Cognitive overload: the hidden bottleneck of AI-powered software delivery
By Jérôme Leroy, Agentic Delivery Coach, SQLI
AI hasn't eliminated the hard part of software delivery, it's relocated it. As code generation accelerates, DORA's 2025 research confirms the real strain now falls on the humans reviewing it, with senior engineers absorbing the load and juniors losing the chance to learn by doing. This piece argues the next competitive edge won't come from writing more code, but from protecting the judgment needed to catch what AI gets wrong.
AI has transformed software development. The DORA 2025 State of AI-assisted Software Development report found that 90% of technology professionals now use AI at work, and over 80% believe it has increased their productivity.
Yet the same research reveals a paradox: DORA's 2025 findings show that higher AI adoption is associated with an increase in both software delivery throughput and software delivery instability. Speed and instability are rising together.
The bottleneck has moved. It's no longer code production, but human attention.
In AI-assisted delivery, success depends less on generating software than on people's ability to understand, validate and govern it.
Writing less, judging more
As AI takes over code generation, developers spend less time writing software and more time reviewing it.
Paradoxically, reviewing AI-generated code often requires more concentration than writing it. That responsibility falls largely on senior engineers. As one engineer told DORA's researchers in 2025: "Reviewing [another's] code is so much harder than writing it. AI tools are increasing the rate at which people can churn out code that needs to be reviewed…"
While dashboards celebrate velocity, experienced developers quietly become overloaded. Reviews become superficial. Code is approved because the tests pass, because it looks correct, or simply because delivery must continue.
The principle of "no merge without understanding" gradually disappears.
An invisible risk
Cognitive overload rarely feels like overload.
Each successful prompt delivers instant progress and encourages another. Developers often describe highly productive sessions that leave them mentally exhausted only afterwards.
The brain mistakes speed for effectiveness, and by the time fatigue becomes visible, review quality has already declined.
This makes cognitive overload more than a wellbeing issue. It's both a delivery and a security risk.
Why juniors are the most vulnerable
The long-term risk is even greater for junior developers.
Software engineers traditionally learned by writing code, making mistakes and receiving demanding reviews. AI removes much of that learning process.
Instead, juniors are increasingly asked to validate code they lack the experience to judge. Faced with confident AI output, automation bias becomes difficult to resist.
DORA's 2025 research calls this the "expertise paradox": lowering the barrier to entry with AI risks bypassing the "productive struggle" needed for deep technical expertise. As the report puts it, current usage patterns "deliver breakthrough productivity while simultaneously blocking skill development, effectively limiting the hands-on problem-solving essential for true apprenticeship."
Without careful supervision, organisations risk producing reviewers who never learned to build.
Judgment, not production, becomes the skill that must be deliberately taught.
Designing delivery around attention
Protecting attention isn't about slowing innovation, but about designing delivery differently.
Three principles matter most:
1/ Humans remain accountable for every production change.
2/ Agent sessions should be deliberately limited rather than multiplied indefinitely.
3/ Automation should eliminate repetitive verification so that people can focus on architecture, risk and engineering judgment.
DORA's own 2025 recommendations point the same way: shifting AI-generated feedback to the author before review, and forcing large AI-generated changes into small, reviewable batches, rather than letting output volume grow unchecked.
Measure attention, not just velocity
Traditional delivery metrics capture throughput; they rarely capture cognitive load.
Early warning signs are easy to spot:
- reviews completed unusually quickly
- approvals without comments
- review responsibility concentrated on a few senior engineers
- juniors who never escalate concerns
- experienced engineers showing signs of fatigue or leaving
These signals reveal far more about the health of an AI-enabled delivery system than velocity alone.
The next competitive advantage
The first wave of AI adoption optimised code generation and the next will optimise human judgment.
In agentic delivery, code is no longer the scarce resource.
Attention is.
Organisations that protect it will not simply deliver software faster, they will deliver it more safely, sustainably and with expertise that continues to compound.