Framework
Human in the Loop
A framework for deciding when AI acts and when humans review.
The most common fear about AI automation is “what if it does something wrong?” The answer isn't to avoid automation — it's to design it correctly. This is a simple framework for deciding which actions AI takes automatically and which ones a human should always review first.
The core principle
Automation earns autonomy. Start with human review on anything that matters. As you build confidence that the system does what you expect, you reduce the review step. You never fully remove accountability — you just move it from “approve every output” to “review exceptions.”
The four questions
Before setting any automation to run without human review, answer these four:
What's the dollar value of this action?
Under $500 routine transaction (sending an invoice, booking an appointment).
Any transaction over $500, any refund, any contract or commitment.
What happens if the AI gets it wrong?
Low consequence — a confirmation email sends early, a follow-up goes out twice.
A customer-facing mistake that costs money, damages the relationship, or can't be undone.
How often do exceptions happen?
Edge cases are rare (under 5% of instances) and you've seen them all before.
New process or automation, or anything where you haven't seen enough examples yet.
Does a customer see this output directly?
Internal ops (moving a job stage, logging a note, updating a field).
Customer-facing communication, pricing, or scheduling changes that affect commitments.
The three modes
The AI drafts; a human approves before anything happens.
Example: New automations. Anything touching customer commitments. Any communication over $1K.
The AI acts automatically, but flags anything outside normal parameters for a human.
Example: Established automations with a solid track record. Internal ops. Low-stakes follow-up.
The AI acts and logs. Humans review the log periodically, not every instance.
Example: High-volume, low-stakes, well-tested automations. Confirmation messages. Status updates.
Rule of thumb: Start every new automation in Full Review. Move to Exception Review after 30 days of clean operation. Consider Full Auto only for things that have run correctly 100+ times. Never fully remove the log — you always want to see what happened.
Your next step
Book your free AI audit.
Every system I build comes with this framework baked in — you decide how much autonomy each automation earns, and you can dial it back any time. The free audit is where we map which of your tasks fit which mode.
Book your free audit →