5 questions every CX Leader should be asking their conversation data

Brightree, a leading medical software company, manages complex contact center operations for more than 50 partners. Known for their high standards of service, their team constantly monitors their AI performance to ensure peak efficiency.

When containment metrics recently signaled room for optimization, Brightree bypassed the slow, conventional diagnostic path —which traditionally requires pulling disparate site reports, sampling recordings, and trying to construct theories from fragmented data. Instead, they proactively leveraged a single Smart Analyst query to instantly optimize their system.

The query revealed that approximately 80% of patient identification and verification failures were actually tracing back to an external source: a newly onboarded healthcare client whose patient records had poor data quality. Because those patients weren't yet appearing in the system, verification failed.

Armed with this continuous monitoring insight, Brightree identified the exact root cause and implemented a clear fix within minutes.

That story is useful not just because it's fast. It's useful because of what it reveals about where customer intelligence actually lives. Brightree's team didn't need to go find the data; it was already there, in every failed call, every transfer, every verification attempt that hit a wall. What changed was the ability to ask it a question.

Your contact center, chat queues, and email threads generate the same kind of intelligence every day. Customers reach out with information your business needs. Most organizations catch fragments of it through sampling and manual review. Smart Analyst makes the rest available in plain language, at a speed that matches the pace of the problem.

Smart Analyst is an AI analytics agent built into the PolyAI platform that lets you query your conversation data across calls, chat, and email in plain language, and get back answers specific enough to act on. Here are five questions worth asking your conversation data every week, and what teams are finding when they do.

"Why aren't my conversations being contained?"

This is the first question most CX leaders reach for when the numbers move. Containment rate tells you that something is wrong. It doesn't tell you where.

At scale, the manual diagnosis is slow by design. Site reports land in different formats from different sources. Recordings have to be sampled. By the time a theory takes shape, the problem has been running for days.

The Brightree query took minutes — and traced the problem to a specific customer integration, across 30+ sites, with the specificity needed to fix it. The data existed before anyone ran the query. What Smart Analyst changed was the time between problem and answer.

"What topics is my knowledge base missing?"

Your knowledge base was built on what you anticipated customers would ask. Customers are already telling you what you got wrong — in every call, chat, and email interaction that didn't resolve.

Conventional KB auditing relies on manual review: pull a sample of non-contained conversations, identify recurring topics, update accordingly. It catches what a reviewer notices, on the schedule a reviewer is available. A topic has to generate visible volume before the process surfaces it.

Two weeks after launching its PolyAI agent, Michael's on East took a different approach. The query: "Analyze the calls since Oct 16 onwards and identify trends or topics in calls that weren't contained. Please include call IDs for each topic and highlight the topics that could be solved by adding an FAQ to the knowledge base."

The output was a prioritized list of KB gaps with call IDs attached, so the team could pull supporting examples for every finding — evidence for each recommendation, not just the recommendation itself.

"Where are my customers getting stuck?"

CSAT scores and transfer rates tell you something is wrong. They don't tell you which moment in the conversation caused it.

Flow problems register as symptoms (a drop in satisfaction, a spike in transfers) without pointing to a location. Fixing the wrong part of the flow wastes the fix: resources spent and the actual problem still running.

At Golden Nugget, Director of Contact Center Operations Brian Jeppesen used Smart Analyst to examine where callers were dropping out of the identity verification flow — pinpointing that DOB mismatches and account lookup failures were the specific friction points driving handoffs, not the broader flow itself. The finding gave the team a targeted fix rather than a broad redesign.

"What are my customers asking for that I'm not offering?"

The conventional path to answering this question is expensive and slow. Commission a survey, wait for responses, analyze the results — and by the time a finding reaches a decision-maker, the trend it captured is months old.

A major global logistics company used Smart Analyst to analyze patterns in their non-contained calls and found that address-related delivery exceptions were the most common friction point (43.5% of exceptions!) with customers consistently looking for online resolution rather than an agent transfer. They worked with the client to build that resolution path. Call containment increased by more than 10%.

That finding didn't come from a commissioned study. Customers had already produced it, one conversation at a time.

95% of CX and contact center leaders agree that their departments would benefit from improvements made to other parts of the customer journey. The contact center along with your chat and email channels is the only function in most businesses generating that first-party intelligence continuously, in the customer's own words.

As Brian Jeppesen at Landry's put it: "With PolyAI's Analyst Agents, we can ask questions in plain language and get answers right away. This gives us a clear view of what's happening across our customer interactions and helps us make smarter, faster decisions to improve our service."

"Where can my automation improve?"

Automation performance is usually evaluated at the output level: containment rate, resolution time, transfer volume. What's harder to see is which specific moments in the conversation are holding those numbers back.

The team at a legacy hotel-restaurant chain used Smart Analyst to analyze conversations around housekeeping automation — specifically to understand what was preventing higher success rates. The findings drove decisions to add intent codes and refine automation logic. Success rate improved 5% quarter-over-quarter.

The 5% came from understanding the conversational patterns causing automation to fail, and acting on them. The insight preceded the metric.

Love's Travel Stop used Smart Analyst after their initial launch to identify what to build next. Rather than guessing at new use cases, the team used conversation data to surface where caller needs were going unmet and which automation opportunities had the most volume behind them. The roadmap came from the conversations themselves.

One more question: what aren't you asking?

The five questions above are starting points. The more durable practice is to ask Smart Analyst what you're not thinking to ask and keep a running list of the queries that return the most useful answers.

The value isn't only in answering the questions already on your list. It's in surfacing the ones that haven't made your list yet: topics accumulating in non-contained calls, friction three turns into a flow you assumed was working, a request pattern no one has escalated because no one had a way to see it at volume.

PolyAI's CEO Nikola Mrkšić has described where Smart Analyst is heading: "Smart Analyst will be detecting trends and pushing them to you. The whole evolution of our agentic tech is going to be working more and more so that you do less. You get more information. You make smart decisions, and you move on."

For now, the questions are yours to ask — and the data has been waiting for them. Build your first agent in 10 minutes on PolyAI’s Agentic Dialog Platform right now to see how to build beyond mere conversational AI .