Skip to content

AI Systems

Human-in-the-loop AI: review is a state, not a pause

Let AI draft the quote, the reply or the summary, and keep the decision with a person. The draft, its sources and its status are saved, so the reviewer can take until morning and nothing is lost. When they change something, the change is recorded too.
Quote Q-212 · J-442drafted from T3's notesWater heater, 50 gal, installed$1,840Labour2 h · $2603 h · $390Haul away old unit$95Total$2,195$2,325operations managerQuote emailed to H.R.saved · waiting 14 h 35 mdrafted17:38awaiting review17:40edited08:12labour 2 h → 3 happroved08:15sent08:15

One step needs judgment; the rest doesn't

Many processes could run on AI except for one step: approving a quote, checking an extracted figure, answering a sensitive message. Human-in-the-loop design lets everything around that step run on its own while the judgment stays with a person, at a defined point, with a record of what was decided.

The pattern is short. Rules or a model prepare a draft; it's validated and saved; the case is marked awaiting review; a reviewer approves, edits or rejects; the decision is saved; the next step runs. Two things make it reliable: the review is stored state rather than a paused script, and the reviewer gets a screen built for deciding.

Review is not an error screen

The reviewer should be able to decide in one look: what's proposed, why, what the AI wasn't sure about, and what approving will do.

Review · Quote Q-212

J-442 · H.R., Northside · drafted 17:38 from T3's notes

  • Water heater, 50 gal, installed$1,840
  • Labour · 2 h$260
  • Haul away old unit$95
  • Total$2,195
  • T3 job notes, 16:55
  • 3 photos
  • Price list v8

Labour estimated from notes: “tight closet, old unit seized”. Check hours.

Approving emails Q-212 to H.R. and sets a follow-up for Thursday.

Approve, Edit, Reject

  1. What's proposed

    The draft itself, laid out the way the customer will see it.

  2. Where it came from

    The notes, photos and price list it was drafted from, one click away.

  3. What it wasn't sure about

    Anything inferred rather than read is flagged, so attention goes where it's needed.

  4. What approving will do

    The consequence, stated before the click.

  5. Three ways out

    Approve, edit, or reject with a reason. None of them is a dead end.

Three decisions, three records

Whatever the reviewer does is saved with the draft, so you can see later what was proposed and what was decided.

Q-212 · approved

status
approved
reviewer
operations manager
at
08:15
changes
none
next
quote emailed to H.R., follow-up set
feedback
counted as an accepted draft

How much review is enough

Set by what a mistake would cost, not by how nervous AI makes everyone.
  1. 01

    Spot-check a sample

    When: mistakes are cheap and easy to undo

    Example: internal job summaries

  2. 02

    Review what it's unsure about

    When: most cases are routine and a few aren't

    Example: extracted invoice totals that don't match the purchase order

  3. 03

    Review every one

    When: it reaches a customer or moves money

    Example: quotes, refunds, apology emails

  4. 04

    Keep it with a person

    When: the judgment is the job

    Example: disputes, safety calls

What human-in-the-loop AI hands off

In

Draft
quote, reply, extracted field
Sources
what it was made from
Why
model output or rule
Review state
awaiting review → decision

Out

Decision
approved, edited, rejected · who · when
Released action
only after the decision
Feedback
edits and reasons, for testing

Worth building where a mistake is costly

Good fit

  • A mistake would be costly, visible to customers, or hard to reverse.
  • There's enough volume that preparing the work automatically saves real time.
  • Someone owns the decision and can be reached.

Not the right fit

  • Every step needs review. That's a manual process with extra software.
  • Reviewers approve everything without looking. The rules need fixing instead.

Questions about human-in-the-loop AI

What if the reviewer doesn't respond?

The case stays in awaiting review with its draft and context saved. Reminders and escalation to a second reviewer are part of the design, on times you choose.

Can reviewers change the draft, or only approve it?

They can edit. The original and the edited version are both kept, so you can see what the AI proposed and what went out.

Doesn't review cancel out the time saved?

Checking a prepared draft is much quicker than writing one, and everything around the decision runs on its own. If reviewers start approving without looking, the review level is wrong for that step.

Where does the review happen?

Where the reviewer already works: an internal screen, a notification that opens the case, or a queue in a tool you already use. What matters is that it shows the context and what approving will do.