# Human in the loop

Human in the loop means a person's decision is required before a specific action completes. The design question is never whether to have one — it is where, because a checkpoint on everything is abandoned within a fortnight and a checkpoint on nothing is discovered after something has gone out. Put it where the action is expensive to undo.

- **Status:** Reference
- **Audience:** both
- **In the app:** #/needs-you, #/approvals
- **Last verified:** 2026-09-10
- **Canonical:** https://connectbyjbrh.com/docs/technology/human-in-the-loop/

## Placing the checkpoint

Two properties decide whether a step deserves one: how hard it is to reverse, and how visible a mistake is to somebody outside the business. Everything else — how clever the step is, how much the model was trusted with — is a distraction.

| Action | Reversible? | Checkpoint |
|---|---|---|
| Reading and classifying incoming mail | Nothing left the building | No |
| Drafting a reply | A draft can be deleted | No |
| Sending to a customer | It has been read | Yes, unless the workspace has chosen otherwise |
| Placing a call | It rang | Yes, at first |
| Recording a memory or a fact | Editable, visible | No, but it must be visible |
| Clearing a do-not-contact entry | Legally and commercially costly | Always, and never by the Assistant |

That is a shape, not a policy. In Connect the policy is yours: [autonomy](/docs/autonomy/) is set per channel and can be narrowed to one mailbox, one number or one contact, and the narrowest scope wins.

## The review-speed problem

A checkpoint only exists if reviews actually happen. A queue that takes a minute per item and receives eighty items a day is not a control — it is a backlog that will be cleared by someone approving in bulk without reading, which is worse than no checkpoint because it produces a record of approvals nobody made.

- **Everything needed to decide in one place.** The message being answered sits above the draft, not on another screen.
- **Ranked, not chronological.** A first-time enquiry held for an hour is a different debt from a held reply on a week-old thread. **Needs You** is ranked for that reason.
- **Edit and approve as one action.** If correcting means rewriting elsewhere, people approve things they would have changed.
- **Self-draining.** Items whose cause has cleared leave on their own; a queue that keeps resolved work trains people to skim it.
- **Narrow scope rather than blanket holds.** Holding one account's replies keeps attention where it is deserved — see [human approval without blocking low-risk work](/research/approval-without-blocking/).

## Does Connect use a human in the loop?

**Used**, and it is a setting rather than a fixed behaviour. The four autonomy modes are `off` (the channel is disabled), `draft_only` (prepare, never send, do not even ask), `ask_before_send` (every outbound needs a human yes) and `autonomous` (act within the gates). They resolve across four scopes — contact, endpoint, channel, workspace — narrowest first.

`draft_only` is the mode people overlook and the most useful one early. It writes what Connect would say and stops, without queueing anything for approval, so a workspace can watch the quality of the drafts for a fortnight without anyone having to keep saying no.

Held work waits in **Needs You**, which is also where operational problems land — voice line health escalates into the same queue for both audiences — so there is one place a person looks rather than several. Every approval and every rejection is recorded against the person who made it in the [decision log](/docs/technology/audit-log/), which answers 'who released this' long after everyone has forgotten.

## How the pattern fails

**Rubber-stamping** — Volume outruns attention and approval becomes a reflex. The fix is fewer, better-targeted holds, not a reminder to read carefully.
**The queue as a bottleneck** — Work that could have gone out in seconds waits for a person who is asleep. Narrow the scope instead of widening the mode.
**A stale approval** — The thread moved on while the draft waited, so approving sends an answer to a superseded message. Regenerating uses the newer messages; editing does not.
**Responsibility without authority** — A reviewer who cannot change what they are approving is a signature, not a control.
**Silent expiry** — Items that vanish unresolved leave the business believing something was handled. Draining must have a visible cause.

## Questions

### Does a held item stop the customer's clock?

No — and this is the part that surprises people. Nothing has been handed to a provider, so there is no message in flight and nothing in your sent folder. A held reply is not a slow reply; it is no reply, and the customer is still waiting.

### Can approval be required for one customer only?

Yes. Autonomy resolves at four scopes and the narrowest wins, so a rule set against a single contact overrides the channel's. That is the usual way to handle an account where every word matters without slowing everything else down.

### What is the difference between draft-only and ask-before-sending?

`ask_before_send` puts the reply in front of a person and waits for a yes. `draft_only` writes it and stops, deliberately without asking. The first is a control on sending; the second is a way of observing what Connect would say.

## Related

- [What Connect may do](https://connectbyjbrh.com/docs/autonomy/)
- [Human approval without blocking low-risk work](https://connectbyjbrh.com/research/approval-without-blocking/)
- [Separating the agent's authority from the person's](https://connectbyjbrh.com/research/authority-separation/)
- [Agent observability](https://connectbyjbrh.com/docs/technology/agent-observability/)
- [Unsupported claims](https://connectbyjbrh.com/docs/technology/hallucination/)

## What this page is based on

- Connect architecture source pack — docs-source/sources/GENERAL.md §5, autonomy, Needs You and the audit trail
- Connect capability registry (docs-source/facts.py) — autonomy_modes, approvals, needs_you, audit_trail
- `backend/app/autonomy.py` — modes and scope order, per the source pack