Human-in-the-loop design: where a person should approve, review or just watch an AI system

Every serious AI deployment eventually asks the same question: when does a person need to be involved? "Human in the loop" is the usual answer, but the phrase hides at least three different designs, and choosing the wrong one produces either a rubber stamp or a bottleneck. This explainer separates them and gives a way to decide.

Three patterns, not one

Approve before action. The system proposes and a person confirms before anything happens. This is the safest pattern and the slowest. It fits payments, outbound messages to customers, deletions and anything touching credentials.

Review after the fact. The system acts, and a person samples or audits the results. This fits high-volume, reversible work such as tagging support tickets or drafting internal summaries, where a wrong answer is cheap to fix.

Monitor and intervene. The system runs on its own while a person watches a dashboard and can pause or stop it. This fits continuous processes where you need an off switch more than a sign-off on each item.

A single workflow often uses all three, at different steps.

What the current desktop agents do

This week's launches show the approve-before-action pattern in practice. Anthropic released Cowork on January 12 as a research preview in its macOS app, initially for Max subscribers. Users grant it access to specific folders and it operates only within them. It asks before significant actions, and it can still carry out destructive actions such as deleting files if instructed ^[2][3]. Anthropic says it has built defenses against prompt injection, in which malicious content in files or web pages tries to steer the agent, but describes securing real-world AI actions as an active area of research ^[2]. In other words, the design relies on scoped access plus confirmation, and the vendor itself does not claim that this removes the risk.

What a reviewer needs to be effective

A checkpoint only adds safety if the person at it can actually do the job. As a design checklist, a reviewer should be able to:

• Understand the system's capabilities and limits, and spot anomalies and unexpected performance.

• Stay aware of automation bias, the tendency to over-rely on the system's output.

• Correctly interpret the output.

• Decide, in any particular situation, not to use the system or to disregard, override or reverse its output.

• Intervene or halt the system safely, such as with a stop button.

The second item deserves emphasis. An approval step that a busy employee clicks through without reading is not oversight, however it looks on a process diagram.

How much autonomy do people actually grant?

Anthropic's January 15 Economic Index report, based on samples of Claude conversations from November 13 to 20, 2025, gives a picture of real usage. On Claude.ai, the augmentation share (a person working with the model) was 52 percent in November 2025, up from under 50 percent in August, while automation was 45 percent. API traffic, which is what companies build into products, is automation-heavy: about three-quarters of interactions are automation ^[4].

The picture is mixed: automation's share on Claude.ai is higher than in January, though the report describes a shift back toward augmentation since August, and software callers on the API delegate far more. Either way, it is important to decide on purpose where the human checkpoint sits.

A simple decision rule

For each step in a workflow, ask three questions:

1. Can it be undone? If not, require approval.

2. Who is affected? If the output reaches a customer, an outside party or a payment system, require approval or a tight review sample.

3. Can the reviewer tell if it is wrong? If the reviewer cannot judge the output, adding them as a checkpoint adds delay without adding safety. Fix the information they see first.

Then measure the checkpoint. Track how often reviewers change or reject the AI's output; a rate near zero for a long time suggests either excellent quality or a rubber stamp, and you need a sample audit to tell which.

The practical implication

Start every new AI workflow with approval before action on anything irreversible or customer-facing, and relax it step by step only when review data shows the outputs hold up. Design the reviewer's screen, what they see and how easily they can override, as carefully as the model itself.