Human checkpoint map

For each AI task your team repeats: the level of oversight, who approves what, what they see and how to stop and undo.

Download

Blank template (.md)Fictional example (.md)

Principles involved

Overview

A checkpoint map turns “human in the loop” into something specific: for each task, where the person is and what they decide. Start with the three to five AI tasks the team does most, and add more as you go.

Choosing the level

Ask three questions for each task. Can it be undone? Who else is affected? What does it cost if it’s wrong? Anything irreversible, external or about a person starts at level 2 (AI proposes, a person approves) or level 4 (you do it). Principle 1 explains the four levels.

Warning signs

01Approvals happen in seconds, every time.
02Nobody has rejected anything in months.
03The reviewer can’t see the actual recipient, amount or wording.
04When the reviewer is away, the action goes ahead anyway.

The template

Each field says what to record. The fictional example shows what a finished record can look like.

Fictional example: the refunds task of the customer success team at Lumen, an invented software company.

01Task: Describe the task in a few words, as the team would say it.

Fictional example

Refunds for billing complaints.
02Level: Choose one: AI drafts, AI proposes, AI acts within limits, or you do it.

Fictional example

AI proposes.
03Why this level: Answer the three questions: can it be undone, who else is affected, and what it costs if it’s wrong.

Fictional example

Refunds are hard to reverse once paid, they affect customers and revenue, and mistakes cost money and trust.
04Who approves: Name the role that approves, and a backup.

Fictional example

The team lead on duty. Backup: the support operations lead.
05What they see before approving: List what the reviewer needs to decide: the exact action, the amount, the recipient and the evidence.

Fictional example

The customer, the amount, the policy clause the agent cited and the draft reply.
06Limits enforced by the system: Write the limits set in permissions or settings, not just in the prompt.

Fictional example

The agent can create refund requests but can’t issue them. Requests over €200 need a second approver.
07If nobody responds: Say what happens when the approver doesn’t answer. Silence never counts as approval.

Fictional example

Requests wait. After four hours, the backup is notified.
08How to stop and undo: Say who can pause the task and how to reverse its effects.

Fictional example

Any team lead can pause the agent. Refunds requested in error are canceled before payment.
09Review date: When will you check that this checkpoint still works?

Fictional example

Monthly, alongside the refund report.

Sources

What each source establishes, and its limits. The practices and recommendations on this page are ours, and the facts come from the sources. See every source we use.

Model AI Governance Framework for Agentic AI Infocomm Media Development Authority, Singapore · Version 1.5, May 20, 2026 (first published January 22, 2026) · Voluntary framework Keeps organizations and their human supervisors accountable for agents’ actions; asks for human approval at significant checkpoints such as deleting data, sending messages and payments; treats very low override rates and very fast reviews as possible signs of rubber-stamping; prefers approvals enforced by system controls over prompts. Limits: Voluntary guidance, not law. Its case studies were supplied by the companies themselves.
Ethics Guidelines for Trustworthy AI High-Level Expert Group on AI, set up by the European Commission · April 8, 2019 · Expert guidelines Describes three kinds of human oversight: human-in-the-loop (intervention in every decision cycle, often neither possible nor desirable), human-on-the-loop (involvement in design and monitoring of operation) and human-in-command (oversight of overall activity, including whether, when and how to use the system). Limits: Not binding, and not an official position of the Commission.
Automation bias: a systematic review of frequency, effect mediators, and mitigators Journal of the American Medical Informatics Association · 2012 (online June 2011) · Peer-reviewed review Defines automation bias as the tendency to over-rely on automation, leading to errors of commission (following wrong advice) and omission (missing problems not flagged), and finds that training, accountability and how advice is presented can reduce it. Limits: Reviews clinical decision support from before generative AI.