What's New in AI Voice Agents: Five Features Worth a Closer Listen

[Justin_Wilburn]

 

People still call. Especially when it matters. Over the last few releases, we’ve focused on making sure your voice agents meet them wherever they are, hears the callers clearly, and make it easy to see what happened. Here's the short version:

• Key Term Dictionary helps your voice agent hear your company terms and acronyms correctly.
• Conversation Activity shows you exactly what happened on any call.
• Portal Integration puts a "call us" button right where your users already are.
• Custom SIP Connections let you bring your own phone setup.
• Voice Analytics in Assistant Designer shows you which agents are pulling their weight.

1. Key Term Dictionary: Teach your agent your language

Every company has words that speech recognition gets wrong. Product names. Internal acronyms. That one system everyone calls by a nickname. When a caller says "Otto" and the transcript says "auto," the whole conversation goes sideways.

The Key Term Dictionary fixes that at the source. Key terms improve speech-to-text accuracy for domain-specific vocabulary that might otherwise be misheard as a similar-sounding common word. Entries are language-specific, and you can add up to 30 key terms per language.

Try it:

1. Open your voice assistant in Assistant Designer and go to the language and voice persona step.
2. Select the Key term dictionary tab, then Add key term.
3. Start with the five words your callers say most that aren't in an everyday dictionary.

Did you know? There's a sibling feature for the other half of the conversation. The Key Term Dictionary improves how your agent hears, while the Pronunciation Dictionary controls how it speaks. You can enter a word and its pronunciation using phonetic spelling or phoneme format. So your agent can say "SQL" like your engineers do.

Please Note: results can vary based on audio quality, caller accent, and context. Test with real calls, then tune.

👉 Create an AI voice assistant

2. Conversation Activity: See every call, turn by turn

“What actually happened on that call?” is the question every admin asks the first week after go-live. Now you can answer it in a couple of clicks.

The Conversations tab in Assistant Designer logs every completed voice interaction across all of your voice assistants. You can use it to review call outcomes, spot problems with tool executions or authentication, and open the full transcript for any call.

Scan the list. Each row is one call. It shows how the call ended (resolved by AI, resolved by AI and a live agent, or not resolved), which voice agents were invoked, the intents detected, authentication status, and the assists consumed.

Drill into one call. Select a row to open its detail report. Here’s what’s inside:

• Performance at a glance: call duration broken into bot talk time, caller talk time, and silence, plus authentication attempts, average response latency, and average tool execution time.
• The transcript: timestamped, with inline markers for agent invocations, tool calls, and authentication events.
• Tool Execution: every tool invoked, with its status, execution time, request parameters, and response.
• Conversation Insights: signals like confusion, effort score, agent empathy, and frustration.

So when a call goes sideways, you can see the exact moment it happened and the reason, not just the outcome.

Try it:

1. In Assistant Designer, open the Conversations tab.
2. Selecting any row from the list.
3. Review the details of the interaction report, including: key metrics, transcripts, event logs, and conversation insights.

You can also troubleshoot from the Conversation tab:

1. Filter by resolution status or authentication status to find the calls that didn’t land.
2. Open one and compare the transcript with the tool execution log. If the agent skipped a step or called the wrong tool, that’s usually a sign its instructions need a tweak.

3. Portal Integration: Put voice where your people already are

Not every conversation starts with a phone number. Sometimes it starts with someone stuck on a portal page, wishing they could just ask.

Now they can. In Assistant Designer, the Web applications option lets you configure a voice call widget on your ServiceNow Portal or Engagement Messenger. No phone, no hold music, no repeating your account number.

What makes this worth your time:

• Guests get in without friction. Unauthenticated portal visitors are automatically authenticated using a cookie-based session, with no extra widget configuration required. They're still subject to the identification and authentication factors you configure.
• Handoffs arrive with context. When a portal call goes to a live agent, it carries the case record, contact record, and the page the customer started from.

Try it:

1. Open your voice assistant and go to Communication channels.
2. Select the WebRTC tab, then Web applications.
3. Configure the voice call widget for your portal or Engagement Messenger.

👉 Enable WebRTC for voice calls

4. Custom SIP Connections: Bring your own phone setup

Not using one of the named contact center providers we have out-of-the-box integrations with? No problem.

You can connect SIP-capable telephony systems to a voice assistant using the Custom provider option. This is also great for organizations that route calls through their own SIP infrastructure.

Translation: your phone system doesn't have to change for your callers to get a better experience.

👉 Integrate with a custom SIP provider

5. Voice Analytics in Assistant Designer: Know what's working

You can't improve what you can't see. Voice Analytics gives you the full picture without leaving Assistant Designer.

Metrics are organized across four tabs (Overview, Performance, Insights, and Assist consumption) covering conversation volume, resolution, tool usage, authentication, conversation quality, and assist consumption. A few views worth bookmarking:

• Agent performance: conversation counts and AI-inferred resolution rates per voice agent, so you can spot high and low performers.
• Empathy: a read on how polite and attentive your agents are, to guide tone improvements.
• Assist consumption: average actions per call and total assists per voice agent.
• Handle time: average duration of the AI-handled portion of calls, excluding time spent with a live agent after transfer.

Try it:

1. Open Assistant Designer and select the Analytics tab.
2. Pick a date range and start on Overview.
3. Find your lowest-resolution agent on Performance, then pull a few of its calls in Conversation Activity. This pairing is where many quick wins come from.

👉 Voice page in Assistant analytics

Honorable mentions

🎙Voice model expansion. We use a flexible model strategy across speech recognition, reasoning, and speech generation, then evaluates complete voice systems before adopting model changes. EVA-Bench, ServiceNow's open evaluation framework for end-to-end voice agent benchmarking, is at the heart of this evaluation. The result is you get access to strong model innovation without taking on the burden of model selection and regression testing.

🗣 Dynamic voice activity detection (VAD). Major enhancements to the technology that tells your agent when a caller has actually finished talking, so it stops interrupting and stops leaving awkward pauses. Note: Available for P1 Languages only today.

🌎 More ways to say hello. Canadian English and Malay have joined the lineup. Voice assistants now support 25+ languages and dialects, from Brazilian Portuguese to Thai to Irish English. You can set one primary language and any number of secondary languages.

Your next step

Pick one feature. Turn it on this week. Then listen to a few calls or browse the analytics to see what changed. 

Not sure where to begin? Try starting with the Key Term Dictionary. It takes ten minutes and your callers will notice.

👉 Explore the full AI Voice Agents documentation