Designing for empathy in contact center AI
A voice agent can apologize for the inconvenience and still ask someone to repeat everything they've already said. Real empathy lives in the full design, from the context the agent has before the call starts to the channel it chooses in the moment.
Most teams designing AI for customer service start in the same place. Warmer words, softer tone, and an acknowledgment when someone sounds frustrated. It makes sense as an instinct, but it’s why so many contact center AI experiences still frustrate the customers they’re meant to serve.
You can apologize for the inconvenience and still ask someone to repeat everything they've already told you. And you can acknowledge frustration warmly and still route them to the wrong place.
Designing for empathy means understanding the reality of the person on the other end of the call. That understanding lives in the full design: the context the agent has before the call starts, how it sequences information, when it speaks and when it waits, and the channel it chooses for any given moment. Getting those decisions right is what determines whether a conversation actually serves the person in it.
The best conversations start before the call
Nothing tells a caller you're not ready to serve them quite like gathering notes about them while they're on the phone. From the phone number alone, a well-designed agent can identify the account, see the last interaction, and anticipate what the customer is likely calling about. The customer feels understood before they've had to say a word.
Listing capabilities on the first turn ("I can help with outages, billing, new service…") triggers IVR muscle memory. The caller starts sorting their issue into predefined categories instead of telling you what's wrong. A context-aware opening works differently. Before the caller speaks, the agent already knows the phone number on record, the account attached to it, and what the last interaction was about. That changes the emotional weight of the call from the first second.
That requires design work that happens long before the conversation starts:
- What context does the agent have going in?
- What does it ask for, and in what order?
- Does the sequence feel intuitive enough that explanation would just slow things down?
Getting that groundwork right means the conversation starts on the right foot.
The cost of over-explaining
Over the phone, a caller can't scroll back, can't screenshot, or pause and re-read. Whatever the agent says has to land in real time, often with the rest of life happening in the background.
Agents built on general-purpose large language models are naturally over-informative. Without the right controls, a voice agent will over-explain every step. For example, after each instruction, it might add, "Let me know when you've done that. Tell me when you're finished," as if the caller needs to be coached through silence. The reflex is familiar: keep the interaction moving, fill the pause, prove the agent is still there. In practice, it makes the caller feel managed instead of helped.
What the difference looks like in practice:
| Over informative | Designed for voice |
|---|---|
| “I need your account number, then your date of birth, then your service address. Once I have all that, I can check your outage status.” | “Let’s start with your account number.” |
| “Let me know when you’ve done that. Tell me when you’re finished.” | Silence. The agent waits. |
| “Your reset code is one, four, eight, seven, two. I’ll repeat that: one, four, eight, seven, two.” | “Your reset code is 14872.” |
Designing the conversation means designing for the channel
The phone is usually the first channel customers reach for when they need support, but it isn't always the best one for the moment. The best-designed conversations know when another channel is better suited. If a customer has a hard to spell surname, an effective voice agent can handle the request, but sometimes it’s easier to send the form via SMS in real time instead. Treating voice and SMS as one conversation gives the agent more ways to meet the customer where they are.
Knowing when to stop trying to solve something over voice at all is a design decision too.
Empathy under pressure
PG&E serves around 15 million customers across Northern and Central California, and call volume tends to spike at the worst possible moments: an unexpected bill, a midnight outage, a safety question during a storm. Before the shift to conversational AI, those calls hit layered touch-tone menus that did nothing to acknowledge how stressful the call already was. Now, even on days when volume exceeds a million calls, the caller is recognized before they need to explain themselves.
Read the full story: How PG&E built a voice agent that earns customer trust
A different kind of pressure shows up in hearing healthcare. Audibel’s customers are often older, often frustrated, and often calling about the device they rely on to hear the world clearly. Long waits and touch-tone menus add friction to a call that already asks a lot of them. After moving to a voice agent built around the caller’s experience, Audibel cut wait times by 87% and brought call abandonment from 46% to 2%. Those numbers come from designing the conversation around what the customer is actually carrying into the call, with speed as a downstream effect.
Real customer behavior is the only honest test
Production is where empathetic design gets proven. The challenge is circular: customers won’t engage with AI self-service until it earns their trust, and you can’t improve the experience without data from real customer behavior. The path is to make the first version good enough that people actually use it, then act quickly on what you learn. Waiting for the system to be perfect before gathering feedback is the surest way to never gets there.
In our recent webinar on building AI agents for the real world , Oliver Shoulson, PolyAI’s Agent Design and Engineering Lead, put it plainly: “You are going to get yourself locked into your imagination of what the end user is going to say, and those will never all be right.” A quiet-room demo doesn’t show you what happens when a customer is calling from their car with a child screaming in the back seat, the TV blaring through the speakerphone, and the radio on.
The best teams don't wait until launch to find this out. They involve real end users early, treat experience changes during early deployment as urgent priorities, and build the agility to respond fast.
Every design decision adds up
Empathy in your contact center AI comes from design decisions made at every level. The context the agent has before the call begins, how it sequences information, how it sets expectations in the first three seconds, the channels it uses, and how it behaves when real life gets loud all determine whether a customer feels genuinely served or just processed.
Good copy still matters, and the right words at the right moment earn trust. A well-designed conversation earns something more lasting.
Make every customer feel heard. Instantly. Speak to our team today.
