Where to put the human in an AI workflow

---
title: "Where to put the human in an AI workflow"
description: "A practical way to place approval points, stop rules and review in an AI workflow, so it can run on its own where it should and nowhere else."
url: https://lidvizion.ai/guides/where-to-put-the-human-in-an-ai-workflow/
site: Lid Vizion
index: https://lidvizion.ai/llms.txt
---

[Field guides](https://lidvizion.ai/guides/) 4 steps

# Where to put the human in an AI workflow

A practical way to place approval points, stop rules and review in an AI workflow, so it can run on its own where it should and nowhere else.

For: Anyone designing or approving an AI agent or automation that touches customers, money or records.

“Human in the loop” is often added as an afterthought: a review step on everything, which slows the work, or on nothing, which removes the safety net. Placing people on purpose is a design task. These four steps decide exactly where a person approves, reviews or takes over.

## Step 1 — Sort the actions by consequence

**Goal:** Separate what is safe to automate from what is not.

1. List every action the workflow can take: read, draft, update, send, pay, delete.
2. Mark each as reversible or irreversible, internal or external.
3. Flag anything touching money, legal commitments, personal data or customers.
4. Everything irreversible, external or flagged starts with an approval step.

**You finish with:** An action list with a consequence level for each.

## Step 2 — Place approval points

**Goal:** Put a named person in front of the actions that need one.

1. Assign each approval to a role, not “the team”.
2. Show the approver the evidence the workflow used, not just its conclusion.
3. Make approving, editing and rejecting one click each.
4. Set a time limit and a fallback when an approval is not answered.

**You finish with:** Approval points with owners, evidence and fallbacks.

## Step 3 — Write the stop rules

**Goal:** Decide when the workflow must stop and hand back to a person.

1. Low confidence: below an agreed threshold, route to review.
2. Missing or conflicting information: stop, do not guess.
3. Out-of-scope requests: hand back with a short summary.
4. Repeated failures: pause the workflow and alert the owner.

**You finish with:** A stop-rule list the workflow enforces automatically.

## Step 4 — Earn autonomy with evidence

**Goal:** Remove approval steps only when the record supports it.

1. Log every decision: what ran, what changed, who approved.
2. Review approval outcomes weekly: how often was the draft accepted unchanged?
3. Remove an approval only after a sustained run of unchanged approvals, and keep spot checks.
4. Re-add the approval immediately if error rates rise.

**You finish with:** A documented, reversible path from supervised to autonomous.

## Common pitfalls

- Reviewing everything, so reviewers rubber-stamp.
- Showing the approver a conclusion without the evidence.
- No fallback when an approver is away.
- Removing approvals because of a good demo instead of a good record.

Want a hand getting started?

## Start with a workflow audit

We pick the workflow with you, set the baseline and write the bounded design.

[See the audit](https://lidvizion.ai/workflow-audit/) [Talk this playbook through](https://lidvizion.ai/contact/?need=ai-workflow&from=%2Fguides%2Fwhere-to-put-the-human-in-an-ai-workflow%2F#write)

## More guides like this

- [Measure the ROI of an AI workflow](https://lidvizion.ai/guides/measure-the-roi-of-an-ai-workflow/)
- [Triage support requests with AI](https://lidvizion.ai/guides/triage-support-requests-with-ai/)

## Related answers

- [What is an AI agent in a business setting?](https://lidvizion.ai/answers/what-is-an-ai-agent-for-business/)
- [How long does it take to automate one business workflow with AI?](https://lidvizion.ai/answers/how-long-does-it-take-to-automate-a-workflow/)

Bring the messy version

## One workflow. A better way to run it.

Show us the tools, the handoffs, and the part that keeps slowing your team down.

[Let’s map your workflow](https://lidvizion.ai/contact/) · [Read our approach](https://lidvizion.ai/approach/)