Webinar Recap: Deploying Agents Across Every Channel

Webinar Recap: Deploying Agents Across Every Channel

Your customers don't all communicate the same way - your agent shouldn’t either. Some want to call. Some want to message. Some want to email. 

Your agent needs to meet them where they are - without you rebuilding it from scratch for every channel. 

In Live Workshop: Deploying Agents Across Every Channel, we showed how you can configure once and deploy everywhere.

Why an omnichannel strategy matters

Customer expectations have shifted. 24/7 support with no hold time is the baseline. When a brand fails to meet that bar, it doesn't just create a bad interaction. It becomes how people talk about that brand.

Companies with a strong customer experience strategy see 1.5x higher revenue growth than those without. When you look specifically at omnichannel execution, that number sharpens: 15% more revenue and 35% more loyalty compared to single-channel approaches.

Companies with strong omnichannel engagement also retain 89% of their customers. Companies without it retain just 33%. This makes it a structural competitive advantage.

AI is already moving the needle on satisfaction scores. 86% percent of service leaders who have deployed AI say it directly improved their CSAT. The question is no longer whether AI can handle customer interactions - the question is whether you are deploying it everywhere your customers actually are.

The omnichannel problem most teams are ignoring

Most companies running an omnichannel strategy are actually running several single-channel strategies stitched together. A different vendor for voice, a different configuration for chat, and another for SMS. Each one optimized in isolation.

Context doesn't transfer between interactions and tone and logic vary depending on where someone reaches you. Technical debt compounds as each channel gets its own maintenance cycle. And when something breaks, it's never clear which integration owns the problem.

The underlying issue is architectural. If an agent is built for one channel and adapted for others, adaptation always means loss - of consistency, of context, of the ability to improve the whole system at once.

The cleaner approach is to build at the logic layer, not the channel layer. One agent, one knowledge base, one set of policies. The channel becomes a rendering surface, not a design constraint.

Voice is worth thinking about specifically here. It tends to get bolted on last, but it's the hardest channel to retrofit well. Conversational voice requires different latency tolerances, different error handling, and different design patterns than text. Teams that treat it as an add-on often find themselves re-platforming when voice becomes a priority - which, given that it consistently outperforms other support channels on customer satisfaction, it usually does.

The question for most teams isn't whether to consolidate. It's when, and how much it will cost to undo what's already been built.

Demo 1: Adding a checked bag over SMS 

Scenario: Sarah is planning a ski trip on a Tuesday evening. She wants to know what it costs to check a ski bag before she commits to bringing one. She texts the airline instead of digging through an FAQ or waiting on hold.

What was shown:

• Sarah sent a text to the agent asking about ski bag fees
• The agent authenticated her using her booking number and last name
• It pulled up her reservation and confirmed the $75 each-way fee for specialty equipment
• The agent added the bag to her booking and generated a secure payment link
• The entire interaction resolved in a single text thread with no wait time

Why it matters: Low-urgency, informational questions like this clog call centers when there is no other option. SMS resolves them instantly, keeps the queue clear for urgent issues, and lets customers get answers on their own time without friction.

Demo 2: Rebooking a canceled flight over voice

Scenario: It is the morning of the trip. Sarah clears security and sees her flight is canceled. She has a tight connection to a mountain shuttle and needs to rebook before she leaves the terminal. This is not a texting moment. She calls.

What was shown:

• Sarah called the same agent that handled her SMS interaction
• The agent authenticated her again using her booking number and name
• It confirmed the cancellation was due to a technical aircraft issue and that she was eligible for rebooking
• The agent surfaced three alternative flights with departure times and arrival details
• Sarah chose the 7:30 PM direct flight to Denver and the agent confirmed the change in real time
• The agent confirmed her seat, confirmed her two checked bags carried over, and sent updated confirmation to her email
• Tool calls were visible throughout, showing the agent checking flight status, querying alternatives, and executing the rebook

Why it matters: When a flight cancels, hundreds of passengers rush to rebook at the same time. A voice agent that can authenticate, query live inventory, and execute changes ends that queue. It also does something a scripted IVR cannot: it responds to urgency. The agent acknowledged Sarah's stress, moved quickly, and delivered a resolution in under two minutes.

Demo 3: Filing a baggage damage claim in-app

Scenario: Sarah lands and finds her snowboard bag has a large tear that was not there when she checked it. She is standing at the baggage carousel with her phone in hand. She opens the app to file a claim.

What was shown:

• Sarah opened the airline app and initiated a voice interaction through the in-app widget
• She described the damage and the agent asked her to upload a photo
• The agent processed the image using a multimodal LLM and assessed the damage: a catastrophic tear across the top of the bag, rendering it non-functional
• It auto-populated a damage claim form with her details and the image attached
• Sarah reviewed and submitted the form directly in the app
• The claim was now trackable in the app without any further contact needed

Why it matters: This demo showed two capabilities that go beyond standard chatbot functionality. First, the agent accepted and interpreted an image in the same conversation. Second, it took action by populating and surfacing the claim form rather than just pointing Sarah toward a help page. The issue moved from discovery to filed claim in one interaction.

In any agent conversation, there are three points where safety must be considered. 

How the agent is actually configured 

All three channels are routed into a single greeting and FAQ router. 

That router triaged each request and handed off to one of three specialized sub-agents:

1. A transaction agent for adding the ski bag
2. A flight delay rebooking agent for the cancellation
3. A baggage claim agent for the damage report.

Each sub-agent has its own conversational goal, its own tool access, and its own guardrails. 

The rebooking agent, for example, only offers rebooking if the delay exceeds two hours. It is the only sub-agent with permission to rebook. 

The baggage claim agent was powered by a multimodal LLM to handle image inputs. 

The transaction agent used a faster LLM optimized for quick responses.

This structure means the agent follows intent, not a decision tree. It behaves the way a well-trained human support agent would: understanding what the customer needs, routing to the right capability, and taking action within defined limits.

Best practices for deploying an omnichannel agent 

1. Configure once, then choose your deployment surfaces. Phone, SMS, WhatsApp, web widget, and mobile apps are all available. You do not rebuild the agent for each one. You choose which surfaces to activate.
2. Use sub-agents to scope permissions tightly. Only the rebooking agent should be able to rebook. Only the refund agent should be able to issue a refund. Narrow tool access by sub-agent prevents the agent from taking actions outside its lane.
3. Match the LLM to the task. Use a multimodal model where image or file processing is needed. Use a faster, lighter model for transactional interactions that prioritize speed. You are not locked into one LLM across the entire agent.
4. Make authentication specific to your risk tolerance. A booking number plus last name is a reasonable starting point. One-time SMS codes, account PINs, or other confirmation steps can be layered in depending on what actions the agent is authorized to take.
5. Use expressive audio tags to make the voice feel natural. ElevenLabs' latest conversational model v3 supports audio tags that add pauses, breaths, and tone variation. Combine that with deliberate prompt writing and the right voice selection to close the gap between AI and human.
6. Think beyond support. Every channel is also a revenue surface. Outbound calls can deliver proactive resolutions with account context already loaded. SMS can carry time-sensitive offers after a purchase. In-app voice can guide a customer through an upgrade mid-checkout. The same agent that closes a support ticket can open the next conversation.

Watch the full session 

Watch the full webinar here.