Human Approval – Arahi AI

What is Human Approval?

Human Approval defines how and when your Agent hands off work—or alerts human team-members—when it can’t confidently proceed by itself.
This ensures that tasks don’t fail silently and human oversight is in place for high-risk or ambiguous situations.

Why Use Human Approval

  • Real-time alerts: When your Agent hits a stuck point, your team gets notified via email, Slack or another channel.

  • Seamless hand-off: Even when the Agent requests approval, the workflow can feel smooth to the end-user.

  • Control & oversight: Define exactly what conditions require approval and how many retries the Agent attempts before asking for it.

Configuring Human Approval in Arahi AI

  1. In the Agent Builder, go to the Human Approval section.

  2. Define the criteria for approval:

    • If a Tool/sub-agent fails a defined number of times.

    • If the Agent’s confidence is below a threshold.

    • If the task involves high risk / high value / VIP customer.

  3. Select the approval channel: email, Slack, Microsoft Teams, internal dashboard.

  4. Choose approval behaviour:

    • Auto-Run: the approval request is sent automatically when criteria are met.

    • Approval Required: the Agent must receive an explicit human “Approve/Reject” before proceeding.

    • Let Agent Decide: the Agent determines whether to request approval or proceed.

  5. Configure retry logic:

    • Enable automatic retries for transient tool failures.

    • Set maximum retries count.

    • Define what happens after max retries: terminate task or request approval.

Visual Workflow

Best Practices

  • Define clear approval thresholds — what counts as “Agent can’t handle it”?

  • Use contextual requests — include relevant input, Agent reasoning, and error logs so humans can act quickly.

  • Monitor approval volume — if you have too many approval stops, the Agent’s scope may be too broad or the prompts/tools unclear.

  • Balance autonomy with oversight — start with more approvals and gradually increase autonomy as the Agent proves reliable.

  • Document approval paths — who is notified, what channel, expected response time (SLAs).

  • Use approval history as learning data — review patterns and update Agent behaviour/prompts accordingly.

Example Human Approval Setup

Agent Name: CustomerSupport-Responder

  • Criteria: If sentiment analysis tool returns “very negative” or Agent confidence < 0.6

  • Behaviour:

    • First, attempt automatic resolution via reply tool.

    • If still unresolved, request Human Approval after 2 attempts.

    • Approval channel: Slack #support-approvals and email to support-manager@company.com

    • Request includes: original message, agent reasoning, reason for approval

  • Post-approval: Human reviews, resolves, tags task “human-approved” and feedback goes to Agent memory for learning.

Updated on
Triggers
Variables