An analysis of 1M+ pull requests across 2,444 engineering organizations. More AI spend. More code volume. More production failures. We measured where AI engineering effort actually goes, how code review has responded to 2.6× volume growth, and why the reactive work treadmill keeps accelerating. The findings: $0.82 of every AI dollar is consumed before a single feature reaches users.
Entelligence Research · May 2026
Token Maxxing Is Making Engineering Teams Slower
Scroll
PRs analyzed
1M+
Across 2,444 engineering organizations
Reactive work, median org
44%
Bugs + maintenance, nearly half of all capacity
Reverts vs PR growth
3.7×vs2.6×
Failures compounding faster than output
01
$0.18reaches users
For every dollar a team spends on AI coding tools, just 18 cents becomes shipped product. The other 82 cents is consumed by the maintenance cycle those same tools accelerate, not because engineers are slow, but because there is no closed loop between production reality and the code being written.
Proportional allocation of AI engineering spend, platform average. The green band is the only part that reaches users.
Reactive · P75 / P90$0.62 / $0.76
Shipped · P75 / P90$0.10 / $0.06
02
Half of all engineering output is reactive.
At the median organization, 44% of every PR is reactive: fixing existing code or keeping systems running. The distribution has a long tail: at the 90th percentile, more than three-quarters of all engineering effort produces no net-new product.
Share of engineering output classified as reactive (bugs + maintenance), by organization percentile. The shaded band is the interquartile range. More AI spend accelerates volume on both sides, features and maintenance alike, so the tail only thickens.
03
1 in 4
lines written each week is thrown away before the week closes. Not planned refactoring, but code that did not survive the sprint it was written in. The AI generates from local context, never from production reality: which patterns failed, which edge cases were already tried and reverted.
Weekly code churn by percentile
Industry 27%
Median
25%
P75
38%
P90
55%
Lines written and discarded within the same week. The median sits right at the Pluralsight/GitPrime industry benchmark; the P90 runs at twice it.
04
Reverts are outpacing output.
Over twelve weeks, weekly PR volume grew 2.6×, but reverted PRs grew 3.7×. The failure rate is climbing faster than the work itself. Each revert spawns a bug-fix PR, which feeds the reactive total. The 44% becomes 50%, then 56%.
Merged PRs · indexed (100 = 2,525/wk)
Reverted PRs · indexed (100 = 10/wk)
Feb 16 - May 4, 2026, platform-wide. No pre-AI baseline exists. The argument rests on the rate difference: if quality scaled with velocity, revert growth would track PR growth. It ran 40% faster.
The compounding loop · larger each turn
More AI spendMore volumeReview can't scaleShips unreviewedReactive work grows↻ back to the start
Seven findings from billions of AI tokens.
Across 1M+ pull requests · 2,444 organizations · May 2026
01
44%
Median org reactive-work share: bugs + maintenance consuming nearly half of all engineering capacity
02
25%
Weekly code churn: one in four lines written is discarded before the sprint closes
03
21.6%
Comments addressed: 4 in 5 review comments are never acted on, across 225,000+ records
04
48.5%
PRs rubber-stamped in under 60 minutes: 10,588 with zero reviewer comments
05
3.7×
Reverted PRs grew 3.7× while PR volume grew 2.6×: failures compounding faster than output
06
11.6d
Average bug lifetime in production before being caught and reverted
07
11,033
High-risk flagged PRs approved and merged anyway: 52.7% of all automated risk flags were ignored
05
Half approved in under an hour. Four in five comments never acted on.
Review hasn't scaled with AI output. 48.5% of PRs are approved in under 60 minutes, faster than any meaningful review. And bug & error comments, the highest-value category, are addressed just 26% of the time.
Avg comments / PR20.8
Bot-generated80%
Comments addressed21.6%
Comment types: share of 225k comments, and how often each is acted on
ShareActed on
Bug / error
32.0%26%
Other
20.2%18%
Security
10.9%21%
Testing
8.5%21%
Performance
8.3%21%
Code suggestion
7.3%16%
Style / nit
5.5%20%
Documentation
4.3%21%
Refactor / design
3.0%21%
Review turnaround time · 1M+ PRs
Under 1 hr
48.5%
1-4 hours
15.4%
4-24 hours
17.2%
Over 24 hrs
18.9%
06
759 of every 1,000 issues are critical or high.
Across organizations with production error tracking connected, three in four issues are serious enough to cause direct user impact. Critical issues fire 3.3 times on average before anyone catches them: failures that already reached users, repeatedly, before being logged.
132
627
189
51
Criticalfires 3.3×
service-breaking, data loss, security
Highfires 1.3×
significant functional failure
Mediumfires 0.6×
non-blocking, limited blast radius
Lowfires 0.2×
cosmetic, no direct user harm
Per 1,000 production issues · 1,141 issues · 1,543 PR match events. “Fires” = average times an issue recurs before it is caught.
Top recurring error classes: flagged in review, what happened next
Fixed
Merged anyway
missing null check
73
unvalidated input to external api
41
unvalidated external input
46
unhandled timeout
39
timeout no retry breaker
28
unhandled activity failure
20
unhandled async operation
13
validation after mutation
12
1,543 match events · 13 organizations. Merged anyway = flagged by Entelligence, then approved and merged without being fixed.
07
Close the loop. Every fix compounds.
Today's tools see half the picture: reviewers see PRs but never production; SRE agents see outages but never the PR that caused them. Entelligence sees both, building a living memory of your org's failures so every review gets smarter than the last.
Review Agentreviews every PR against real incident history, citing the past failure your diff resembles
Instrument Agentadds the right logs, traces and alerts at merge time, so a break explains itself
RCA Agentroot-causes production failures and drafts the fix PR
Monitor Agentwatches logs after deploy and confirms the fix held
