Strip away the vendor keynotes and the doom threads, and the demonstrated, bankable wins for AI in software engineering are strikingly consistent: they're narrow, high-volume, and unglamorous. Framework upgrades. Language-version migrations. Boilerplate services. Test scaffolds. Dependency patches. The work every backlog carries for years because it's important, boring, and never urgent — until the audit or the CVE makes it urgent all at once.
That's not a disappointing consolation prize. It's the best news an engineering organization has had in a decade. Migration debt is the tax every platform pays quarterly; it's measured in developer-years, and it's exactly the shape of problem — repetitive, verifiable, pattern-heavy — that current AI tooling handles brilliantly. I run coding agents against my own infrastructure for precisely this class of work: module upgrades, manifest migrations, config sweeps across repos. The judgment stays mine. The toil doesn't.
Make it a platform capability, not a personal experiment
Most organizations are adopting coding AI as a thousand individual experiments: every developer with a different assistant, different settings, no shared guardrails, and no way to answer "what did AI write and who reviewed it?" That's the same mistake as letting every team hand-roll its own CI in 2015, and it has the same fix: make it a paved road. Sanctioned tools with governed model access. AI-written changes flowing through the same review, testing, and provenance gates as human ones — the pipeline doesn't care who typed it, and that's the point. And productize the big wins: a migration agent the platform team runs as a service against a hundred repos beats a hundred developers each negotiating with an assistant.
The guardrail question matters more with volume. Code review was designed for human error rates and human patch sizes; an agent that can produce fifty plausible pull requests a day will find every weakness in a rubber-stamp culture. The teams that win treat verification as the product: test coverage, behavioral checks, canary deploys — the paved road doing the skepticism so humans can spend theirs where it counts.
What's left is the good part
The uncomfortable reframe for engineers is also the liberating one: if the rote work drains out, what remains is the part that was always the actual job — understanding the problem, choosing the architecture, deciding what "correct" means, and owning the consequences. The developers who struggle in this shift are the ones whose value was fluency in toil. The ones who thrive were always doing judgment work, and just lost their least favorite chores. Engineering leaders should plan for that explicitly: measure the migration debt AI can burn down this year, build the guardrails to let it, and be honest with teams about which skills are appreciating and which just got automated.