Human-In-The-Loop Isn't The Safety

A few weeks ago I posted about blast radius control in agentic AI. Here's a pattern I keep coming back to.

Fixing the underlying authorization stack is the right long-term answer. But it's a significant, multi-year undertaking. Most enterprise systems were never designed for agents. Meanwhile, AI is letting us move faster than ever. The pressure to ship is real.

So one pragmatic pattern I've been thinking about: use human-in-the-loop as a deliberate blast radius containment strategy.

Let the agent do all the work. Research, cross-reference, reason across systems, surface the relevant context. Be completely transparent about how it reached its conclusion. What signals it looked at, what tradeoffs it weighed, what action it's recommending and why.

Then let a human make the call on whether to actually execute.

It's pragmatic. It buys you time to fix the stack properly. And it feels safe.

But here's the problem: we're already watching this pattern fail in a different context.

AI-generated code. The PR looks clean. Well-structured, passes all the checks, reads coherently. The reviewer skims it, hits +1, and moves on. Nobody actually stress-tested the logic. The human was in the loop. Just not meaningfully.

Now take that same dynamic and apply it to an agent recommending updates to a batch of payroll entries. Evaluated performance metrics for an entire site population. A workforce action that affects who gets a PIP, a promotion, or a headcount cut.

The stakes are not a bad merge. They're a paycheck that doesn't clear, a performance record that follows someone's career, an employment decision that can't be easily undone.

The real bar for human-in-the-loop isn't "did a human see it." It's: can the human reliably catch a bad recommendation?

If the answer is no, because there are too many updates, too much context, or the cognitive load is too high, you haven't added a safety layer. You've added a rubber stamp with extra steps.

This feels like one of the most underrated UX problems in enterprise AI right now. Not "how do we make agents capable" but "how do we make human oversight actually meaningful at the speed agents operate."

A few questions I'm sitting with:

• What's the right unit of human review? One action? A plan? A batch?
• How do you surface the critical signal without asking humans to audit an entire reasoning chain?
• At what volume or velocity does meaningful oversight become practically impossible?

Curious if others are hitting this in production and what's actually working.