How to Build an AI Agent with Human-in-the-Loop on Azure

An AI agent with human-in-the-loop on Azure routes proposed high-stakes actions to a human reviewer before they execute. This guide covers all six steps: scoping decision authority, standing up the agent on Azure AI Foundry or Copilot Studio, and wiring a real approval queue in Teams or Power Automate.

What you need before you start

Check the full list of AI implementation guides if you are still deciding on an approach. For this guide specifically, confirm you have the following in place before starting:

• Azure subscription with Contributor or Owner access to the resource group where you will deploy.
• Azure AI Foundry project provisioned in your subscription, or a Microsoft 365 license with Copilot Studio entitlement for the low-code path.
• Azure AI Search instance at Standard tier or above, with your organization's data already indexed.
• Azure AD tenant admin rights to register an app and grant Microsoft Graph consent if the approval surface runs in Teams.
• Power Automate Premium license for the flow owner who manages the Approval connector.
• A classified action list: every action the agent can take, labeled as auto-execute (safe, reversible) or requires-human-approval (irreversible, customer-facing, or financially consequential).
• An audit log target: Azure Monitor workspace or a dedicated database table where every HITL decision will be written.

Step-by-step: Building an AI agent with human-in-the-loop on Azure

1. Scope the agent's task and classify actions by risk. List every action the agent can take and label each one: auto-execute or requires-human-approval. Sending a customer email, moving money, and modifying production records all belong in the second category. This classification becomes your HITL policy and drives every design decision that follows. If you skip this step, you will end up retrofitting checkpoints into a system that was not built to pause.
2. Stand up the agent on Azure AI Foundry or Copilot Studio with grounded search. In Azure AI Foundry, create a new project, select a GPT-4o deployment from the model catalog, and attach your Azure AI Search index as the grounding data source. For the low-code path, open Copilot Studio, create a new copilot, and connect it to the same index using the built-in data source connector. Either approach gives you an agent that retrieves and reasons over your organization's data before proposing any action.
3. Define the approval checkpoint: the agent pauses and routes a draft action to a reviewer. This is the HITL gate. When the agent determines a high-stakes action is required, it composes a structured draft (action type, target entity, supporting reasoning, confidence level) and stops. In Azure AI Foundry, implement this as a tool call that returns a "pending human review" status instead of executing. In Copilot Studio, use a pause-and-escalate topic node that hands off to a Power Automate flow. The draft must contain enough context for a reviewer to make a real decision, not just rubber-stamp it.
4. Wire the human review surface in Teams or Power Automate. In Power Automate, build an Approval connector flow: it receives the agent's draft action, sends it to a named reviewer with approve, edit, and reject options, and returns the decision to the agent. On approval, the flow passes a "proceed" signal back. On rejection, it returns a "cancel with reason" payload the agent can use to explain the outcome to the end user. On edit, it passes the modified action back for execution. Wire all three branches before testing.
5. Log the agent's reasoning, proposed action, and reviewer decision for audit. Write a structured record to Azure Monitor or a dedicated table for every HITL event: timestamp, session ID, proposed action payload, reviewer identity from Azure AD, their decision (approve, edit, or reject), any edits made, and the final action taken. An agent with no audit trail has no accountability, and this log is what you show a compliance team when something goes wrong.
6. Test the rejection and edit paths before going live. Run explicit test cases where the reviewer rejects the action, where the reviewer edits the proposed payload, and where no reviewer responds within your timeout window. Verify the agent handles all three gracefully, that the audit log captures every outcome, and that the end user receives a clear explanation in each case. Testing only the happy path is the most common failure mode before launch.

Choosing between Azure AI Foundry and Copilot Studio for HITL agents

The right platform depends on who builds the agent and how much custom logic the approval workflow requires.

If your team runs on Microsoft 365 and the approval workflow maps cleanly to standard Power Automate connectors, Copilot Studio is the faster path to production. If you need full control over the reasoning loop or your compliance requirements demand a custom audit schema, build on Azure AI Foundry.

Where this gets tricky

The HITL pattern works well for actions that are high-stakes and infrequent enough that waiting on a human is acceptable. Two categories break it.

First: agents that auto-execute irreversible actions without a gate. Moving money, sending customer-facing messages, and modifying production records without any human checkpoint are exactly the failure mode this pattern prevents. If your classification exercise (Step 1) places genuinely irreversible actions in the auto-execute bucket, go back and reclassify. The temptation to approve "just this one" action automatically erodes the entire control framework.

Second: latency-sensitive flows. If the agent must respond in under two seconds, inserting a human review step that takes minutes or hours breaks the user experience. These flows need a different design: either the agent acts within a narrow, pre-approved policy envelope without pausing, or the high-stakes action defers to a background queue while the user receives an immediate acknowledgment. Do not force HITL into a real-time loop where it does not fit.

A third failure mode is a shallow audit log. Reviewers who approve without seeing the agent's full reasoning leave behind a log entry that is useless for debugging or compliance review. Require structured reasoning in the draft payload, not just the proposed action.

How QServices builds HITL agents on Azure

QServices' AI Agent Development practice treats human-in-the-loop governance as a default, built in from the start rather than bolted on after launch. A typical engagement runs 6 to 12 weeks and costs between $15,000 and $85,000, depending on integration complexity and whether compliance documentation is in scope.

The work divides into three phases: HITL policy design (which actions require sign-off and what the review surface looks like), agent build on Azure AI Foundry or Copilot Studio with grounded Azure AI Search, and approval workflow integration in Power Automate or Teams. We also build the evaluation harness before launch, a step most teams skip and regret when the first unhandled rejection path reaches a real user.

Our Smart PM Assistant for an IT services company automated meeting transcript capture and Azure DevOps backlog creation, with humans keeping sign-off authority over sprint commitment decisions. For a detailed cost breakdown, see our AI Agent Development cost page.

Do I need a separate approval system, or does Power Automate cover it?

Power Automate's Approval connector covers most internal HITL scenarios: it sends the proposed action to a named reviewer, captures their response, and returns the decision to the calling flow. For more complex cases, including multiple approval tiers, time-limited response windows, or reviewers outside your Azure AD tenant, you will need a custom queue backed by Azure Service Bus or a dedicated workflow tool.