What good looks like: CX benchmarks from 5 enterprise deployments

What good looks like: CX benchmarks from 5 enterprise deployments

The hardest question to answer before a CX deployment: Is the number you're being shown actually any good?

A vendor says 70% containment, but what does that even mean? 

• Is that strong? 

• Average? 

• Do they measure it the way you'd measure it? 

Most published benchmarks come from surveys of self-reported averages, but that only tells you what people say they achieve. However, it’s not necessarily what a specific deployment produced.

We’ll help you out. In our Innovator’s Report, we talked to CIOs and CTOs from known enterprise companies about their deployments and benchmarks. Here are the real numbers.

And where a company reported direction (instead of a figure), that’s noted.

The enterprise deployment numbers (CX benchmarks)

Two of the five enterprises reported direction rather than magnitude. That happens in regulated industries because specific operational figures are sensitive, but it’s good to keep in mind when you’re comparing against them.

Containment: What's realistic (and what to keep in mind)

Containment rate is the percentage of customer conversations an AI agent handles from start to finish without handing off to a human. If your voice agent fields 1,000 calls and 720 never reach a live agent, your containment rate is 72%.

Eltropy's up to 70% containment comes from voice deployments at community banks and credit unions, handling high-volume inquiries like balance checks and transaction history. Voice accounts for 65% of all customer communication at those institutions, so it's a large sample of a very specific kind of work.

Treat that as an upper band for well-scoped, transactional voice automation. The scoping is why it's achievable, so it won't transfer to a general-purpose agent.

Major caveat: Containment counts conversations the AI handled without escalating. It also counts conversations where the customer gave up and stopped replying. Those show up identically in your data.

71% of consumers will abandon a conversation entirely if an AI fails to recognize who they are. Every one of those abandonments improves your containment rate. That’s why we recommend splitting containment by outcome before you benchmark against anyone.

Speed benchmarks

In Twilio's Inside the Conversational AI Revolution report, 92% of consumers say AI typically replies in under 30 seconds. So a sub-30-second response is really the expectation rather than an achievement.

Really, patience is the more interesting number.

• 39% of consumers won't wait at all for an AI agent, while 88% will wait for a human

• Consumers will wait 8.5 minutes for a person against 4.3 minutes for AI

Your AI has roughly half the patience budget a human agent gets, and that changes what a reasonable handle time looks like.

Maybe that’s because consumers hope an AI can help them but know a human can? It also translates to real time costs because ineffective AI agents waste 81 seconds per interaction on failed attempts before escalating.

Satisfaction benchmarks

We have satisfaction benchmarks from two different Twilio studies that asked the same question and found almost exactly the same result:

1. The 2026 Customer Insights Series: 86% of companies believe customers are satisfied with their AI-powered service, while only 51% of consumers agree.

2. Inside the Conversational AI Revolution: 90% of organizations believed their customers were satisfied, but only 59% of consumers agreed.

Those are different sample audiences from different years, so we’re not necessarily drawing a trend line across them. Ultimately, it’s about consistency and reliability. We’re seeing internal satisfaction estimates run 30 to 35 points above what real customers report.

This means if you benchmark your AI CSAT against an industry average (against other enterprise companies), you’re probably just comparing one optimistic number against a bunch of other optimist numbers.

How to benchmark your own deployment

Now, before you compare yourself to anyone, we recommend the following, because most major benchmark disappointments are just measurements mismatches rather than performance gaps:

• Split containment by outcome. Resolved and abandoned are different results.

• Measure resolution on the errand. Whether the customer's underlying issue closed, across sessions and channels, not whether one session ended.

• Track repeat contact within seven days. No survey required, and nobody can game it.

• Measure abandonment by stage. This is the population your CSAT can't see.

• Time the handoff separately. 76% of consumers say the human agent they reach has little or no context, so escalated conversations deserve their own numbers.

Compare, measure, improve

The useful thing about the benchmarks above isn't their size. It's that each one is attached to a named company, specific industry, and scope you can evaluate against your own. Pick two or three of these, check whether the scope resembles yours, and measure the same way.

Read each deployment in full in the Innovators Report. Ready to measure your own? Start with Twilio for free.

Frequently asked questions

What is a good containment rate for an AI agent? 

Eltropy reports up to 70% average containment across dozens of voice deployments at community financial institutions. Treat that as an upper band for well-scoped voice automation, not a general benchmark.

Is containment rate a reliable measure of AI performance? 

Not on its own. Containment counts conversations the AI handled and conversations the customer abandoned identically. Twilio research found 71% of consumers abandon a conversation when an AI fails to recognize them.

How fast should an AI agent respond? 

Twilio research found 92% of consumers say AI typically replies in under 30 seconds. Patience is the tighter constraint: consumers will wait 8.5 minutes for a human and 4.3 minutes for AI.

What CX metrics should I track for AI deployments? 

Track containment split by outcome, resolution measured on the customer's underlying issue, repeat contact within seven days, abandonment by conversation stage, and separate timing for escalated conversations.

Why is my AI satisfaction score higher than customer sentiment? 

Internal estimates consistently run high. Twilio research found 86% of companies believe customers are satisfied with AI service while only 51% of consumers agree, largely because post-interaction surveys only reach customers who stayed.