Tech-Improvement Work Keeps Getting Killed. Here's What Survives.
When product managers control requirement creation and measure everything against revenue impact, engineering teams lose the ability to pay down technical debt through official channels. The workaround is to treat tech-improvement as internal hygiene—done in small phases without competing for business scheduling—and to lean on AI for the grunt work that makes the cost too small to argue about.
Across multiple companies, the same pattern holds: architecture upgrades, infrastructure work, and performance fixes get indefinitely postponed or cut outright. The calculus changed around 2022 when traffic growth stopped covering the cost of internal engineering work. Managers now ask for specific data—exactly how many ANRs will this fix, and in which scenarios—and most tech-improvement proposals can't answer.
The few that survive are infrastructure optimizations with immediate, visible results and performance/stability work that produces quantifiable metrics. Code-style nitpicking is a waste of a ticket; it belongs in pre-commit hooks. Architecture overhauls carry the highest risk of being abandoned mid-flight.
AI tooling shifts the equation slightly. Refactoring, test generation, and documentation that once took days can finish in hours, making the labor cost low enough that leadership stops scrutinizing. But AI-generated code also introduces uneven quality and new review overhead, so the net effect is that infrastructure and performance work gain breathing room while broad rewrites remain dangerous.
The core conflict is structural: product managers control the requirement pipeline and optimize for revenue KPIs, so any work without a direct revenue line is filtered out before engineering can even evaluate it.
Treating tech-improvement as internal team hygiene—like cleaning the house—sidesteps the gatekeeping problem entirely, but it requires discipline to carve out time without official tickets.
The shift from qualitative to quantitative justification is a net positive. If a team cannot articulate the concrete impact of a refactor, it probably should not be done.
Architecture migrations that take a year are a bet on organizational stability that most companies cannot make right now; the risk of cancellation mid-way is priced too high.
AI changes the economics of tech-improvement not by making it more strategic, but by making it cheaper—when a refactor costs hours instead of days, the approval bar drops.
The hidden cost of AI-assisted tech-improvement is the review burden: generated code that looks plausible but is subtly wrong creates a new category of technical debt.