ASAPP’s Human-in-the-Loop Agents (HILA) capability extends the power of GenerativeAgent by introducing human judgment at critical moments in customer interaction. GenerativeAgent is built to resolve most customer queries on its own. However, when it encounters situations it can’t handle—ambiguous requests, compliance reviews, or actions beyond its automation scope—it can request targeted assistance from a HILA:
  1. In your task configuration, you specify when GenerativeAgent should request assistance.
  2. GenerativeAgent requests an expert human agent, who provides guidance, information, or performs an action on its behalf.
  3. GenerativeAgent directly responds to the customer without disrupting the customer experience.
This hybrid model of automation and supervision enables:
  • Scalable support: One HILA can assist multiple conversations simultaneously.
  • Continuous learning: Each interaction provides information to improve the system.
  • Reduced escalations: Lightweight human input reduces the need for full escalations.
  • Seamless customer experience: The customer experience remains uninterrupted.

​
Approver Mode

HILA can also be used in Approver Mode, where an additional oversight layer is needed, such as during sensitive rollouts. In this mode, every message from GenerativeAgent is reviewed by the HILA agent before it is sent to the customer.

Approver Mode

Learn how human agents review and refine generated responses to ensure safe, accurate, and on-brand customer interactions.

​
When to Use HILA

Consider enabling HILA in your GenerativeAgent deployment when you need to:
  • Handle unknown, ambiguous, or complex cases: When GenerativeAgent doesn’t have enough information or encounters requests outside its Knowledge Base Content or API Sources.
  • Meet compliance or regulatory requirements: For scenarios that require human review or approval before proceeding, such as financial transactions or sensitive data handling.
  • Perform actions not available via API: When certain tasks can only be completed by a human agent.
  • Minimize full escalations: Use HILA to resolve issues with lightweight human input and reduce the need to transfer entire conversations to live agents.

​
How HILA Works

HILA combines GenerativeAgent with your Agent Desk Platform.
  1. Trigger: GenerativeAgent encounters a situation, defined in your configuration, that requires human input (e.g., missing API, sensitive action).
  2. Ticket Creation: GenerativeAgent generates a structured help request and sends it to the digital support platform.
  3. HILA Assignment: The request is routed to a dedicated HILA queue.
  4. Agent Action: HILAs review the context and provide necessary input (answers, corrections, approvals).
  5. Conversation: GenerativeAgent uses the input to proceed with the customer conversation.
  6. Escalation (if needed): HILAs can recommend escalating to a live agent.
HILA Flow

​
Set Up HILA

To set up HILA, you need to integrate the flow into your Agent Desk Platform and configure GenerativeAgent to initiate the HILA request:
1

Integrate HILA with your Agent Desk Platform

ASAPP offers seamless integration into your existing agent desk through an iframe-based HILA application.HILAs operate from dedicated queues within your platform, receiving GenerativeAgent consultation requests directly. This ensures they work within a familiar interface, while you retain full control over routing and assignment.
To set up HILA in your environment, contact your ASAPP Implementation Manager.
2

Configure Ticket Assignment and Routing

HILAs are assigned tickets based on the routing and assignment configurations in your support platform.If no HILA is assigned within the configurable threshold (default: 60 seconds), the case is automatically escalated to a live agent.
3

Define When to Request Human Input

In your GenerativeAgent task configuration, define when human input is required.You can also have GenerativeAgent populate parameters by including them in the function call.To instruct GenerativeAgent to call HILA, use the APICALL expert_assistance() command in your task instructions. For example:
- If the account status is valid, consult an expert by using the command `APICALL expert_assistance(request_type='OFFER_NEEDED', question: str=)` for promotional offers.
  - In your question, make sure to describe why the customer is interested in promotions, as well as the product or service they are interested in.
In this example, GenerativeAgent creates a ticket to check for promotional offers once it has confirmed the account validity. There are two required parameters:
  • request_type: Classifies the consultation reason (any string).
  • question: The relevant question for HILA.
4

Try out HILA

With HILA configured, you can try it out by having a conversation with GenerativeAgent that uses HILA.You can try out HILA by using the Previewer, which enables you to iterate on the task configuration while developing the integration between your Agent Desk Platform and ASAPP.

​
Administrator Controls

There are additional controls you can configure to further customize the HILA experience:
  • Response Timeout: If no response is sent within a configurable timeout (default: 300 seconds), the customer is escalated to a live agent.
  • Agent Assignment Timeout: If no HILA is assigned within a configurable timeout (default: 60 seconds), the conversation escalates to a live agent.
  • Escalation Message: When HILA initiates an escalation, a configurable message is sent to the customer, ensuring a smooth and transparent transition.

​
HILA Workflow

Once the HILA capability has been set up, your agents will start responding to HILA requests:
1

Agent Receives HILA Request

The agent receives a request to review the conversation with GenerativeAgent.This is determined by the routing and queue configuration you set up in your Agent Desk Platform.
2

Agent Reviews HILA Request

HILAs are presented with key information to help them answer the GenerativeAgent query.
Screenshot of HILA Full UI
The HILA application includes:
  1. Transcript of the customer interaction
  2. Summarized context and key intents
  3. Customer information (if authenticated)
  4. Real-time conversation thread
  5. Assignment timer tracking ticket duration
3

Agent Responds to the HILA Request

The agent can respond to the HILA request in one of three ways:
Screenshot focused on Agent Response
  1. Provide a response to the customer (closing the ticket)
  2. Request more information from the customer (keeping the ticket open)
  3. Escalate the conversation to a live agent if needed by clicking the transfer button in the top right corner of the transcript panel.
4

GenerativeAgent Continues the Conversation

GenerativeAgent uses the input from HILA to proceed with the customer conversation, maintaining a seamless experience.If the agent responds with “Info needed”, GenerativeAgent will continue the conversation by requesting the additional information from the customer.The HILA application can also send real-time notifications to ensure that HILA doesn’t miss any new messages from GenerativeAgent.
Screenshot of HILA Multi-turn

​
Next Steps

After setting up Human-in-the-Loop, you are ready to speed up customer replies and resolve their inquiries. You may find one of the following sections helpful in advancing your integration:

Approver Mode

Learn how human agents review and refine generated responses to ensure safe, accurate, and on-brand customer interactions.

Configure GenerativeAgent

Learn how to configure GenerativeAgent for your use case.

Task and Function Improvements

Learn how to improve GenerativeAgent’s tasks and functions.

Go Live

Follow steps and best practices to launch your GenerativeAgent deployment in production.