Human-in-the-Loop: When AI Should Ask Before Acting

The most expensive mistake we see in AI deployments isn't an AI doing the wrong thing — it's an AI doing the wrong thing irreversibly. "Fully autonomous" sounds efficient right up until it empties a customer's account, sends a refund to the wrong person, or emails a whole list the wrong price.

The frame that works: some actions should be fully automated, some should be drafted by AI and approved by a human, and some should never touch AI in the first place. Here's how we decide.

The three tiers

Tier 1: fully automated

Reversible, low-stakes, high-volume actions. Send an acknowledgment SMS, log a lead in the CRM, propose a meeting time, answer a FAQ. If the AI gets it wrong, the cost is a confused customer for ten minutes, and nothing is permanent.

Tier 2: drafted by AI, approved by a human

Actions that are reversible but costly if wrong, or that touch tone and relationship. Refunds below a threshold, personalized pricing quotes, sensitive customer replies, invoice generation. The AI does 95% of the work; a human hits "send" or tweaks.

Tier 3: never autonomous

Irreversible, high-stakes, or legally binding. Large refunds, contract execution, terminating accounts, sending legal notices. The AI can gather context, but it doesn't act.

The decision framework

Four questions, in order:

1. Is it reversible? If no, it's at least tier 2.
2. What's the cost of a single wrong action? If it's high, tier 2 or 3.
3. How often does it happen? High volume + low stakes = tier 1, even with occasional errors. Low volume + high stakes = tier 3.
4. Does a regulator care? If yes, a human signs off. Every time.

Approval UX matters

A human-in-the-loop approval step works when the UX is fast. A queue that requires logging into a dashboard and reading six screens of context gets abandoned. The approval step should be a Slack message with the full draft and two buttons: Send / Edit. If that's the friction, people actually use it.

The goal isn't pure autonomy

It's the highest ratio of autonomous-to-approved actions that your business can actually tolerate. That ratio goes up over time as the system proves itself — start conservative, earn autonomy.

This is how we structure every system we install — see the pattern on our AI automation page.