Human-in-the-loop patterns that don't slow you down

“If a person has to approve everything, why automate at all?” It’s a fair question, and the answer is that people shouldn’t approve everything. They should approve the right things, quickly, with everything they need in front of them.

Here are the five patterns we use most.

1. Approve by consequence, not by step

Don’t ask “which steps need a human?” Ask “which outcomes would be expensive to undo?” Money leaving the business, records deleted, the first email to a new customer, anything a regulator would care about. Those get an approval step. Reading, sorting, drafting and looking things up don’t.

In practice this means most runs never touch a person. In an invoice workflow, matching a $40 invoice to a purchase order happens alone; paying a $4,800 invoice from a new vendor waits for someone in finance.

2. One click, full context

An approval that takes five minutes of digging gets rubber-stamped or ignored. We design every approval so the decision takes seconds:

• the proposed action in one sentence,

• the evidence next to it (the email, the PDF, the order),

• why it was escalated (“over your $2,000 threshold”, “new vendor”),

• three buttons: approve, edit, reject.

We put it where the approver already works — Slack, Teams, the help desk — never in a new dashboard.

3. Thresholds you can move

Every rule that sends work to a person has a number in it: an amount, a confidence score, a customer tier. We make those numbers settings, not code, and we review them with you after the first month. If finance approved every invoice under $500 without a change, raising the threshold is a five-second decision backed by data.

4. Confidence is a signal, not a verdict

Models can report how sure they are, and it’s useful — but only as one input. We combine it with plain checks: does the total add up, does the vendor exist, does the PO match. A confident model that fails a plain check still goes to a person. A low-confidence result on a $12 receipt might not need one.

5. Every override teaches the system

When a person edits or rejects what the agent proposed, we log the original, the change and the reason. Those cases go into the evaluation set, so the next version of the prompt or the workflow is tested against them. Over a few months, the share of approvals that need a change drops — and that’s your signal to relax a threshold, not before.

What this looks like in numbers

On a typical finance workflow we see something like this after two months:

• 70–80% of runs complete without a person,

• the rest reach an approver in under a minute,

• the approver changes the proposed action in fewer than one case in ten.

Your numbers will differ. The point is that a well-placed person costs seconds per decision, and in exchange you get an automation you can actually trust with money.

The rule we don’t break

Agents propose; people approve anything that’s expensive to undo. We’d rather ship something that asks too often in week one than something that acts alone and has to be switched off in week three.