Today’s customer support agents are extremely powerful—Salesforce just bought Fin for $3 billion, and the market is continuing to expand quickly. 

But these agents aren’t perfect. When they, it’s because they're answering questions they weren’t trained on, don’t have data for, or weren’t expecting. .

When a user opens a support chat and types "I can't export my report," the agent hears that sentence but it may not see that the user has tried the export flow three times in the last 20 minutes, failed at the same step each time, and is now frustrated enough to type in all caps. So it returns a generic walkthrough the user already tried, and a support ticket gets submitted anyway.

The missing insight between what a user says to agents, and what they've been doing in your product, is where ticket deflection breaks down. But CS teams are beginning to close this with behavioral insights. 

Why smart agents still create support tickets

Most AI agent deployments start with the same setup: a knowledge base, a chat interface, and a reasonable expectation that the two together will resolve most user issues. For simple, one-dimensional questions, that works fine.

The problem is that most support interactions aren't simple. They're specific. A user stuck at step three of a six-step workflow doesn't need a general answer about the feature. They need help with what went wrong at step three. An agent without behavioral context can't make that distinction. Every user looks the same, and every question is treated as if it arrived with no history.

Why ticket volume stays flat after agent deployment

This is why you can deploy a genuinely capable AI agent and still see ticket volume hold steady. Agents are just missing the layer of information that would let them respond to what's actually happening.

Why ticket volume stays flat after agent deployment

This is why you can deploy a genuinely capable AI agent and still see ticket volume hold steady. Agents are just missing the layer of information that would let them respond to what's actually happening.

Here's what that looks like in practice: 

  1. A user encounters friction in a workflow. 
  2. They open the agent chat and describe their problem in the most general terms they can find, because they don't know the technical name for what went wrong.
  3. The agent matches keywords, returns a response that addresses the general topic, and misses the specific failure entirely. 
  4. The user tries the suggestion, it doesn't help, and they submit a ticket.

The ticket could’ve been avoided if your agents had what it needed to prevent it.

Behavioral signals: The context agents have been missing

Product behavioral data gives agents what knowledge bases can't: information about what users have actually been doing in your product before conversations begin. 

The four signals that drive ticket deflection

There are four categories that matter most:

  1. Repeated attempts tell the agent that a user has tried the same action multiple times without success. That's different than a user asking about the feature for the first time, and it calls for a different response.
  2. Drop-off point tells the agent exactly where in a workflow a user got stuck, not just that they're stuck. If a user has completed steps one through four and abandoned at step five every time, the answer lives at step five, not step one.
  3. Frustration signals, such as rage clicks, repeated inputs, and rapid back-and-forth navigation, indicate that the user's patience is already gone before the conversation opens. The first response needs to land. There's no runway for a miss.
  4. User context tells the agent whether it's talking to someone in their first week with the product or a power user hitting an edge case. A new user needs onboarding guidance. A power user needs a specific fix. Those are different conversations.

How context changes agent interactions

When an agent has access to these signals before the conversation starts, the agent already understands the situation and can respond directly. It’s all about enabling better conversations with agents to bypass tickets.

Pendo Agent Analytics is what captures and surfaces these behavioral signals, giving your agents the context layer they've been missing.

Pendo for AI Agents
Govern, measure, and prove ROI on your entire AI portfolio.
See how it works

Agent observability: the metrics that tell you if deflection is actually happening

Once behavioral data is informing your agent, you need visibility into whether it's paying off via agent observability. 

Resolution rate tells you how often agents close conversations without a human handoff, but not why agents fail when they do. For that, you need to go a level deeper.

The agent metrics that matter for IT leaders

  • Rage prompt rate measures prompts containing profanity, all caps, or repeated requests. A high rage prompt rate means users are already frustrated before the agent has a chance to help. If this number is climbing, something in the pre-conversation experience is creating friction that the agent is then inheriting.
  • Drop-off rate within the agent interaction tells you where users abandon. If users consistently exit after the agent's second response, the problem is the second response. That's a trainable issue once you can see it, but you have to be able to see it first.
  • Task completion rate answers the question that resolution rate skips: did the user actually do the thing the agent helped them with? A user can close a chat without submitting a ticket and still not have their problem solved. 

These numbers tell you whether your agent investment is defensible at the budget level. 

Command Center gives you the portfolio-level view of how every AI tool is performing, so agent metrics don't live in a silo. They connect to the broader picture of what's working across your AI stack.

AI ROI in practice: 86% of conversations resolved without a human

A user hits friction in a workflow. The behavioral signals are captured in real time. When that user opens the agent chat, the agent already has context: what they were doing, where they got stuck, how frustrated they appear. The response addresses the specific situation. The issue gets resolved before it becomes a ticket.

That’s what Teachable built with Fin, powered by Pendo data: 

"Because Fin is powered by Pendo data, it's catching customers at critical moments and resolving incoming questions before they become support tickets. We're already seeing Fin solve 86% of conversations without looping in a human, all while maintaining a CX score of 4/5 stars."

— Kathleen Ross, Product Support Associate Director, Teachable

Why that’s a business metric, not a support metric

Every conversation Fin resolves without a human handoff is a measurable cost avoided. Holding a 4/5 CX score while doing it means quality didn't drop to hit the number. The agent reads the actual context, so it doesn't have to guess.

The Improvement Loop: How Deflection Gets Better Over Time

The initial deployment is only part of it. Behavioral data arms the agent before a conversation and shows you what to fix after. When rage prompts a spike or drop-off rate at the same point repeatedly, you adjust the agent's response, retrain it on the specific failure, or surface the workflow gap to the product team. Deflection improves because the feedback loop is continuous, not because the configuration was right on day one.

Behavioral data is what makes deflection actually happen

Most AI agents are deployed with good intentions and underperform for a fixable reason: they're missing context. Product behavioral data is that context. It arms agents before the conversation starts, surfaces where they're failing after, and gives IT leaders the visibility to prove the investment is working.

The 86% is simple science: agents that know what users have been doing resolve more than agents that only know what users are saying right now.

See how Pendo helps you govern, measure, and prove ROI on your entire AI portfolio.