk.
What changed
Writing a change used to be the slow, expensive step. Review, testing and release were sized around that pace. AI moved the cost of writing close to zero, but every other step still runs at human speed.
Nearly 90% of technology professionals now use AI at work (DORA 2025). In Faros AI's telemetry, the share of AI suggestions developers accept rose from 20% to 60% (Faros AI, 2026).
The result is a queue. Changes arrive faster than senior engineers can read them, faster than test suites grow, and faster than anyone can confirm each one does what its ticket asked. The gap is not that AI writes bad code. It is that nobody can tell, quickly and with evidence, which changes are safe.
The evidence
Independent sources show the same pattern: output rises, and the signals of risk rise with it.
| Signal | Finding | Source |
|---|---|---|
| Output | 98% more pull requests merged per engineer | Faros AI, 10,000 developers |
| Review load | Average review time up 199.6%, median time in review up 441.5% | Faros AI, Apr 2026, about 22,000 developers |
| Unreviewed merges | Up 31.3% | Faros AI, Apr 2026 |
| Incidents | Incidents per pull request up 242.7% | Faros AI, Apr 2026 |
| Defects at review | 10.83 issues per AI co-authored PR vs 6.45 human-only; security issues up to 2.74 times | CodeRabbit, Dec 2025, 470 open-source PRs |
| Security | 45% of AI coding tasks produced code with a security flaw | Veracode, Jul 2025, 100+ models |
| Production | 43% of AI-generated changes still needed debugging in production after QA | Lightrun via VentureBeat, Apr 2026, 200 SRE and DevOps leaders |
| Stability | Instability keeps rising with AI adoption while throughput improves | DORA 2025 via InfoQ, about 5,000 professionals |
| Guidance | 24% of developers say their employer has published AI tool guidance | Stack Overflow 2026 via IT Brief |
These sources differ in method: telemetry, code analysis and surveys. Several come from vendors with products in this space, so read each figure as directional. What matters is that they point the same way.
Where the gap shows up
The gap is not one failure. It is five places between the pull request and production where verification falls behind.
- Review bandwidth. Senior reviewers are the scarcest resource. A copy change and an authorization change wait in the same queue and often get the same attention.
- Test coverage. AI writes the change quickly but rarely the integration, security or edge-case tests that would prove it. Passing unit tests can hide a missing cross-tenant test.
- Intent. A change can be correct code and still not do what the ticket asked. When the ticket is vague, nobody can check alignment.
- Blast radius. Changes to authorization, data boundaries, payments or shared libraries carry far more risk than their size suggests. Diff size is a poor proxy for impact.
- Ownership and history. The reviewer may not know a touched service has a history of incidents, or who owns it. That context lives in incident logs and people's heads, not in the diff.
Example
An AI agent opens a pull request that updates tenant permission checks. It is small, its unit tests pass, and it looks routine. But it changes who can see which customer's data, it has no cross-tenant test, the ticket never says what denied access should do, and two affected services have prior incidents. Each fact is knowable. Today a reviewer has to find all of them by hand.
What careful teams do today
DORA calls AI an amplifier: it magnifies the strengths of strong teams and the problems of struggling ones (InfoQ). The capabilities DORA links to good AI outcomes include small batches, strong version control and a quality internal platform (TechTarget). Faros AI's 2026 data adds a warning: mature practices alone did not stop downstream quality from slipping.
- Keep AI changes smallOne concern per pull request. Smaller changes are faster to verify.
- Tier review by risk, not sizeRoute authorization, data access, payments and shared libraries to a named owner.
- Require evidence, not green checksFor risky areas, ask for the specific test that proves the behavior.
- Write checkable ticketsState expected behavior, including the failure case.
- Bring history into reviewFlag services with recent incidents so reviewers look harder there.
- Track unreviewed merges and reworkRising churn or incidents per PR are early warnings.
None of these remove risk. They make it visible, so limited review time goes where it matters most.
Brief 02 is out.
AI Pull Request Scoreboard, Edition 1: we followed 361,296 AI agent pull requests to see how they get merged. More than half of the merged ones had no approving review on record.
Read brief 02- Method: public reports and articles, checked on November 4, 2026. EngiSignal has not yet collected its own data. Where a figure came through a news report, the report is named.
- Faros AI, The AI Productivity Paradox
- Faros AI, The AI Engineering Report 2026, ten takeaways, April 12, 2026
- CodeRabbit, State of AI vs Human Code Generation, December 17, 2025
- InfoQ, DORA State of AI-assisted Software Development 2025, September 29, 2025
- TechTarget, Google DORA on software delivery and AI coding tools
- Veracode, 2025 GenAI Code Security Report, July 2025
- VentureBeat on Lightrun, 2026 State of AI-Powered Engineering Report, April 2026
- IT Brief on the Stack Overflow 2026 Developer Survey, October 2026