Implementing Speech and Text Analytics Integration with Quality Evaluation Trigger Policies

How-To-Articles
1

Implementing Speech and Text Analytics Integration with Quality Evaluation Trigger Policies

What This Guide Covers

This guide details the architectural implementation of automated Quality Management (QM) workflows in Genesys Cloud CX. You will configure Speech and Text Analytics to generate quality events that trigger specific Quality Evaluation Forms based on content analysis, sentiment thresholds, or keyword detection. The end result is a system where high-risk or non-compliant interactions automatically route to specialized QA forms, while standard interactions route to generic forms, eliminating manual tagging errors and ensuring consistent compliance auditing.

Prerequisites, Roles & Licensing

  • Licensing:
    • Genesys Cloud CX 1, 2, or 3 license.
    • Speech and Text Analytics Add-on (required for transcription and analytics processing).
    • Quality Management Add-on (required for creating and assigning evaluations).
  • Roles & Permissions:
    • Administrator role with Quality > Evaluation Form > Edit and Quality > Policy > Edit permissions.
    • Speech and Text Analytics Administrator role with Speech and Text Analytics > Model > Edit and Speech and Text Analytics > Topic > Edit permissions.
    • Architect role with Architect > Flow > Edit permissions (if using flow-based triggers).
  • External Dependencies:
    • A configured Speech and Text Analytics model with trained topics or keywords.
    • Pre-built Quality Evaluation Forms mapped to specific business processes.

The Implementation Deep-Dive

1. Establishing the Analytics Foundation: Topics and Keywords

Before Quality Management can react to interaction content, the Speech and Text Analytics engine must be able to identify that content. The most common failure mode in this integration is relying on generic, out-of-the-box topics. Generic topics lack the specificity required for compliance triggers, leading to high false-positive rates in automated evaluations.

You must create custom Topics or Keywords that map directly to your QA criteria. For this example, we assume a financial services context where we need to detect “Mandatory Disclosure Failure” and “Hostile Agent Sentiment.”

The Trap: Configuring a Topic with broad keywords like “money” or “problem.” This causes the analytics engine to flag nearly every interaction, overwhelming the QA queue with low-value evaluations. The downstream effect is QA fatigue, where agents ignore automated flags because they perceive them as noise.

Architectural Reasoning: We use Topics for semantic understanding (e.g., “Did the agent explain the fee structure?”) and Keywords for exact compliance phrase matching (e.g., “Your right to cancel…”). Topics require a training period of 100-500 interactions to achieve high confidence scores. Do not rely on untrained topics for automated triggering.

Configuration Steps:

  1. Navigate to Admin > Speech and Text Analytics > Models.
  2. Select your active model and go to the Topics tab.
  3. Create a new Topic named Compliance_Disclosure_Failure.
  4. Add example phrases that represent missing information or incorrect explanations. Note that Genesys Cloud uses a positive classification model; you are training it to identify the presence of specific concepts. To detect failures, you often need a separate topic for “Successful Disclosure” and invert the logic in the trigger, or train a specific topic for “Apology for Error” which often follows a failure.
  5. Create a second Topic named Agent_Hostility. Train this with phrases indicating sarcasm, interruption, or aggressive tone.

API Reference for Topic Creation:
If you are scripting this deployment, use the following endpoint.

POST /api/v2/analytics/models/{modelId}/topics
Authorization: Bearer {access_token}
Content-Type: application/json

{
  "name": "Compliance_Disclosure_Failure",
  "description": "Identifies instances where mandatory fee disclosures are omitted.",
  "active": true,
  "type": "topic"
}

2. Configuring Quality Evaluation Forms

The Quality Evaluation Form is the payload that gets assigned to an agent. You need at least two forms: a Standard_QA_Form and a Compliance_QA_Form. The Compliance_QA_Form should contain rigorous, pass/fail criteria related to the topics identified above.

The Trap: Using a single, massive evaluation form with conditional sections. While Genesys Cloud supports conditional logic in forms, it is computationally heavier and harder to maintain than separate forms. More critically, if you use a single form, you cannot easily segment reporting by “Compliance Risk” versus “General Service.”

Architectural Reasoning: Separate forms allow for different scoring weights. A compliance failure might be a zero-tolerance event, whereas a service gap might be a 10% deduction. By separating the forms, you can assign different reviewers (e.g., Compliance Officers vs. QA Managers) based on the form type.

Configuration Steps:

  1. Navigate to Admin > Quality > Evaluation Forms.
  2. Create Standard_QA_Form. Include general courtesy, hold time, and resolution questions.
  3. Create Compliance_QA_Form. Include specific questions: “Did the agent read the mandatory disclosure verbatim?” and “Was the tone professional throughout the hostile customer interaction?”
  4. Note the Evaluation Form IDs for both. You will need these for the Policy configuration.

3. Creating the Quality Policy with Analytics Triggers

This is the core integration point. A Quality Policy defines when an evaluation is created and which form is used. We will create a policy that listens for Speech and Text Analytics events.

The Trap: Setting the trigger threshold too low. For example, triggering a compliance evaluation if a Topic confidence score is > 0.5. At 0.5 confidence, the model is essentially guessing. This results in evaluations being generated for interactions where the topic was not actually present, leading to agents being penalized for errors they did not make. Always require a confidence score of at least 0.8 or 0.9 for automated triggers.

Architectural Reasoning: We use Event-Based Triggers within the Quality Policy. Genesys Cloud processes these triggers asynchronously after the interaction is transcribed. This decouples the real-time telephony path from the QA assignment path, ensuring that QA assignments do not impact call setup times or network latency.

Configuration Steps:

  1. Navigate to Admin > Quality > Policies.
  2. Click Add Policy. Name it Auto_QA_Compliance_Trigger.
  3. In the Trigger section, select Speech and Text Analytics.
  4. Configure the trigger conditions:
    • Condition: Topic
    • Operator: Is In
    • Value: Select Compliance_Disclosure_Failure AND Agent_Hostility.
    • Confidence Threshold: Set to 0.85.
  5. In the Action section:
    • Evaluation Form: Select Compliance_QA_Form.
    • Reviewer: Assign to a specific group, e.g., Compliance_Team.
    • Deadline: Set to 7 days (standard SLA).
  6. Save the Policy.

Alternative Approach: Using Architect for Complex Logic
If your logic requires multiple analytics results (e.g., “If Topic A is present AND Sentiment is Negative”), the native Quality Policy UI may be limited. In this case, use an Architect Flow.

  1. Create a new Interaction Flow.
  2. Add a Trigger for Speech and Text Analytics Event.
  3. Use a Split node to evaluate the analytics payload.
  4. Use the Create Quality Evaluation block in Architect. This block allows you to dynamically select the evaluation form based on complex business logic.

API Reference for Policy Creation:

POST /api/v2/quality/policies
Authorization: Bearer {access_token}
Content-Type: application/json

{
  "name": "Auto_QA_Compliance_Trigger",
  "description": "Triggers compliance QA for high-risk topics",
  "enabled": true,
  "trigger": {
    "type": "speechAndTextAnalytics",
    "conditions": [
      {
        "field": "topicName",
        "operator": "in",
        "values": ["Compliance_Disclosure_Failure", "Agent_Hostility"]
      }
    ],
    "confidenceThreshold": 0.85
  },
  "actions": [
    {
      "type": "createEvaluation",
      "evaluationFormId": "{compliance_form_id}",
      "reviewerGroupId": "{compliance_team_id}",
      "deadlineOffset": "P7D"
    }
  ]
}

4. Mapping Analytics Results to QA Scores (Optional Advanced Integration)

For a fully automated QA process, you can bypass the human reviewer entirely for low-risk items or use analytics scores to pre-fill QA scores. This requires enabling Automated Scoring in the Quality settings.

The Trap: Fully automating scoring for compliance-critical items. Regulatory bodies (PCI-DSS, HIPAA, FCC) often require human verification of compliance checks. Automating a “Pass” for a mandatory disclosure based solely on a keyword match can lead to audit failures. Use automated scoring only for soft skills (e.g., Tone, Pace) or as a pre-screening step for human reviewers.

Architectural Reasoning: We use Analytics Score Mappings to translate a 0-1 confidence score into a 0-100 QA score. This requires a linear mapping configuration.

Configuration Steps:

  1. Navigate to Admin > Quality > Settings.
  2. Enable Automated Scoring.
  3. Map the Agent_Hostility topic to a QA question “Professionalism”.
  4. Set the mapping:
    • If Topic Confidence < 0.1, Score = 100 (No hostility detected).
    • If Topic Confidence > 0.9, Score = 0 (High hostility detected).
  5. This allows the QA dashboard to show a preliminary score before human review.

Validation, Edge Cases & Troubleshooting

Edge Case 1: The “Late Transcription” Gap

The Failure Condition: An agent completes a call at 5:00 PM. The QA evaluation is expected to be assigned by 5:05 PM. However, the evaluation does not appear until 6:30 PM. The QA manager reports that the automation is broken.

The Root Cause: Speech and Text Analytics processing is not instantaneous. It depends on the queue depth of the transcription service, the length of the call, and the complexity of the model. For long calls (>15 minutes), transcription can take 10-20 minutes. The Quality Policy only triggers after the analytics event is published. If the policy has a strict SLA of “24 hours,” a 20-minute delay is acceptable. If the SLA is “1 hour,” this delay causes SLA breaches.

The Solution:

  1. Adjust the Deadline in the Quality Policy to account for transcription latency. Add a buffer (e.g., set deadline to P1D instead of PT4H).
  2. Implement a Monitoring Dashboard that tracks the delta between Interaction_End_Time and QA_Assignment_Time.
  3. Consider using Real-Time Analytics (if licensed) for critical compliance checks, which can interrupt the call in progress, rather than post-call QA.

Edge Case 2: False Positives from Background Noise

The Failure Condition: The Compliance_Disclosure_Failure topic is triggered frequently, but upon review, the agent did read the disclosure. The analytics engine missed it due to background noise or a poor microphone.

The Root Cause: The transcription accuracy is low, causing the semantic model to fail to recognize the keywords. The model is confident that the topic is not present, but if your trigger is based on the absence of a “Success” topic, it fails. Alternatively, if the noise creates false keywords, it triggers the topic incorrectly.

The Solution:

  1. Filter by Audio Quality: In the Quality Policy trigger, add a condition for Audio Quality Score. Only trigger QA if the Audio Quality Score is > 70. This prevents QA assignments for calls with unintelligible audio.
  2. Human-in-the-Loop Review: Configure the QA assignment to route to a “Review Queue” rather than directly to an agent. A QA supervisor verifies the analytics flag before assigning the evaluation to the agent.
  3. Retrain the Model: Use the false positive examples to retrain the topic. Mark the interactions as “Not This Topic” in the Analytics UI to improve model precision.

Edge Case 3: Cross-Channel Consistency

The Failure Condition: A customer starts a chat, then calls. The chat is flagged for Compliance_Disclosure_Failure, but the voice call is not. The QA system creates two separate evaluations, but the agent is penalized twice for the same underlying issue.

The Root Cause: Genesys Cloud treats each channel interaction as a separate entity. The Quality Policy triggers independently for each channel. Without cross-channel correlation, the system does not know that the voice call is a continuation of the chat.

The Solution:

  1. Use Interaction Links in Architect. When the chat converts to a voice call, link the interactions.
  2. Create a Custom Attribute on the interaction that propagates the analytics result from the chat to the voice call.
  3. Modify the Quality Policy to check for this custom attribute. If the attribute indicates a compliance failure in the prior channel, suppress the duplicate evaluation or merge the QA results.

Official References