Human-in-the-Loop Workflow: Where to Put Approval Gates for AI Agents
By Kirill Stolbushkin
Short answer: A human-in-the-loop workflow is a process where an AI agent does the routine work and a person approves, edits or rejects at chosen steps. Put approval gates where a mistake is expensive or hard to undo: money above a limit, anything sent outside the company, deletions, policy exceptions and low-confidence cases. Leave the rest to the agent. Too many gates and people stop reading; too few and you hear about mistakes from customers.
What is a human-in-the-loop workflow?
A human-in-the-loop workflow is a business process where an AI agent handles the steps it can do reliably and pauses for a person at the steps that need judgment or accountability.
I'm Kirill, a Vevos cofounder, and approval gates are something we've spent a lot of this year building, so I'll admit the bias up front. Everything below works on a whiteboard too. To sketch your own version from a sentence or two, the free process mapper runs without an account.
The person isn't there to redo the agent's work. They're there to make the decision the company wants a human to own. There are three common setups, and they line up with the three levels in Singapore IMDA's framework for agentic AI:
Setup | What the person does | IMDA level | Good fit |
|---|---|---|---|
Human in the loop | Approves before the agent acts, at every step or at chosen gates | Agent proposes, human operates; or agent operates, human approves | Money, customers, irreversible actions |
Human on the loop | Watches the agent work and can step in or stop it | Agent operates, human observes | Routine work that's easy to reverse |
Human out of the loop | Not involved while it runs; reviews results later | Not one of the three IMDA levels | Only very low-risk, fully rule-based work |
Most teams should start with a person approving every action on a new agent, then move one process at a time to gates at critical steps only.
Human in the loop AI examples
Human in the loop shows up wherever an agent prepares the work and a person owns the decision. Here are five common ones.
Process | What the agent does | Where the gate is | Approver |
|---|---|---|---|
Refunds | Checks the order, the policy and past refunds, then drafts the refund | Before any refund above the agent's limit | Support lead |
Vendor onboarding | Collects documents, checks the tax form and restricted list | Before the vendor is set up for payment | Procurement lead |
Contract review | Compares the contract with the standard terms and flags changes | Before any non-standard clause is accepted | Legal counsel |
Content moderation | Scores posts and hides clear spam | Before removing anything borderline or banning an account | Trust and safety reviewer |
Hiring screen | Summarizes applications against the posted criteria | Before any candidate is rejected or advanced | Hiring manager |
Notice the pattern: the agent does the reading and sorting, the person makes the call that affects money, rights or reputation. The hiring row matters most. Using AI in employment decisions is one of the EU AI Act's high-risk areas, so keep a person on every decision there.
Where to put approval gates
Put an approval gate at any step where a wrong action costs real money, reaches a customer, or can't be undone.
Check these five places first:
Money above a limit. Payments, refunds, credits and purchases above the agent's authority limit.
Anything that leaves the building. Emails to customers, messages to vendors, filings and public posts, at least until the agent has a track record.
Irreversible actions. Deleting records, closing accounts, cancelling orders.
Policy exceptions. Anything the process says is outside the normal rules.
Low confidence or missing data. When the agent can't find what it needs or the inputs conflict.
IMDA's framework makes the same point: define significant checkpoints that need human approval, such as high-stakes or irreversible actions. In the EU, Article 14 of the AI Act requires high-risk AI systems to be built so people can effectively oversee them while they're in use; after the 2026 AI Omnibus amendment, those high-risk rules apply from 2 December 2027 for most stand-alone use cases.
Where not to put approval gates
Don't put a gate on a step where the reviewer can't add anything, because a gate nobody reads is worse than no gate.
If a person approves 200 routine items a day, they'll start clicking approve without looking. That's automation bias, and IMDA calls it out as a growing concern as agents get more capable. Skip gates on steps that are cheap to reverse, fully rule-based, or already checked later in the process. Spend people's attention where it matters.
How to design an approval gate
A good approval gate tells the approver exactly what they're deciding, gives them the evidence, and says what happens if they don't answer.
Design question | Example answer |
|---|---|
Who approves? | The AP lead |
Who's the backup? | The AP manager, after one business day |
What's the final escalation? | The controller, after a second business day |
What do they see? | The invoice, the purchase order, the price difference and the agent's recommendation |
What can they do? | Approve, edit the amount, reject with a reason, or send back for more information |
What if nobody answers at all? | It stays in the controller's queue. It never auto-approves |
What happens after a rejection? | The agent notifies the requester and closes the request with the reason |
The non-response rows are the ones people forget. Decide them before the first item waits.
How pause and resume works
When an agent reaches an approval gate, it stops, saves where it is, asks a person, and only continues once that person decides.
In a well-built workflow the sequence looks like this. The agent finishes the steps before the gate and packages what the approver needs: the request, the evidence and its recommendation. The workflow saves its state, meaning the data so far and the exact step it stopped on, so nothing is lost if the wait takes an hour or a week. The approver is notified through whatever channel your team uses. While it waits, the agent does nothing else on that item.
When the decision comes in, the workflow checks it's from someone allowed to make it. If approved, the agent resumes from the saved step with the approved values. If rejected, it takes the rejection path and stops. If the deadline passes, the item moves to the backup approver instead of continuing. Each item resumes once, so a double click can't pay an invoice twice.
Human-in-the-loop workflow diagram (BPMN)
In BPMN, an agent's work is a task, the decision to ask a person is an exclusive gateway, the person's approval is a user task, and the deadline is a timer boundary event.
That's enough to draw almost any human-in-the-loop workflow, including agentic ones. BPMN doesn't care whether a task is done by software or a model; the diagram shows what happens, in what order, under which rules. Use lanes to show who does what: one for the agent, one for each approver role. Our post on swimlane diagrams covers lanes, and the BPMN symbols cheat sheet covers every shape.
The same diagram does two jobs. People read it to understand the process, and agents follow it as the approved version.
Example: an invoice price exception
Here's a human-in-the-loop workflow for one common AP problem: the invoice price doesn't match the purchase order. In BPMN, the agent steps sit in the agent lane and step 7 sits in the approver lanes.
Start: a vendor invoice arrives.
Agent: reads the invoice and finds the matching purchase order.
Agent: compares unit prices and quantities.
Gateway: does the invoice match within 2%?
Agent, if yes: queues the invoice for payment under its authority limit.
Agent, if no: prepares a summary of the difference and a recommendation, then pauses.
AP lead: approves, edits or rejects. Timer: after one business day it goes to the AP manager, and after a second day to the controller.
Gateway: approved or rejected?
Agent, if approved: resumes and queues the invoice with the approved amount.
Agent, if rejected: drafts a note to the vendor for the AP lead to send, and closes the item.
End.
One approval gate, placed at the point of consequence, with two escalation levels. The agent does the matching, which is most of the work. For the full AP picture, see our invoice approval workflow guide.
How to measure whether your gates work
Measure four numbers per gate: how often it fires, how long items wait, how often people change the agent's recommendation, and how often they approve without opening the evidence.
Signal | What it tells you |
|---|---|
Gate rate | Share of items that stop for a person. Very high means the limit may be too low |
Wait time | How long items sit. Long waits mean the wrong approver or no backup |
Override rate | How often people change or reject the recommendation. Near zero for months may mean you can raise the limit |
Blind approvals | Approvals without opening the evidence. A rising count means the gate is becoming decoration |
Review these monthly with the process owner and move limits on evidence, not on feel.
Human-in-the-loop workflows in Vevos
In Vevos, approval gates and authority limits sit on the approved playbook, so every agent that follows the playbook stops at the same steps.
You start by describing the process or uploading an SOP or a recording, and Vevos drafts the BPMN 2.0 map and its documentation. On that map you mark which steps are approval gates and who approves them, and you set the agent's authority limits. The process owner signs off on the version before any agent uses it, and version history with a diff shows what moved when you change a limit later. Outside AI agents reach the same playbook through Unison over MCP, so they hit the same gates. If something looks wrong, the stop switch halts the agent.
If you also want an app around the process, Conductor Agents can build and deploy one from a reviewed map, starting on the Tempo plan; see pricing.
One limit worth saying plainly: a gate only works if the right person gets it and has time to read it. No software fixes a bad choice of approver.
FAQ
Short answers to common questions about human-in-the-loop workflows.
What's the difference between human in the loop and human on the loop?
Human in the loop means a person approves before the agent acts at defined steps. Human on the loop means the agent acts on its own while a person watches and can step in or stop it. Use in the loop for risky steps and on the loop for routine, reversible work.
What are examples of human in the loop AI?
Refunds above a limit, vendor setup before payment, non-standard contract clauses, borderline content removals and hiring decisions. In each, the agent prepares the work and a person makes the final call.
How do you build human in the loop into an agentic workflow?
Map the process, mark the steps where a wrong action is costly, and make those user tasks with a named approver, a deadline and a backup. Make sure the workflow saves its state while it waits and resumes only on an authorized decision.
How many approval gates should a workflow have?
As few as cover the real risks. Most processes need one to three: one at the money limit, one before anything goes outside the company, and one for exceptions.
Should an approval ever auto-approve when nobody responds?
Not for high-stakes steps. Send it to a backup approver and then to a manager's queue. Auto-approving on timeout quietly removes the human from the loop.
Related reading
These guides cover the rest of the picture.
AI agent governance framework: owners, limits, gates and a register
AI automation playbook: map, document and automate step by step
Draw your first gate
The quickest way to start is to map one process and mark where a person must decide.
Describe it at vevos.ai/free to get a BPMN map and documentation draft, and check it with our guide on how to review an AI-generated process model. When you're ready to put gates in front of your agents, see plans and pricing.
Related blog posts
- Approval Matrix Template: Thresholds, Approvers, and Authority Limits for AI Agents — A copyable approval matrix template for finance and operations: who approves what at which amount, with backups and deadlines, step-by-step setup in Excel or Google Sheets, examples by process, and a column for what an AI agent may do on its own.
- AI Automation Playbook: How to Map, Document, and Automate Your First Processes — A practical AI automation playbook for operations teams: eight steps from listing processes to measuring ROI, a scored sample shortlist, a 30/60/90-day rollout, a worked vendor setup example and the mistakes to avoid.
- AI Agent Governance Framework: How to Govern AI Agents With Approved Processes — A practical framework for governing AI agents at work: six parts, the main agent risks, a technical controls layer, a copyable agent register, a crosswalk to IMDA, NIST and ISO/IEC 42001, and the metrics that show it works.