Abstract
Putting a person in charge of an AI workflow sounds simple until you try it. Ask for approval on everything and the person drowns. Ask for approval on nothing and nobody is in charge. This paper describes the four levels we use to decide where a person signs off, what makes an approval quick to do well, how a step earns the right to run on its own, and which steps never do.
1. Two ways approvals go wrong
The first failure is the bottleneck. Every action waits for a person, so the workflow saves nobody any time. People give up on it and go back to doing the work by hand.
The second failure is quieter and worse. The agent gets good, the person approves a hundred correct drafts in a row, and approving turns into clicking. The approval screen is still there, so everyone believes a person is checking. Nobody really is.
Both come from treating every step the same. The fix is to match the amount of checking to what's at stake in each step.
2. The four levels
| Level | What it covers | Examples | How a person is involved |
|---|---|---|---|
| 1. Looks things up | Reading and summarizing, nothing changed | Finding a customer's last order, summarizing an email thread | Runs on its own, with a record of what it looked at |
| 2. Works inside your tools | Changes only your team sees | Updating a CRM record, preparing a draft | Checked before or after, your choice |
| 3. Reaches your customers | Anything a customer or vendor will see | Sending a quote, a payment reminder | Approved before it goes out, until you decide otherwise |
| 4. Judgment calls | Decisions with real consequences | Approving an admission, offering a discount | Always a person |
The level belongs to the step, not to the agent. Sending a payment reminder is level 3 whether the agent has sent ten or ten thousand.
3. Making approvals quick to do well
An approval is only worth something if the person can make a real decision in a short time. We design every approval screen to answer three questions without opening another system:
- What did the agent do? The draft, the record change, or the message, exactly as it will go out.
- What did it base that on? The document, email, or record it used, next to the draft.
- What wasn't it sure about? Anything missing, unusual, or outside the normal pattern, called out at the top.
When an approval shows all three, a careful review takes seconds. When it doesn't, people skip the check, and the approval becomes a formality.
Batching helps too. A bookkeeper reviewing twenty payment reminders at once, sorted with the unusual ones first, does a better job than one interrupted twenty times a day.
4. When a step can run on its own
Some level 2 and level 3 steps can eventually run without approval. We think that decision should be:
- Made by a person. You choose, step by step. It never happens automatically.
- Based on a track record. The step has run under approval long enough, with few enough corrections, that checking it no longer catches anything.
- Easy to undo. If the agent's work starts needing corrections again, the step goes back to requiring approval.
Even when a step runs on its own, every action is still recorded. You can always look back at what happened and why.
5. Steps that never run on their own
Level 4 steps stay with a person permanently. Approving a home health admission or deciding whether a claim is covered carries consequences someone has to answer for, and that someone should be a person.
The agent's job at level 4 is to make the decision easier, not to make it. It gathers the facts, lays out the options, and flags what's unusual. Then a person decides.
6. What we haven't proven yet
This model comes from how we design workflows, not from years of measured results. We haven't yet published data on how long steps typically take to earn independence, or how often they get moved back. We'll share that as real workflows build a track record.