Human-in-the-loop: where to put approval gates in AI workflows
Human-in-the-loop means designing AI workflows so a person reviews or approves certain steps before they take effect. Place approval gates where actions are hard to reverse, affect customers or money, carry legal risk, or where the AI is uncertain. Low-risk, reversible steps can run automatically with monitoring. Good approval screens show what will happen, why, and make approving or correcting quick.
What does human-in-the-loop mean?
In a human-in-the-loop design, an AI system does most of the work but a person checks or approves specific steps. It sits between fully manual processes and fully automated ones. The term applies to AI agents, AI automations and any system where AI decisions have consequences.
Where should approval gates go?
| Put a gate when the step… | Example |
|---|---|
| Cannot be easily undone | Sending an email to a customer, deleting records |
| Moves money | Refunds, payments, discounts above a limit |
| Has legal or compliance impact | Contract terms, hiring decisions, credit decisions |
| Affects a customer directly | Account changes, cancellations |
| Falls outside normal patterns | Unusual amounts, new suppliers |
| Comes with low AI confidence | Ambiguous requests, conflicting data |
How do you decide, step by step?
- Map the workflow: List every step from trigger to outcome.
- Score each step: Rate impact if wrong and how easily it can be reversed.
- Set thresholds: For example: refunds under a set amount run automatically, larger ones need approval.
- Let AI flag uncertainty: Route low-confidence cases to a person automatically.
- Review regularly: Loosen gates where people rarely change anything; tighten where errors appear.
How do you design approval screens people actually use?
- Show the action: Exactly what will happen, in plain language.
- Show the reason: Why the AI proposed it, with the key evidence.
- Make editing easy: Let people adjust before approving, not just accept or reject.
- Batch similar items: Approve routine items together to save time.
- Capture feedback: Record why items were rejected to improve the system.
- Respect attention: Too many low-value approvals lead to rubber-stamping.
These patterns build on our UX principles for AI design.
What is approval fatigue?
When people are asked to approve too many trivial items, they start approving without reading. That makes the gate meaningless. Keep approvals for decisions that matter, batch the rest, and track how long reviewers spend and how often they change anything. If the override rate is near zero for months, the gate may be unnecessary; if it is high, the AI needs improving before you automate further.
Do regulations require human oversight?
For some uses, yes. The EU AI Act requires human oversight for high-risk AI systems, such as those used in hiring or credit decisions, with obligations now due from December 2027. Our EU AI Act guide covers the timeline. Data protection laws can also give people rights regarding automated decisions.
What does this look like in different teams?
| Team | Runs automatically | Needs approval |
|---|---|---|
| Sales | Lead research and enrichment, draft emails | Sending outreach, changing deal stages |
| Support | Ticket classification, order lookups, draft replies | Refunds above a limit, complaint responses |
| Finance | Invoice data extraction, matching payments | Paying suppliers, adding new suppliers |
| HR | Interview scheduling, policy answers | Shortlisting and hiring decisions |
| Marketing | Drafting posts and reports | Publishing, paid spend changes |
The pattern is consistent: AI prepares, gathers and drafts; people decide where money, customers, legal exposure or reputation are involved.
How should gates change over time?
Start with more approval than you think you need. As you collect data on how often reviewers change the AI’s proposals, you can raise thresholds, switch some steps to after-the-fact sampling and keep approvals only where they add value. Record every change to the gates, with the evidence behind it, so you can explain your oversight approach to managers, auditors or regulators.
What should you measure?
- Approval volume and time spent per approval
- Override rate: how often people change or reject AI proposals
- Errors caught at the gate versus errors found later
- End-to-end time compared with the manual process
- Reviewer satisfaction and workload
What mistakes should you avoid?
- Requiring approval for everything, which slows work and causes fatigue
- Removing all approvals after a successful pilot without monitoring
- Approval screens that hide the reason or evidence
- No record of who approved what and when
- Assigning approvals to people without time or authority to review them
A short written policy that lists each gate, its threshold and its owner keeps the system understandable as it grows.
Done well, human-in-the-loop design lets teams automate far more than they could safely automate otherwise, because people trust a system they can see into and correct.
Review points matter most in regulated sectors; our guide to AI in healthcare shows how clinicians stay in control of every clinical decision.
How UIUX Studio helps
We design approval flows and review screens for AI agents and automations, balancing speed with control. Talk to our Agentic AI & Automation and product design teams.
Frequently asked questions
What is human-in-the-loop AI?
A design where people review or approve certain AI steps before they take effect.
Does every AI action need approval?
No. Gate actions by risk and reversibility; let low-risk steps run with monitoring.
What is human-on-the-loop?
The AI acts on its own while a person monitors and can step in.
How do we avoid approval fatigue?
Approve only meaningful decisions, batch routine items and track override rates.
Is human oversight legally required?
For high-risk uses under the EU AI Act, yes; check rules in your markets.