Human-in-the-loop guardrails
Human in the loop means a person reviews, approves or can stop an automated process at a defined point. It serves two purposes named in the framework: reliability (catching errors before they reach people) and trust (employees and leaders knowing that a person is accountable).
It is also becoming a legal expectation. Examples covered in Lesson 1.3 include the EU AI Act’s human- oversight requirements for high-risk systems, data-protection limits on solely automated decisions, and California’s SB 947, which from 1 July 2027 requires a human reviewer to corroborate discipline or termination decisions that rely primarily on an automated system.
Guardrail patterns
| Pattern | How it works | Use when |
|---|---|---|
| Approval gate | The workflow pauses until a person approves. | Anything sent outside HR; anything that is hard to undo. |
| Edit before send | The reviewer can change the content, and the edited version continues. | AI-drafted messages. |
| Exception routing | Standard cases proceed; unusual ones go to a person. | High volume with clear rules for what counts as unusual. |
| Uncertainty flag | The AI step is told to return a marker such as NEEDS REVIEW when unsure or when input is missing; a filter routes those to a person. | AI classification or extraction. |
| Sampling audit | A person reviews a random sample of completed runs on a schedule. | Low-risk, high-volume automations that run without approval. |
| Two-person rule | Two people must approve. | Changes to pay or status. |
| Safe timeout | If nobody responds, the workflow stops and alerts someone. It never proceeds by default. | Every approval gate. |
| Off switch | A named person can pause the automation immediately, and knows how. | Every automation. |
How Zapier’s Human in the Loop works According to Zapier’s documentation, the app offers two actions and a trigger. Request Approval pauses the run and asks a reviewer to approve or decline, with an option to let them edit the content. Collect Data pauses the run to ask a person for additional information. New Approval Requested lets a second Zap send the request through another app. Reviewers can be notified by email or Slack, you set how long to wait and what happens on timeout, and the outcome can be used in later filters. See Zapier’s overview, its Request Approval instructions and its guide.
Making the review meaningful
- Show the reviewer the inputs alongside the draft, so they can check one against the other.
- Give them time. If approvals arrive faster than a person can read them, the design is wrong.
- Give them authority to edit and decline, and a clear place to send problems.
- Watch the override rate. If a reviewer has approved hundreds of items without a single change, either the automation is excellent or the review has become a formality. Sample some to find out.
- Rotate or refresh reviewers to counter automation bias.
Matching oversight to risk
| If the automation… | Minimum oversight |
|---|---|
| Only moves or files internal data, and is easily reversed | Error alerts and a sampling audit. |
| Sends routine reminders or notifications | Tested template, sampling audit, off switch. |
| Sends personalised or AI-drafted content to a candidate or employee | Approval gate with edit before send. |
| Classifies or prioritises requests using AI | Uncertainty flag, exception routing, sampling audit; a person acts on the result. |
| Affects a person’s pay, status, opportunities or record | A person makes the decision. Automation may prepare information only. Legal review. |
Documentation that builds trust Keep a one-page runbook for each automation: purpose, owner, trigger, steps, data used, checkpoints, what to do when it fails, how to switch it off, change log, last review date. If you cannot explain an automation to an employee it affects, it is not ready.
Jot down key ideas in your own words. Saves automatically.