Choosing the interface

The interface is where people meet the agent. Choose it for the audience and the kind of work, not for novelty. The same agent can serve several channels, but private and public audiences should not share one.

At a glance

Interface	Best for	Typical audience
Copilot	A full working view: conversations, testing, escalations, human takeover	Your own team
Live chat and SDK	Conversations on your website, portal or product	Customers, prospects
Email and SMS	Asynchronous questions and agent-sent messages	Customers and staff
Voice	Spoken conversations by phone	Customers
Microsoft Teams	Questions inside the tool staff already use	Your own team
Claude and ChatGPT over MCP	Reaching a raia agent from a person's AI assistant	Your own team
A custom front end	Structured tasks such as forms, uploads and reviews	Anyone

Copilot

Copilot is the right choice when your team needs a full working interface rather than a small chat window. It's also where you test agents, see escalated conversations, and step in to take over. When an escalation happens, a person can continue the same conversation.

Tools still run when you test in Copilot. Test with non-production systems, or with tools limited to safe actions, until you're confident.

Learn more: raia Copilot and Human-in-the-loop.

Live chat and the SDK

Use the embeddable live chat widget for conversations on your website or in your product. The SDK can pass context into the conversation, such as a person's name, their plan, the page they're on, or a CRM contact reference, so the agent doesn't have to ask.

Context passed from the browser personalizes a conversation, but it doesn't prove who someone is. Anyone can change it. Before an agent reads private records or makes changes for a person, verify their identity on your own servers. See Live chat security.

Learn more: raia Chat and SDK and embed guide.

Email and SMS

Agents can have their own email addresses, sending domains and SMS numbers. Number availability depends on country and provider setup.

When a Loop needs to send an email, use the explicit "Send an email" step with a recipient, subject and body. Asking an agent in a prompt to "email someone" isn't the same as a guaranteed send, even if its email channel is switched on.

A sent step means the email channel accepted the message. It doesn't prove the message reached an inbox. If delivery matters, plan for evidence from your mail provider. See Loops and outbound.

Voice

Voice agents have their own instructions and knowledge. Keep the knowledge focused and answers short, because people on the phone notice delays. Test number setup and behavior for your region before launch. Learn more: Voice.

Microsoft Teams, Claude and ChatGPT

Many staff already work inside Microsoft Teams or an AI assistant. A raia agent can be reached from these tools, so knowledge, permissions and cost controls stay in one place while people work where they already are. One validated agent can answer the same questions for many people, instead of each person loading documents into their own chats.

Learn more: Microsoft Teams, Add an agent to Claude and Add an agent to ChatGPT.

A custom front end

Some work isn't a conversation. Uploading receipts on a phone, previewing extracted fields, and correcting a form are easier on a purpose-built screen than in chat. In these cases, raia agents and Loops do the AI work and the workflow behind a lightweight custom application.

The custom application is a separate deliverable with its own owner. Adding agents and Loops doesn't create the screen itself. Plan who will build and maintain it. See Operations, data and finance scenarios.

Choosing between them

Start with where the audience already is. Customers on your website need live chat; a support team living in Teams needs Teams.
Separate audiences when information is private. An internal agent and a customer agent should be two agents, even if they share some public knowledge.
Prefer the simplest surface that does the job. Add channels after the first one is working well.
Choose a custom front end only when the task is structured enough that chat would slow people down.