Conversational analytics — the capability formerly known as Contact Lens — analyses conversations between customers and agents, or customers and conversational AI, across voice, chat, and email, using natural language processing. It performs sentiment analysis, detects issues, automatically categorises contacts, transcribes and redacts recordings, and surfaces a set of conversation metrics.[1][2]
The value isn't the feature being on — it's knowing what to check for. Below are the signals worth watching, including the ones you asked about (and an honest answer on the "is this even possible?" ones), plus the wider behavioural signals that help you shape the experience rather than just report on it.
Call length & handle time
Average handle time is the obvious one, but the real insight is length by contact driver — the conversational analytics dashboard shows average handle time per call driver (e.g. "Where's my stuff?"), so you can see which journeys run long.[3]
Why it matters: consistently long calls for a specific driver flag a complex or broken journey — a candidate for better self-service, clearer flows, or agent guidance.
Agent vs customer talk time
The talk-time graph shows the distribution of talk vs non-talk time across the call, with talk time split into agent and customer.[1] There's also an "agent talk time percent" metric (agent talk time as a percentage of total conversation).[4]
Why it matters: an agent dominating the conversation can mean over-scripting or not listening; a customer dominating can mean confusion or venting. The healthy balance varies by call type — watch for outliers.
Sentiment — and the sentiment trend
Conversational analytics scores customer sentiment as Positive, Neutral, or Negative, both as an overall distribution and — more usefully — as a trend across the call, showing how sentiment changes as the contact progresses.[1]
Why it matters: the shape beats the average. A call that starts negative and ends positive is a recovery win; one that starts fine and ends negative tells you exactly where the experience broke. Real-time sentiment can even alert a supervisor to join a call that's turning negative.[1]
Outcomes & issues
Conversational analytics detects issues and lets you automatically categorise contacts using rules, so you can group interactions by outcome and spot trends.[1][9]
Why it matters: categorisation turns thousands of unstructured conversations into countable outcomes — resolved, escalated, repeat contact, complaint — so you can measure what actually happened, not just how long it took.
"Tone of voice" — what's actually possible
Honest answer: there's no single "tone" or "emotion" label. But conversational analytics gives you the signals that together indicate tone: sentiment and its trend, interruptions (how often one party talks over the other),[2] and talk speed (words per minute — comprehension suffers if someone speaks too fast).[4a]
Why it matters: read sentiment + interruptions + talk speed + non-talk time together and you get a reliable picture of how an interaction felt — a rushed agent talking fast and interrupting a negative-sentiment customer is a coaching moment, even without a literal "tone" field.
Hold time & non-talk time
Yes, this is possible — via non-talk time, which is the combined duration of hold time plus any silence where neither participant speaks for more than 3 seconds (this threshold can't be customised).[5][7] It appears in the talk-time distribution.[1]
Why it matters: dead air and long holds are a direct driver of frustration. High non-talk time often points to agents hunting for information — a signal to improve knowledge access, agent assist, or screen-pop so they're not leaving the customer in silence.
The questions customers ask
The dashboard surfaces why customers are contacting you (contact drivers) and the trend of those drivers over time, and intelligent search lets you find conversations by content.[3] That's your raw material for understanding the real questions being asked.
Why it matters: this is gold for enriching self-service. The questions customers actually ask — in their words — should feed your knowledge base, bot utterances, and Q&A content, so more of them get answered without an agent. It also pairs directly with building a production-grounded Lex test set and eval set.
More Behavioural Signals Worth Watching
Beyond your seven, conversational analytics exposes several more behavioural signals — and reading behaviour (not just outcomes) is what lets you proactively shape the experience.
| Signal | What it tells you |
|---|---|
| Interruptions | How often the agent talks over the customer — a listening/empathy and compliance signal[2] |
| Talk speed (words per minute) | Whether an agent is rushing; fast talk speed hurts comprehension, especially with distressed customers[4a] |
| Agent greeting time (chat) | How quickly the agent engages after joining a chat — a slow greeting can seed early negative sentiment[1][4] |
| Agent & customer response times (chat) | Lets you compare the experience of the bot vs the agent and find where responsiveness drops[1][8] |
| Themes / contact-driver trends | What's emerging across many conversations over time — spot a spike in a new issue early[3] |
| Compliance keywords/phrases | Whether required phrases were said (or risky ones), via categorisation rules and real-time alerts[10] |
| Bot vs agent experience | For chat, investigate the customer's interaction with a bot versus a human to find self-service gaps[1] |
| Contact summaries | Generative summaries of what happened, so trends can be read without listening to every call[1] |
Privacy note: conversational analytics can redact sensitive data (name, address, card details) from transcripts and recordings.[1] Turn redaction on — behavioural insight shouldn't come at the cost of storing PII you don't need.
Turning Signals Into Experience Changes
The point of all this isn't a prettier dashboard — it's action. Reading behaviour tells you what to change:
- Long calls for one driver → simplify that journey, add self-service, or give agents a step-by-step guide.
- Sentiment dropping at a specific point → fix what happens there (a confusing step, a long hold, a bad handoff).
- High non-talk time → improve knowledge access and agent assist so agents aren't leaving dead air.
- Agent interrupting / talking too fast on negative calls → targeted coaching, not blanket training.
- The real questions customers ask → feed them into your knowledge base, bot, and Lex optimisation so containment rises.
- Emerging themes → get ahead of a developing issue before it floods the queue.
Make it a loop. Behavioural analytics is most powerful as a continuous cycle: observe the signals, change something specific, and watch the signals move. That's the same self-improvement discipline behind great self-service — see the self-service flywheel.
Frequently Asked Questions
What is Amazon Connect conversational analytics?
Conversational analytics (formerly Contact Lens) is the Amazon Connect capability that analyses conversations between customers and agents or conversational AI, across voice, chat, and email, using natural language processing. It performs sentiment analysis, detects issues, automatically categorises contacts, transcribes and redacts recordings, and surfaces metrics like talk time, non-talk time, interruptions, talk speed, and response times.
Can conversational analytics detect tone of voice?
Not as a literal "tone" or emotion label, but it detects the signals that together indicate tone: sentiment (positive, neutral, negative) and how it trends across the call, interruptions, talk speed in words per minute, and non-talk time. Reading those together gives you a reliable read on how the interaction felt without a single "tone" field.
Does conversational analytics measure hold time?
Yes, via non-talk time. Non-talk time is the combined duration of hold time plus any silence where neither participant speaks for more than 3 seconds. It is shown in the talk-time distribution for a call and helps you spot dead air and long holds that hurt the customer experience.
How does conversational analytics help improve the customer experience?
By revealing behaviour, not just outcomes. You can see why customers contact you (contact drivers/themes), how sentiment moves through a call, where agents talk over customers, where dead air and long holds occur, which questions get asked (to enrich self-service and Q&A content), and which journeys run long or end badly — all of which point to specific changes that shape a better experience.
Conclusion
Amazon Connect conversational analytics is only as useful as the questions you bring to it. Check call length by driver, the agent/customer talk-time balance, sentiment and its trend, outcomes via categorisation, the tone signals (sentiment + interruptions + talk speed + non-talk time), hold/non-talk time, and the real questions customers ask — then widen out to interruptions, response times, themes, and compliance. Most importantly, treat every signal as a prompt to change something specific. Behaviour you can see is behaviour you can shape.
Sources
- AWS, Analyze conversations using conversational analytics (sentiment, trend, talk time, redaction, chat response times, bot vs agent)
- AWS, Conversational analytics metrics (interruptions, agent talk time percent, greeting time)
- AWS, Conversational analytics dashboard (contact drivers, AHT per driver, trends)
- AWS, Contact analytics data in the data lake (sentiment, talk/non-talk time, interruptions, talk speed, response times)
- AWS, Call analytics — talk speed (words per minute) [4a]
- AWS, Investigate non-talk time (hold + silence >3s)
- AWS, Investigate response time during chats
- AWS, Real-time alerts based on keywords and phrases
Content was rephrased for compliance with licensing restrictions.