Contact Lens QA & Compliance Analytics

The problem

Manual QA sees 1-2% of calls — the rest is a blind spot

Supervisors can only manually review a tiny fraction of calls, so most coaching opportunities go unseen and most compliance risk stays invisible until it becomes a complaint, a regulator finding, or a churned customer.

This solution scores every call for sentiment, disclosure and script adherence, and compliance risk — surfacing what needs a human's attention instead of asking humans to find it.

Primary objective

Give every call the scrutiny only 1-2% used to get — without adding QA headcount.

How it works

Capture call → score → flag risk → route to coaching/audit

Capture call

Contact Lens transcribes & analyses

→

Score

Sentiment & script adherence

→

Flag risk

Compliance & disclosure gaps

→

Route

Coaching queue or audit trail

What it does

✓ 100% call coverage scoring
✓ Sentiment & tone tracking
✓ Script & disclosure adherence
✓ Compliance risk flagging
✓ Auto-built coaching queues
✓ Trend dashboards by team

Built with

✓ Amazon Connect Contact Lens
✓ AWS Lambda + EventBridge
✓ Amazon CloudWatch
✓ Amazon DynamoDB
✓ QuickSight-based reporting

Guardrails

✓ Human review of every flag
✓ No automated penalties
✓ PII masking in transcripts
✓ Configurable rule sets

Under the hood

Risk flags land with a supervisor, not a scoreboard

Contact Lens analytics fire an event the moment a call ends. Flagged calls go straight to a supervisor's review queue — the model surfaces risk, a person decides what it means.

connect/contact_lens_flag.py PYTHON

import boto3

connect = boto3.client("connect")
dynamodb = boto3.resource("dynamodb").Table("qa-review-queue")

RISK_TERMS = ["cancel policy", "refund guarantee", "not disclosed"]

def handle_contact_lens_event(event: dict) -> None:
    analytics = connect.get_contact_attributes(
        InstanceId=event["instanceId"], InitialContactId=event["contactId"])
    sentiment = analytics["Attributes"].get("OverallSentiment")
    transcript = analytics["Attributes"].get("Transcript", "")

    # Flag for a human — never auto-score an agent's record
    risk_hits = [t for t in RISK_TERMS if t in transcript.lower()]
    if sentiment == "NEGATIVE" or risk_hits:
        dynamodb.put_item(Item={
            "contact_id": event["contactId"], "sentiment": sentiment,
            "risk_terms": risk_hits, "status": "pending_review"})

Outcomes

100%

Of calls reviewed

Minutes

Not weeks, to surface risk

Fewer

Compliance findings

Faster

Coaching cycles

See every call, not just the sample

Deploy standalone, or pair it with Conversational Voice AI to score bot and human interactions the same way.

Start a conversation →