Contact center analytics: 6 metrics for AI and human agents

Contact center analytics was built for a world where people answered every call. Can you imagine? No more rerouting and Kenny G hold music (I wasn’t complaining, Kenny). Now AI agents handle a share of conversations and human agents take what's left.

However, the contact center metrics you've tracked for years can mislead you about both.

Let’s look at which metrics matter when AI and human agents share the work, how AI changes each one, and how to measure them with Twilio.

What is contact center analytics?

Contact center analytics is the practice of measuring and analyzing customer conversations and agent activity to understand service quality, efficiency, and customer experience across voice and digital channels.

Traditionally, that meant queue stats and agent productivity:

• How long customers wait

How long customers wait

• How long calls take

How long calls take

• How many hang up

How many hang up

With AI agents in the mix, your call center analytics has to answer some new questions:

• Which conversations did the AI agent resolve on its own?

Which conversations did the AI agent resolve on its own?

• Where did customers give up on it?

Where did customers give up on it?

• Are the conversations that reach a human agent harder than they used to be?

Are the conversations that reach a human agent harder than they used to be?

Why containment rate shouldn’t be your priority metric

Containment rate is the share of conversations an AI agent or self-service flow handles without a human agent. It's the headline number for most AI agent deployments, but it says nothing about whether the customer got what they needed.

Simply put, it’s not enough. It doesn’t tell the full story, and a half-story can be more confusing and misleading than no story.

A customer who gets a clear answer counts as contained. But so does a customer who gives up and hangs up, or who tries again tomorrow on a different channel. Push containment hard enough, and you can make it climb by making a person harder to reach.

That’s not the goal, and it doesn’t make for very happy customers.

That's why the metrics below focus on outcomes. Track those for AI and human agents alike, and you'll know whether your AI agent is helping customers or deflecting them.

These six metrics cover most of what contact center leaders report on today. For each, here's what it measures, how AI agents change it, and how to track it with Twilio.

1. Customer satisfaction (CSAT)

CSAT measures how satisfied customers are with an interaction, usually through a short post-conversation survey. When AI agents handle part of the journey, measure CSAT separately for AI-only and human-handled conversations, or a strong human team can hide a weak AI agent.

CSAT is something you collect and load. In Twilio Flex, you can load survey scores into Flex Insights ascustom conversation measures (up to ten numeric values per conversation).

Twilio's tutorial onpost-task surveys with Flex Insights walks through the build, from a Studio survey to survey reports. Prefer a ready-made option? Flex integrates withMindful Feedback for post-call IVR, SMS, email, and web surveys.

Most customers skip surveys, though. To estimate satisfaction on every conversation, write a Conversation Intelligence custom operator that scores it from the transcript, and track that score alongside your survey results.

2. Sentiment

Sentiment shows how customers felt during a conversation, and it covers every conversation, including the customers who skip the survey. That matters for AI agents, since customers who leave an AI conversation frustrated rarely stop to rate it.

I mean, would you? Probably not.

Twilio measures sentiment natively.Conversation Intelligence includes aTwilio-authored Sentiment operator that classifies each conversation as positive, negative, neutral, or mixed, across voice and messaging channels.Conversation Insights (currently in public beta) can then break down conversations by Sentiment label. For a trend you can average, write a custom operator that returns a numeric sentiment score and register it as a custom measure.

In Flex, Agent Copilot (currently in public beta), captures customer sentiment in its wrap-up notes for human-handled conversations. Flex Insights (also in public beta) reportsAgent Copilot metrics like conversations with negative, neutral, or positive sentiment.

Learn how to build it yourself with Twilio'sConversation Intelligence tutorial forNode.js—it captures sentiment, a summary, and script adherence for every call and stores the results in your own database.

3. Resolution

Resolution measures whether the customer's issue got solved. It's the clearest test of whether an AI agent is doing its job, but it's hard to capture from structured data alone.

Twilio handles it two ways:

• In Flex, Agent Copilot generates disposition codes in its wrap-up notes, which agents can review and edit, and you can pass your own outcome attributes to Flex Insights. Twilio's Flex project template includes a Dispositions feature that writes the agent's wrap-up code to the Outcome field in Flex Insights.

In Flex, Agent Copilot generates disposition codes in its wrap-up notes, which agents can review and edit, and you can pass your own outcome attributes to Flex Insights. Twilio's Flex project template includes a Dispositions feature that writes the agent's wrap-up code to the Outcome field in Flex Insights.

• With Conversation Intelligence, you can score resolution with acustom language operator. Twilio'scustom operator examples include a Conversation Scoring operator that weights problem resolution at 35% of an overall quality score. For a dedicated resolution metric, have your operator return a separate numeric resolution field and register it as a custom measure.

With Conversation Intelligence, you can score resolution with acustom language operator. Twilio'scustom operator examples include a Conversation Scoring operator that weights problem resolution at 35% of an overall quality score. For a dedicated resolution metric, have your operator return a separate numeric resolution field and register it as a custom measure.

4. Average handle time (AHT)

AHT measures how long agents spend on a conversation, including talk, hold, and wrap-up time. AI agents tend to shift this number. When they take the routine contacts, the conversations that reach human agents are often more complex, so human AHT can rise even as overall efficiency improves.

Flex reports AHT natively. Thereal-time Queues View shows average handle time, talk time, and wrap-up time for the last 30 minutes and today. Flex Insights adds historical handling time, which includes wrap-up work. For voice AI agents, theConversation Relay Insights dashboard shows how long virtual agent and human agent calls run, grouped into duration buckets.

5. Abandonment and drop-off

Abandonment measures customers who leave before an agent answers. Drop-off covers customers who leave an AI or self-service conversation partway through.

Both matter, and they're easy to confuse.

Flex tracks queue abandonment natively, including abandoned tasks, percent abandoned, and short abandons in the Queues View, andabandoned conversation rates in Flex Insights. The Flex project template's Hang Up By Reporting feature records who ended each call, which keeps abandonment reporting accurate.

For voice AI agents, Conversation Relay Insights shows whether the customer or the virtual agent ended the call, and the share of silent calls, which helps you spot where customers gave up. A full AI drop-off rate is something you define from those signals.

6. Intent

Intent analysis shows why customers contact you, which tells you what to automate next and where AI agents fall short.

Conversation Intelligence doesn't include a built-in intent operator, but a custom operator with classification output can label each conversation's intent. Conversation Insights can count conversations by that label, so you can see your top contact reasons and how they shift over time. The same results are available by webhook or from theOperator Results API if you'd rather analyze them in your own tools.

How to build contact center analytics with Twilio

Most teams assemble analytics from four layers, starting with what's built in:

• Flex built-in metrics: The real-time Queues View covers AHT, talk and wrap-up time, abandonment, SLA, and average speed of answer. Flex Insights, in public beta, adds historical dashboards, threshold alerts, and conversation assessments for quality reviews.

Flex built-in metrics: The real-time Queues View covers AHT, talk and wrap-up time, abandonment, SLA, and average speed of answer. Flex Insights, in public beta, adds historical dashboards, threshold alerts, and conversation assessments for quality reviews.

• Voice AI agent monitoring:Conversation Relay Insights, generally available as part of the paid Voice Insights Advanced Features, tracks time to first audio, interruptions, errors, silent calls, turns per call, and who hung up, with asingle-call view for debugging.

Voice AI agent monitoring:Conversation Relay Insights, generally available as part of the paid Voice Insights Advanced Features, tracks time to first audio, interruptions, errors, silent calls, turns per call, and who hung up, with asingle-call view for debugging.

• Conversation scoring: Conversation Intelligence runs Twilio-authored operators like Sentiment and Script Adherence, plus custom operators you write for resolution, intent, or quality scores, across voice and messaging.

Conversation scoring: Conversation Intelligence runs Twilio-authored operators like Sentiment and Script Adherence, plus custom operators you write for resolution, intent, or quality scores, across voice and messaging.

• Aggregation and comparison: Conversation Insights, in public beta, is an API that aggregates conversation and operator data. Results arrive about 15 minutes after conversation events, ready for your own dashboards or BI tools.

Aggregation and comparison: Conversation Insights, in public beta, is an API that aggregates conversation and operator data. Results arrive about 15 minutes after conversation events, ready for your own dashboards or BI tools.

Start with Flex's built-in metrics, add AI agent monitoring when you deploy voice AI agents, and layer in scoring and aggregation as your definitions mature.

How to compare AI and human agent performance

Twilio compares AI and human agent performance through Conversation Insights, which can break down conversation data by channel and by whether an AI agent, a human agent, or both took part in the conversation. The goal is benchmarking your workforce (both human and AI) on the exact same metrics.

That comparison answers the questions single-agent dashboards miss. Register a numeric score your own operator produces, like a quality score or a resolution score, as a custom measure. You can then average it by channel, compare it between AI and human agents, or track it week over week. A conversation handed off from an AI agent to a human agent counts as both, so you can compare handoffs against conversations either one handled alone.

Measure your contact center on what actually matters

AI agents didn't make contact center analytics obsolete. Instead, they made the old dashboards incomplete. The teams that get this right track AI and human agents with the same scores, measure outcomes over deflection, and watch what happens at the handoff.

Twilio gives you the pieces to make it happen:

• Built-in Flex metrics

Built-in Flex metrics

• Voice AI agent monitoring

Voice AI agent monitoring

• Conversation scoring you can shape to your own definitions

Conversation scoring you can shape to your own definitions

• An API that compares AI and human agents across voice and messaging

An API that compares AI and human agents across voice and messaging

Start for free and start scoring conversations with Conversation Intelligence, ortalk to sales about analytics for your Flex contact center.

Frequently asked questions

What platforms provide agent performance dashboards showing containment rate, CSAT, and average handle time?

Twilio Flex reports average handle time natively, and survey CSAT scores load into Flex Insights as custom measures. Twilio Conversation Intelligence adds what containment misses: sentiment labels and resolution scores from custom operators.

What tools support conversation analytics including intent detection, drop-off rates, and resolution scoring?

Twilio Conversation Intelligence classifies sentiment and supports custom operators you write for resolution and intent. Conversation Relay Insights shows who hung up on voice AI calls, and Flex tracks queue abandonment.

Is containment rate a good metric for AI agents?

Twilio Conversation Intelligence can measure what containment rate misses: whether the customer's issue got resolved and how they felt. Containment alone counts customers who hang up in frustration as successes.

How do you measure CSAT for AI agents?

Twilio Flex loads post-conversation survey scores into Flex Insights as custom measures. Survey AI-only conversations separately, or score satisfaction from the transcript with a Conversation Intelligence custom operator.

Can you compare AI agent and human agent performance?

Twilio Conversation Insights breaks down conversation data by channel and by whether an AI agent, a human agent, or both took part. Conversation Relay Insights also splits voice call volume and handling time between virtual and human agents.