AI Coding Gains Real But Overstated, Warns Eng Director
28 Jul 2026
A director of engineering with more than three years in the role has spent the past year compiling notes on what actually changes when large language models enter an engineering org's daily workflow — and what doesn't. The notes, edited with help from Gemini 4, land at a moment when newer AI models launched in Q4 2025 have eclipsed the capabilities of prior generations. The author's conclusion: the cost of producing plausible-looking code has collapsed permanently, but several popular claims built on top of that fact don't hold up.
What's real
According to the report, the author is unambiguous that something fundamental has shifted. Producing code that looks plausible is now dramatically cheaper than before, and that change appears permanent. AI gains are clearest in greenfield work, boilerplate, and unfamiliar territory — areas where engineers are building something new or working outside their expertise.
What's overstated
But the author pushes back hard on three widely repeated claims:
- That AI tooling has made engineering orgs "dramatically faster" overall — this remains unproven.
- That code review, documentation, and onboarding are now obsolete — the author calls this simply wrong.
- That teams can run the same roadmap with half the headcount — the author frames this as a bet, not a fact.
Notably, the report indicates that AI's gains fade or even invert once engineers are doing deep work on systems they already understand well — precisely the kind of work that experienced teams spend much of their time on.
The pipeline problem
The most pointed warning in the report concerns junior engineers. The author states plainly that there is no known method for training engineers in the current environment, and that this is an unsolved problem industry-wide. If AI absorbs the "practice work" that junior engineers traditionally use to develop into senior engineers, the pipeline that produces senior talent breaks — but the effects wouldn't be visible for an estimated three to five years, since that's roughly how long it takes today's juniors to become tomorrow's seniors.
Risks flagged in the report
- Cutting headcount based on the unproven assumption that AI lets teams run the same roadmap with fewer people.
- Treating code review, documentation, and onboarding as obsolete, which could undermine engineering quality and knowledge transfer.
- The junior pipeline problem may not surface for years, creating a delayed but significant talent gap.
- Overestimating AI's impact on deep, familiar systems work could lead to misallocated engineering resources.
Opportunities
The report also points to upside: AI tooling may genuinely help with greenfield work, boilerplate, and unfamiliar territory. Organizations that get ahead of junior-engineer training in an AI-assisted environment could build a long-term talent advantage. And rather than scrapping code review, documentation, and onboarding, reassessing how they're done could help teams adapt without losing rigor.
What's missing
The report notes several gaps worth flagging. No specific data, metrics, or studies are cited to support or refute the productivity claims. The size, industry, or type of engineering org isn't specified, and there's no detail on what concrete management practices actually changed. It's also unclear how "unproven" claims were evaluated, or what a solution to the junior pipeline problem would even look like. Aside from Gemini 4 being used for editing, the exact AI models driving the Q4 2025 capability shift aren't named.
Why founders should care
For early-stage founders, the implications are largely about calibration rather than panic. It's plausible that AI tooling meaningfully speeds up greenfield and boilerplate work, but it's far less certain — and arguably unlikely, per this report — that the same team can sustain its full roadmap at half the headcount. Founders who treat AI productivity claims as hypotheses to test, rather than settled facts, are more likely to avoid over-cutting engineering capacity prematurely.
The junior pipeline risk deserves particular attention precisely because it's slow-moving: a startup that stops investing in junior engineer development today may not feel the consequences for three to five years, by which point the talent gap could be significant and hard to reverse. Founders relying heavily on AI for deep, familiar codebases should also watch closely for whether expected gains actually materialize — the report suggests they may fade or invert in exactly this context. Finally, even as AI reshapes how code gets written, documentation and onboarding processes still appear to matter — startups that abandon them entirely may be trading short-term speed for longer-term knowledge and quality risks.