AI confidence is rising. Accuracy remains a blind spot.

Good metrics help you predict whether you’re moving toward the outcomes you want. They also show you how far you’ve come, or how far you’ve drifted from where you intended to go.

But AI has changed how customer service gets delivered, and we suspected many organizations hadn’t stopped to ask whether their metrics still did either of those things. Were they looking in the right places? Was what they were measuring still relevant?

So we partnered with Metric Sherpa to find out. We surveyed 531 U.S. customer experience (CX) and contact center professionals and 238 executives for the 2026 CX Observability Study. We asked what they measure across human and AI interactions, how much they trust those numbers, and what happens when they find a problem.

The study showed where traditional CX metrics fall short. Speed, containment, and transfer rates can tell you whether an AI agent finished the job, but not whether its answer was right.

Confidence grows, but gaps remain

What stands out is that confidence rises as AI becomes more embedded, even though measurement gaps don’t shrink. Operators report blind spots at roughly the same rate across adoption stages, and executives remain concerned about undetected AI errors throughout.

Although 94.7% trust that their current metrics reflect the real customer experience, 89.5% can name at least one area they can’t see. Even operators with the highest trust report as many blind spots as those with the lowest. That confidence comes with some healthy discomfort. They know AI is doing more than their dashboards reveal, and they aren’t satisfied with a partial view.

That partial view reflects the scorecard CX teams inherited. They’ve spent years refining measures such as average handle time, containment, and agent conduct, but those measures were largely designed for human-centered operations. They don’t automatically capture the ways an AI agent can fail.

Accuracy is the least-measured AI risk

The clearest example is factual accuracy.

Among the 367 operators whose organizations use AI, 78.5% report having no metric for hallucinations or factual accuracy. Even among operators who say they’re very confident in their AI, 76.6% report the same missing metric.

The gap continues at the executive level. Only 27.9% of executives say hallucination or factual-accuracy data reaches them, the lowest rate for any safety or compliance measure included in the study. The ones who do get it value it: 60.9% report high trust in that data.

You might expect regulated industries to be different, but they aren’t. 79.3% of operators in healthcare, financial services, insurance, government, and utilities report no metric for hallucinations or factual accuracy, compared with 78.0% elsewhere.

A missing accuracy metric puts you at risk on two fronts: compliance and reputation. Regulations like the EU AI Act are raising expectations that companies can prove their AI is accurate, and “we didn’t know” won’t be an acceptable answer much longer.

Your brand is on the line, too. Customers notice a wrong answer even when your dashboard doesn’t. And when an AI error can repeat across customers before anyone catches it, that lack of visibility becomes a ticking time bomb.

How CX observability helps you find your blind spots

Defusing that risk takes CX observability: complete visibility into every customer interaction, whether it’s handled by a human agent, an AI agent, or both.

The term is still new. Only 32.0% of operators and 30.6% of executives said they were very or extremely familiar with it. But once it was defined, more than half of both groups called it very or extremely valuable. That tracks with what we saw on accuracy: executives trust that data when they get it; most just aren’t getting it yet.

With CX observability in place, you can set a visibility standard for each stage of your AI deployment. Before expanding an AI agent’s role, your team should be able to answer a few basic questions: Which interactions does the AI handle? How is factual accuracy measured? How quickly are errors detected, assigned, and corrected? Who can pause further expansion if the necessary evidence doesn’t exist? Does accuracy data reach the executives responsible for CX outcomes? If your team can’t answer those questions yet, close those gaps before moving forward.

Read the full 2026 CX Observability Study to find your organization’s CX blind spots, assess your current measurement practice using the five-part CX Observability Index, and learn where to focus next.