Where the human stays in the loop: a decision rule for sales automation
“Human in the loop” is a phrase everyone agrees with and nobody defines. Which step, and which human?
If you can’t answer that for each step, reps either approve whatever the AI produces or ignore it entirely.
Ask three questions about every step
1. What does a mistake cost?
Not “could AI get this wrong.” Everything can get something wrong. The question is what happens when it does.
A wrong industry tag on a lead record costs nothing. Someone fixes it later. A wrong company name in a first-touch email to a target account costs you the account, and probably a screenshot in someone’s Slack.
2. Can the mistake be undone?
Routing a lead to the wrong rep is reversible in minutes. A sent email isn’t. A discount quoted on a call isn’t.
3. Does the step need context that isn’t in your systems?
Some judgment lives in data: firmographics, intent signals, past opportunities. Some lives only in a rep’s head: what the champion said off the record, why the last deal stalled, who actually signs.
Turn the answers into a verdict
| Mistake cost | Reversible? | Context in systems? | Verdict |
|---|---|---|---|
| Low | Yes | Yes | Automate |
| Medium or high | No | Mostly | Assist: AI drafts, a human approves |
| High | No | No | Human: software stays out |
Run your process through this and a pattern shows up fast.
Routing and logging land in automate. First drafts land in assist. Anything involving pricing or a strategic account stays human.
Assist is where most of the value is
The automate column gets the attention because it sounds like headcount savings. The bigger win is usually assist.
An SDR who spends 40 minutes researching an account and writing a first email can review an AI-drafted version in five. That’s not replacing the rep. It’s giving them back most of their day for the part only they can do: having real conversations.
The catch is that assist only works if the review is real. Build the checkpoint into the workflow. If approving a draft is one click with no friction, it’ll get approved without being read.
Move the line on purpose
The line isn’t permanent. As a step proves itself, with low error rates and no surprises, promote it from assist to automate. If something goes wrong, pull it back.
What matters is that every move is a decision someone made and wrote down, not something that drifted because nobody was watching.
The goal isn’t the most automation. It’s the right automation, with a human exactly where a human earns their keep.
If you want a second set of eyes on where your line sits today, book a call and bring one workflow.