Without workflow redesign, teams get a 10-20% boost instead of the dramatic gains they expected. Developers use AI as a suggestion engine. Money spent, minimal return.

AI code without tests is AI liability. Tests aren't just for humans. They're the agent's feedback loop. Without them, the agent can't verify its own work and you can't verify it either.

Unsupervised juniors with agentic tools ship plausible-looking code they can't actually evaluate. The fix is structure, not gatekeeping. Start juniors on AI-ASSIST tasks with mandatory senior review. Have them write prompts and approve plans before the agent codes. They'll build judgment fast, but they need guardrails while they do. I've seen this go both ways: juniors left alone produce a mess, juniors paired with a senior who coaches them through the review process level up remarkably quickly.

AI generates subtle vulnerabilities that pass tests. Without automated scanning from day one, you accumulate security debt at the same rate you're shipping features, which is much faster than before.

AI works best with existing codebases that have patterns, tests, and conventions. Starting from zero gives the agent nothing to learn from. You lose the institutional knowledge embedded in your code. When I port a system, I put the original source code in a reference folder and set up automated checks so the agent verifies its implementation against the original for consistency. The old code IS the specification.

AI makes LOC meaningless. Teams tracking LOC reward verbose AI output. Measure features shipped, bug rates, and cycle time instead.