Sometimes you have to grab the reins. AI doesn't know when to stop — and if you're not paying attention, it'll over-engineer a one-character fix into an hour-long deployment.
Last week I needed to add a space to a column header in a critical executive report. One space. Numbers were validated. UAT was approved. Purely cosmetic.
I could've opened VS, edited it, and deployed in five minutes. Done it a thousand times.
Instead, I watched AI spin up a full merge request through CI/CD. Gateway approvals. Code review. Pipeline promotion back through Dev into UAT. An hour-long process for a single character.
I stopped it. "Why are we routing this through the full pipeline?"
Same week, a design change came in: pivot the entire report layout from horizontal to vertical. I knew it was doable — I've done it before. But the specific steps? Fifteen years dusty. One prompt with a screenshot describing the target layout. AI knocked it out in minutes.
Then Friday — two reports approved for production. I triggered my CR workflow. AI immediately started re-running validations. Checking data. Comparing outputs. Prompting me for approvals on tests we'd already passed.
Cancelled. "We already validated everything. Users signed off. Just package and ship."
Three lessons in one week:
→ Sometimes AI's thoroughness is waste. A space doesn't need an hour-long pipeline. Catch it early or just do it yourself.
→ Sometimes it's exactly what you need. When details have faded but judgment hasn't, AI fills the gap perfectly.
→ Sometimes AI doesn't know the work is done. It will re-validate indefinitely unless you say stop.
I'm calling this "AI Creep" — the tendency to over-explore, over-validate, and over-engineer unless someone with domain knowledge keeps it focused.
The irony? The more you know about the task, the more you feel the creep. A junior might not notice unnecessary re-validation. You notice because you know what's already been done.
Here's what makes AI powerful: it handles the technical details you'd normally spend 20 minutes Googling to refresh. Syntax you haven't touched in years. Steps you've done before but don't keep loaded in memory. It fills those gaps instantly — which means experience compounds instead of fading. That's true whether you have 30 years or 3.
But that same AI will burn an hour on a space and re-test what's already approved if you let it.
The expertise isn't in remembering every technical detail anymore. It's in knowing when to let AI run — and when to pull it back.