A few months ago I sat with an engineering leader whose team had gone all-in on AI-assisted coding. PR volume had doubled. Cycle time had not budged. Engineers were shipping faster than ever and, by their own admission, understanding less of what they shipped. “It feels like we are running downhill,” he told me, “but the map is getting blurrier every sprint.”
That conversation stuck with me because the data tells the same story. AI coding assistants are now table stakes. Adoption has peaked. The frontier has shifted from whether to use AI to whether your organization can survive the side effects.
I think the side effects cluster into three debts: technical, cognitive, and intent. The numbers are stark, and most teams are only measuring the first one.
Debt 1: Technical Debt in Review Queues
Let us start with what shows up in dashboards. AI coding tools reliably speed up the inner loop. Across 250,000 developers in 60-plus enterprise organizations, active coding time dropped 48–58% after AI adoption.1 ROI typically pays back in one to three months when usage is actively managed.1
But the work does not disappear. It moves downstream.
LinearB analyzed 8.1 million pull requests across 4,800 teams in 42 countries and found that AI-generated PRs sit in review queues 4.6 times longer than human-written PRs before anyone picks them up.2 Once a reviewer starts, they blast through AI PRs twice as fast, presumably because the patterns are easier to spot.2 The net result is still slower delivery, because queue time dominates.
The quality signals are mixed in ways that should worry leaders. AI-assisted code shows slightly higher test pass rates (93% versus 92%) and lower bug escape rates per thousand lines of code (2.8 versus 3.1).1 But the same data set shows 15–18% more security vulnerabilities in AI-generated code, and code duplication rises from 10.5% to 13.5%.1
The acceptance rates are the clearest signal of trust erosion. In LinearB’s data, AI-generated PRs are accepted at 32.7% versus 84.4% for human code.2 AI PRs contain 1.7 times more issues on average, with critical issues up 40% and logic errors up 75%.3
Reviewers are becoming the safety net for AI output, but review capacity has stayed flat while code volume has grown. Opsera calls this the wait-time anomaly: without automated review, testing, and trust controls, AI shifts work downstream rather than eliminating it.1
Debt 2: Cognitive Debt in People
The second debt lives in heads, not repositories. Augment Code surveyed 219 engineering leaders running AI-forward teams and found that 48% of their shipped code is now AI-generated.4 Yet 55% of those same leaders are concerned or very concerned about losing shared understanding of their own codebase.4
This is what Thoughtworks calls codebase cognitive debt: the growing gap between a system’s implementation and a team’s shared understanding of how and why it works.5 As AI increases change velocity, especially with multiple contributors or coding agent swarms, teams can lose track of design intent and hidden coupling. Weaker system understanding also reduces developers’ ability to guide AI effectively, making it harder to anticipate edge cases and steer agents away from architectural pitfalls.5
O’Reilly’s Radar recently named this comprehension debt: the hidden cost to human intelligence when teams rely on AI-generated code they do not fully understand.6 The risk is not that the code is messy. The risk is that it looks clean, passes tests, and still erodes the team’s mental model.
At the largest organizations in Augment’s survey, 201 to 1,000 engineers, 89% of engineering leaders say their engineers are actively raising fears about skill relevance to their managers.4 The same report notes that many teams described their emotional state with contradictory words: “excited, anxious, invigorated."4
I have felt that tension. AI makes me faster. It also makes me read more code I did not write, produced by a process I did not witness. The old contract was that writing code built understanding through friction. When that friction disappears, something has to replace it, or the understanding disappears too.
Dr. Margaret-Anne Storey at the University of Victoria proposes a triple-debt model: technical debt lives in the code, cognitive debt lives in people, and intent debt lives in artifacts.7 AI may reduce the first while accelerating the other two.7
Debt 3: Intent Debt in Processes
The third debt is the hardest to see because it looks like business as usual. Intent debt is the absence of externalized rationale, goals, and constraints in the specs, role definitions, and onboarding processes that have not been updated to account for AI.7
Augment’s survey exposes this sharply. The number-one hiring priority for these AI-forward leaders is now evaluating AI-generated code. Traditional coding ability has dropped to fifth place.4 Yet only 19 of the 219 surveyed organizations have formally updated their role definitions to match a job that has fundamentally changed.4
We are hiring for a different craft and pretending the job description still fits.
Jellyfish’s 2026 State of Engineering Management Report, based on 636 engineering leaders and platform data from over 1,000 companies, confirms the gap. Eighty-four percent of respondents said engineering productivity is a top management concern, and three in four believe it is a strategic concern for their business.8 Sixty-four percent believe they are achieving at least a 25% increase in developer velocity using AI.8
But the returns scale with commitment. Teams with very high AI adoption report dramatically better outcomes than low adopters: 97% of high adopters say AI boosts productivity versus 60% of low adopters, and 71% report higher job satisfaction versus 45%.8 Only 10% of respondents reported strong enablement combined with high adoption, which means the upside is significant for the other 90%.8
The tool landscape itself is shifting faster than procurement cycles. A year ago GitHub Copilot dominated at 42%. In 2026 it ranks third, displaced by Claude Code and Gemini Code Assist, with twelve more tools close behind.8 No single tool dominates the way Copilot once did. The most advanced teams in Jellyfish data see 1.8× PR throughput, and autonomous agents now generate 21% of PRs for those teams.8
The challenge leaders cite most is not the model or the IDE. It is managing the cost, change, and complexity of putting AI to work at scale.8
Closing the Gaps
I do not think the answer is to slow down AI adoption. The data says the opposite. High adopters outperform low adopters on every measure Jellyfish tracks: productivity, efficiency, job satisfaction, and growth outlook.8
The answer is to treat organizational adaptation as a first-class project, not a side effect.
For technical debt, that means investing in automated review, security scanning, and CI/CD checks that run faster than humans before expanding AI usage further.1 Opsera’s data is explicit: speed that cannot clear review is wasted capacity.
For cognitive debt, teams need explicit countermeasures: walkthroughs where developers explain code they did not write, knowledge transfer during onboarding and offboarding, and retrospectives focused on rebuilding shared mental models.7 Thoughtworks recommends tracking team cognitive load and using architectural fitness functions to enforce key constraints as AI accelerates output.5
For intent debt, the lowest-hanging fruit is honest job definitions. If your hiring bar now centers on evaluating AI output, say so in the job description. If your onboarding assumes engineers learn by writing code from scratch, redesign it for a world where they learn by reading, reviewing, and orchestrating.4
What I Am Doing Differently
At Wawandco and Symbol, we have started three practices that cost almost nothing and help immensely.
First, every AI-generated PR above a trivial size gets a written explanation from the human who prompted it, not just a description of what changed, but why the AI chose the approach it did. This forces the author to understand the output before asking someone else to review it.
Second, we rotate a “comprehension check” into sprint retrospectives. One person picks a recently merged module they did not write and explains it to the team. If they cannot, we treat that as a signal, not a personal failure.
Third, we updated our engineering ladder. The expectation is no longer “writes clean code.” It is “produces correct systems, whether the lines are typed by human or machine, and can explain the tradeoffs to the rest of the team.”
The tools will keep changing. Claude Code may not be number one next year. The debt that matters is not the kind a linter catches. It is the gap between what your tools can do and what your team understands. That is the gap that determines whether AI makes you faster or just makes you busier.
References
[1] Opsera, “AI Coding Impact 2026 Benchmark Report,” February 2026. https://ajoconnell.com/wp-content/uploads/2026/02/opsera-2026-AI-coding-impact-benchmark-report.pdf
[2] LinearB, “2026 Software Engineering Benchmarks Report,” 2026. https://linearb.io/resources/engineering-benchmarks
[3] byteiota, “LinearB 2026: AI PRs Wait 4.6x Longer—Queue Crisis,” 2026. https://byteiota.com/linearb-2026-ai-prs-wait-4-6x-longer-queue-crisis/
[4] Vinay Perneti and Emma Webb, “Excited, anxious, invigorated: what 219 engineering leaders told us about going AI-native,” Augment Code, May 2026. https://www.augmentcode.com/blog/ai-native-survey-2026
[5] Thoughtworks, “Codebase cognitive debt,” Technology Radar, April 2026. https://www.thoughtworks.com/en-us/radar/techniques/codebase-cognitive-debt
[6] O’Reilly Radar, “Comprehension Debt: The Hidden Cost of AI-Generated Code,” 2026. https://www.oreilly.com/radar/comprehension-debt-the-hidden-cost-of-ai-generated-code/
[7] Dr. Margaret-Anne Storey et al., “What kinds of new debt are teams accumulating with AI?” RDEL Newsletter, 2026. https://rdel.substack.com/p/rdel-137-what-kinds-of-new-debt-are
[8] Jellyfish, “2026 State of Engineering Management Report,” 2026. https://jellyfish.co/2026-state-of-engineering-management/