Decide Where a Human Must Stay in the Loop

Why this matters

Not every step should be automated or assisted. Some steps exist precisely because judgment, accountability, or context is required. The mistake is deciding this by convenience — whatever the tool can do gets done by the tool, and the human presence shrinks by accident. Human checkpoints should be deliberate, not accidental leftovers.

What you will be able to decide after this

Which steps carry judgment or accountability that cannot be delegated.
Where a tool may assist, and where it may not decide.
Whether the step's error cost justifies keeping a human in it.

Core lesson

For each major step in the map, ask four questions:

What happens if this step is wrong? Trace the failure one step downstream. If the answer is "someone catches it later," that later someone is the real control point, and you should say so.
Is the cost of error high enough that a human must own it? A step that only moves a ticket along costs little when wrong. A step that commits money, health, safety, or reputation costs a lot. The cutoff is yours to set — but set it deliberately, per step.
Does this step require information the system cannot reliably have? The unrecorded context — the customer's history, the exception someone mentioned in passing, the unwritten policy. If the deciding information lives outside the system, the system should not decide.
Is this a control point? Approvals, exception handling, final accountability. These exist to be checked. Automating the check away is not an improvement; it is a redesign of who is accountable.

Then classify each step:

Human-required — judgment, accountability, or context that cannot be delegated. The step stays human, full stop.
AI-assist candidate — a draft or recommendation is fine, but a named human accepts, revises, or rejects before the work proceeds.
Full automation candidate — error is cheap, information is complete, and no control point sits here.

Example constraint language for a human checkpoint:

This step remains human-owned.
The system may prepare a draft or recommendation.
A named person must review against [specific criteria] and explicitly
accept, revise, or reject before the work proceeds.
No silent auto-advance.

The last line matters most. "No silent auto-advance" is the sentence that turns a checkbox into a checkpoint.

Worked example

A renewal workflow maps to three steps: eligibility check, pricing, and confirmation to the customer. The eligibility check reads only recorded data and errors are caught by the pricing step — full automation candidate. The pricing step generates a quote that commits money: the system may draft the numbers, but a named person reviews against the pricing criteria and explicitly accepts, revises, or rejects before the quote goes out. No silent auto-advance. Confirmation to the customer carries tone and relationship risk the system does not understand: human-required, one sentence: "A wrong confirmation reads as a broken promise, and the system cannot reliably know which customer that matters to."

Practice

Take the map from Module 1. Run the four questions on the three most important steps. For each, write what happens if the step is wrong, and whether the deciding information lives inside or outside the system.

Apply — produce the artifact

Take the Module 1 map and mark every step: Human-required / AI-assist candidate / Full automation candidate. For every Human-required step, write one sentence explaining why. Make the AI-assist candidates explicit about who reviews and what happens if they do not act.

Verify

Review the marked map with someone who will be accountable if the workflow fails. If they disagree on any high-stakes step, resolve it before moving on.

Sources

This module is original practice guidance based on the authoring standard and does not depend on a specific external factual claim. Editorial review is still required.