Guide · Approve-first
Human-in-the-loop automation
Human-in-the-loop automation is a pattern where AI prepares the work and a person approves it before it takes effect. The machine drafts, researches and assembles; a named human makes the decision that reaches a customer, spends money, or publishes in public. The question is never whether to have a human involved — it is exactly where the work stops for one.
In the loop, on the loop, out of the loop
Three positions, and most teams use all three at once on different work. Naming which one applies where is most of the job.
| Position | What it means | Where it fits | What it costs |
|---|---|---|---|
| In the loop | The work stops. A person reviews and approves before the action happens. | Anything that reaches a customer, spends money, or is published in public | One review step; the work waits for a person |
| On the loop | The work proceeds. A person watches, can intervene, and reviews the record afterwards. | High-volume, low-consequence steps a team has already watched for a while | Mistakes are caught after the fact, not before |
| Out of the loop | The work runs unattended. Nobody sees it unless something breaks loudly. | Internal, reversible, well-bounded tasks | You find out from a customer, a platform, or an invoice |
Why teams want a gate before send, spend and publish
Not because the model is bad at writing. Because these are the five failures that are cheap to prevent and expensive to undo.
The wrong claim, in public
A confident sentence about a capability, a price or a result that is not exactly true. Cheap to catch in review; expensive to retract.
The wrong recipient
A sequence that goes to a competitor, a customer who already complained, or the same person twice. Nothing about the copy is wrong — the routing is.
Budget on an untested variant
Ad spend behind creative nobody looked at. The loss is not the impressions; it is the week you spent believing the test.
Voice drift
Everything is defensible and none of it sounds like the company. It shows up gradually, which is why it survives.
No record
It went out, it worked or it did not, and the reason lives in a chat thread nobody will open again.
Where the gate belongs
At the boundary where work leaves the company. Three boundaries cover almost everything: before an email or sequence sends, before budget is spent, before content is published to a public destination.
Everything upstream of those — research, drafts, captions, ad variants, summaries — can run without a gate, and should. That is where the hours are, and none of it is irreversible. Gating preparation is how teams end up feeling like the approval step is the problem.
The trust ladder: off, then on where it is earned
Autonomy is a setting, not a belief. The useful version moves in one direction at a time and can always move back.
01
Start with the gate on everything
Every send, every spend, every publish stops for a person. This is the default because it is the state you can reason about.
02
Watch one desk for a few weeks
Approve the work and notice what you are actually changing. If you approve thirty items without editing one, that is data.
03
Widen where the evidence is
Turn autonomy on for the desk or loop that has earned it — all, some, or none. Keep the gate everywhere else.
04
Keep the record either way
Approved or automatic, the same evidence gets written. That is what makes the next widening a decision instead of a hope.
05
Move it back when something changes
A new market, a new claim, a new person on the desk. The gate is a setting, not a stage you graduate from.
What to keep as evidence
An approval that leaves no record is a feeling. The minimum worth keeping: what was prepared, who approved it, when it went out, and what the destination said back.
That last part is the one teams skip. Approve is not the finish line — a send can bounce, a post can fail, a platform can accept the call and publish nothing. Reading the destination afterwards is what turns an approval into evidence, and it is what makes widening autonomy a decision rather than a guess.
Related: Destination verification · CRM as operating evidence.
How 77os implements it
AI prepares content, sales outreach, ads, creative work and operations. Approvals and decisions arrive in one place — Needs you — and the rest stays quiet. Nothing sends, spends or publishes without Human Approve.
Autopilot is Off by default and can be turned on per desk or per loop as trust builds — all, some, or none — and changed at any time. Destination verification and CRM interaction records close the loop where supported, so the same evidence exists whether a person approved the action or a setting did.
77systems installs the system, trains the team on their real work, and verifies outcomes landed. Software is how that is delivered, not what is sold first.
Related: What is a business operating system? · 77os vs AI SDR tools · Sales outreach with Human Approve.
FAQ
- What is human-in-the-loop automation?
- Human-in-the-loop automation is a pattern where software or AI prepares work and a person approves it before it takes effect. The machine does the preparation; a named human makes the decision that reaches a customer, spends money or publishes in public.
- What is the difference between human in the loop and human on the loop?
- In the loop means the work stops for approval before the action happens. On the loop means the work proceeds while a person supervises and can intervene, reviewing the record afterwards. In the loop prevents mistakes; on the loop catches them.
- Does an approval step slow the work down?
- It adds a review, and it removes retraction, apology and rework. The practical answer is to put the gate only where the consequence is external — sends, spend and publishes — and leave preparation ungated, which is where most of the time goes anyway.
- Where should the approval gate sit?
- At the boundary where the work leaves the company: before an email or sequence sends, before budget is committed, before anything is published to a public destination. Internal drafts, research and variants do not need a gate.
- Is human-in-the-loop the same as turning autopilot off?
- It is the same idea stated as a setting. In 77os, Autopilot is Off by default, which means every live action waits for Human Approve. It can be turned on per desk or per loop as trust builds — all, some, or none — and changed at any time.
- How do you know when to widen autonomy?
- When the approvals on a given desk stop changing anything. If a team approves a run of work without edits, and the destination evidence agrees with what was expected, that desk has earned a wider setting. Keep the evidence either way so the next decision is based on a record.
- Does this work with any AI model?
- Yes. The pattern is about where the decision sits, not which model does the preparation. 77os is flexible across providers, so the gate and the evidence stay the same as the tools underneath change.
See the gate on your own work
A demo runs one of your loops end to end: prepared, Approved, and verified at the destination.
77systems · 77os
Santa Monica. founder@77systems.ai