The first time I watched a developer delete 2,000 lines of code because an AI agent wrote them wrong, I stopped believing the hype. That was August 2024. Here's what I actually think now.
AI in software development is genuinely useful. It's also genuinely overhyped in ways that get expensive fast. The gap between those two facts is where most teams are struggling right now, and honestly, it's a harder gap to close than most guides will tell you.
Bad specs are now your biggest problem. Full stop.
Before agents, a vague spec meant a slow sprint. Now it means an agent writes 500 lines of confident, coherent, completely wrong code. Fast. The bottleneck didn't disappear, it moved. Product clarity, not engineering bandwidth, is the ceiling now. That's a management problem more than a technical one. Weird, right?
We've watched business analysts become unexpectedly critical, the people who can write a user story tight enough that the agent doesn't go sideways halfway through. Nobody predicted that particular outcome two years ago. But here we are.
What actually accelerates, and what doesn't
Test generation, boilerplate, legacy mapping. Those work. Really well. We had a client last year, fintech, 8-year-old codebase, nobody fully understood the billing module anymore. Agents mapped the dependencies in two days. A human team would have taken three weeks just to feel confident enough to touch it.
But architecture? Never architecture. When the same client asked the agent to redesign the module, that went sideways fast. You have to know which jobs to hand over. That judgment doesn't come from the tool.
Code review is where the ROI is hiding
Not enough people talk about this. Automated review agents, not generation, review, cut routine comment cycles by 60–80% in two out of three projects we ran in 2025. PRs that sat for three days got same-day feedback. Not exciting to demo. But it's the thing that actually moved the needle.
💡 From RSC production data: Automated code review, not code generation, was the highest-ROI AI use case across our 2025 EU client projects. Cycle time dropped from days to hours. Review quality went up.
The governance problem nobody warned you about
Here's the failure mode I keep seeing. Speed goes up short-term. Then, three months in, you start finding architectural inconsistencies, undocumented dependencies, security assumptions baked into agent-generated code that nobody reviewed carefully enough. Faster debt accumulation. That's all it was.
The fix isn't slowing down. It's adding structure before you scale: approved patterns for how agents touch the codebase, repository context that's explicit enough for agents to work well, and review standards that account for the fact that agents are confident even when they're wrong. That last one matters more than people expect.
So what should you actually do?
Start with one contained workflow. Not "AI strategy." One thing. Measure it before you expand it. Build the context layer first, repo structure, architecture docs, clear API contracts, because agents work proportionally to how explicit your context is. That discipline is also just good engineering. Two birds.
The honest version: it's working, it's messier than the sales decks suggest, and the teams winning are the ones who redesigned around it. Not the ones who plugged in a tool and waited. If you want to talk through what that looks like for your team, reach out.