← Back to blog

51% Say AI Feels Like a Second Job for Engineers

AI promised to free engineers. Instead, 51% say it's become a second job. Discover why AI tools backfired for many teams.

51% Say AI Feels Like a Second Job for Engineers

The Promise vs. The Bill

The pitch was clean: AI assistants would write the boilerplate, generate the tests, scaffold the migrations, and free engineers to think about the interesting problems. Two years in, the receipts tell a different story. The boilerplate got faster. The interesting problems got significantly harder.

In our work helping product teams ship software, the pattern we keep seeing is consistent. Velocity on the easy 60% of a project — CRUD endpoints, form validation, basic UI states — went up sharply. Velocity on the hard 40% — architecture decisions, debugging legacy systems, reviewing code nobody fully owns — went down or stayed flat. The net result for many teams is faster output and more exhausted engineers.

Put simply: AI did not flatten the difficulty curve. It steepened it.

What the Research Actually Says

The data backs up what teams have been feeling. According to Boston Consulting Group's study covered by CBS News, researchers surveyed about 1,500 workers and found that people constantly bouncing between multiple AI tools reported more decision fatigue and more errors. About one in seven workers said they had experienced mental fatigue from juggling AI tools at work — a phenomenon workers themselves are now calling "AI brain fry."

Julie Bedard, the BCG managing director who co-authored the study, framed the issue plainly: "The AI can run out far ahead of us, but we're still here with the same brain we had yesterday." She added that intensive oversight of AI "causes a lot of sort of cognitive, just exhaustion."

The identity dimension is equally loud. In a 2025 global LinkedIn/Censuswide survey of more than 19,000 professionals, 51% said learning AI feels like another job, and 41% said the current pace of AI change is taking a toll on their wellbeing. The Martian Chronicles analysis from Evil Martians puts it bluntly: "AI-assisted coding grows the workload exponentially — increasing both the amount of tasks and the intensity of work."

Real numbers: when a non-AI engineer might complete one steady chunk of work in a four-hour block, the AI-assisted counterpart can ship roughly twice the output — at roughly twice the cognitive intensity. The math is impressive on the dashboard. It is brutal on the calendar.

Why the Hard Parts Got Harder

The "hard parts" of software engineering were never typing speed. They were three things: holding the right mental model, making correct trade-offs under uncertainty, and verifying that what shipped actually does what it should. AI changed how each one feels.

The Mental Model Now Belongs to a Machine

When an engineer writes a feature from scratch, they build the mental model by typing it. By the time the pull request is open, they understand every branch. When an AI writes the same feature in 90 seconds, the code exists — but the model in the engineer's head doesn't. They now have to reverse-engineer their own pull request. That is a fundamentally harder cognitive task than authoring code: the same way reading someone else's essay and grading it is harder than writing your own.

This is why teams report what one developer told CBS News: the temptation to keep iterating because "the next best thing is possible" leads to spending more time writing the perfect prompt than the perfect code. The judgment work moved up the stack, and the stack got noisier.

Context Switching Multiplied

Software engineering was already losing to context switching before AI. Research from the University of California, cited by MindspaceX, found it takes an average of 23 minutes and 15 seconds to fully regain focus after an interruption. Reclaim's analysis adds that every context switch costs roughly 20% of cognitive capacity.

Now add three or four parallel AI agents that each need supervision, prompt refinement, and output review. The engineer is no longer working on one feature with occasional interruptions — they are managing a small team of unreliable junior contributors who never get tired and never stop generating work. The traditional "deep work block" becomes nearly impossible to defend.

In our experience with engineering teams, this is where the velocity story collapses. Output graphs go up. Bug rates also go up. The team ships more code and trusts less of it.

Identity, Not Just Workload

Honest take: the burnout AI is creating is not the burnout from too many tickets. It is the quieter, more corrosive kind described in the LinkedIn analysis by Henrik Werdelin — engineers comparing themselves against machines that do not get tired, do not feel, and do not pause. The World Economic Forum has a term for this: FOBO, the Fear of Becoming Obsolete. It shows up as anxiety, irritability, and a constant sense of "I'm not doing enough" — even on days when output is up significantly.

For technical leaders, this matters because high performers freeze first. The engineers who derived identity from being the person who solved the impossible bug are the ones most disoriented by an AI that solves a passable version of it before lunch.

The Productivity Paradox in One Sentence

The BCG study captured the paradox cleanly: AI can both reduce burnout and create it. When workers use AI to offload repetitive tasks, stress drops. When workers have to supervise multiple AI systems or juggle several tools at once, mental strain rises sharply.

The dividing line is not "how much AI" but "what AI is replacing." Replace the boring part of the job — boilerplate, repetitive refactors, test scaffolding — and engineers feel relief. Replace the part that was the cognitive rest between hard problems, and you remove the recovery time without removing the hard problems.

This is the trap that hits ambitious teams hardest. Visier's analysis of AI exhaustion borrows a phrase from healthcare: "practicing at the top of their license." When AI handles every easy task, engineers are constantly operating at peak cognitive demand. There is no longer a slow Friday afternoon spent renaming variables. Every hour is a hard hour.

Alternatives and Trade-offs

There is no version of the next five years where engineering teams roll back AI tooling. The honest question is what shape the workflow takes. We see three patterns in practice, each with trade-offs.

Pattern 1: Maximum Velocity

The team adopts every available AI tool, runs parallel agents, and optimizes for output. Throughput goes up significantly. Code review queues balloon. Senior engineers become full-time reviewers of AI output and stop writing code themselves. Bug rates often climb because reviewers cannot hold mental models for code they did not author. Burnout shows up around month six.

This works for short bursts — MVPs, prototypes, hackathon-style projects. It does not scale to long-lived products.

Pattern 2: AI as a Strict Assistant

The team uses AI only for clearly bounded tasks: generating tests against an existing specification, writing one-off scripts, drafting documentation, translating between languages or frameworks. Engineers keep authorship of architecture and core logic. Velocity gains are smaller — perhaps 15-30% on the easy parts — but cognitive load stays manageable and code quality holds.

This is the most sustainable pattern we see for teams maintaining production systems with real customers.

Pattern 3: Async-First with Closure Practices

Borrowing from how GitLab runs its 1,300-person remote team, this pattern combines AI tooling with strict asynchronous communication, time-blocked deep work, and what MindspaceX calls "closure practices" — taking 2-3 minutes before switching tasks to document current status and next steps. The team accepts slightly lower raw velocity in exchange for protecting the focus blocks that hard engineering problems require.

What this means for your project: if you are building anything that needs to run for more than 18 months, Pattern 2 or 3 will cost less in total than Pattern 1. The savings show up in bug rates, employee retention, and the absence of an architecture rewrite in year two.

Here Is What We Recommend

Three concrete moves for engineering leaders trying to keep velocity gains without paying the burnout tax.

Separate generation from judgment. Let AI generate aggressively in bounded domains — tests, fixtures, scaffolding, documentation. Keep judgment work — architecture, security reviews, data model changes — in human hands without AI as a first draft. This protects the mental models that matter most.

Budget for review time, not just generation time. A reasonable rule we use: if AI produces a feature in 30 minutes, budget at least 30 minutes for an engineer to read, understand, and validate it before merge. Teams that skip this step ship faster for two months and then spend the next six debugging code nobody owns.

Defend at least one deep work block per day. With context switching costing 23 minutes per interruption, an engineer with five interruptions has already lost two hours of effective focus before counting actual work. Make one four-hour block per day non-negotiable. No AI agents to supervise, no Slack, no review queue. This is where the hard 40% of the work actually happens.

Key Takeaway for Business

Three conclusions for technical leaders making decisions in the next quarter.

First, AI productivity gains are real, but they are concentrated in the parts of engineering that were already cheap. The expensive parts — judgment, debugging, system design — got more expensive, not less. Plan budgets accordingly.

Second, "AI brain fry" is not a soft concern. With one in seven workers reporting cognitive fatigue and 51% calling AI a second job, the retention risk for senior engineers is material. The cost of replacing a senior engineer who burns out is significantly higher than the cost of a slightly slower release cycle.

Third, the teams that win the next two years will not be the ones that adopt the most AI. They will be the ones that adopt AI most deliberately — choosing what to automate, what to protect, and where to insert humans back into the loop on purpose.

Put simply: AI made typing cheap and thinking expensive. Price your projects around the thinking.

Frequently Asked Questions

How do developers debug and modify code they didn't write and don't fully understand?

The honest answer is: slowly, and with more bugs than code they authored themselves. The standard practice is to treat AI-generated code as if it came from an unknown contractor — read it line by line, write tests against it before modifying, and rely on observability to catch behavior the reviewer missed. The 30-minute review rule mentioned above exists for this reason.

How can teams prevent "vibe coding" where developers iterate on prompts until output seems right without verifying logic?

Two practices help. First, require a written specification or test case before any AI generation — if the engineer cannot describe what correct output looks like, they cannot validate what comes back. Second, enforce code review by someone other than the prompter. Vibe coding survives self-review because the prompter already believes the output. A second pair of eyes breaks the spell.

Does AI-generated code require more cleanup time than manual code, despite being faster to produce?

In many cases, yes. AI tools optimize for plausible output, not correct output. Code that compiles and passes obvious tests can still contain subtle bugs, security issues, or architectural mismatches that surface weeks later. Teams should measure end-to-end time (generation + review + bug-fix tail) rather than just generation speed when evaluating tooling.

This article is based on publicly available sources and may contain inaccuracies.

Related articles

✦SqueezeAI
  1. AI didn't flatten the difficulty curve for engineers — it steepened it. Velocity on routine work (CRUD, forms, scaffolding) rose sharply, but velocity on the hard 40% (architecture, legacy debugging, code review) stayed flat or declined, leaving teams with faster output and more exhausted engineers.
  2. Overseeing AI output is itself cognitively expensive: BCG research found that intensive AI oversight causes significant mental fatigue ('AI brain fry'), because the human brain hasn't changed even as AI runs far ahead of it.
  3. AI-assisted engineers produce roughly twice the output in the same time — but at roughly twice the cognitive intensity, meaning the productivity gains visible on dashboards come at a direct cost to engineer wellbeing.
  4. When AI writes the code, engineers lose the mental model they would have built by writing it themselves — they inherit responsibility for code they don't fully understand, making verification and ownership harder than before.

Powered by B1KEY