AI in Samvaads
Two kinds of AI, and the difference is the point.
Copilot works for the person replying. AI agents talk to customers themselves. Both read from your knowledge base rather than the open web, and both end with a human able to take over.
Copilot works for the agent
Copilot sits beside the reply box. It summarizes a long thread before someone picks it up, drafts the response, rewrites it shorter or warmer, and answers questions from your published articles with the sources shown. Nothing it produces reaches a customer until a person sends it.
- Summarize a thread, draft a reply, rewrite what is already there
- Answers cite the knowledge base article they came from
- Call transcription and other assists on the same panel
- Everything stays a suggestion until someone presses send
AI agents work for the customer
An agent is a flow you draw. Steps send messages, ask questions, branch on conditions, look an answer up in your knowledge base, or call a tool. You simulate it, publish a version, and deploy that version to the channels you choose.
- Message, question, choice, condition, set field, AI resolve, handoff, and end steps
- Chat, email, and voice modes from the same flow engine
- Versions with simulate and publish, so a channel only runs what you approved
- Handoff carries a written summary to whoever takes over
Agents can do things, not just say things
A flow step can call your own HTTP endpoint or an MCP tool, so an agent can look up an order, check a warranty, or start a journey mid-conversation. Samvaads also ships its own MCP server, which lets MCP-aware clients work inside your desk.
- HTTP tool steps against your endpoints, with the response available to later steps
- MCP tool steps for anything already exposed over MCP
- Samvaad MCP wraps the external API for clients like Claude and Cursor
- Every tool stays bounded by the API key scope you issued
What the AI is allowed to see
Contact email and phone are masked in the workspace until someone explicitly reveals them, and that reveal is recorded. Sensitive strings in inbound messages are redacted before they are stored, which means before an AI agent reads them.
- Answers grounded in your knowledge base and allowlisted tools, not the open web
- Revealing a masked contact detail is an audited action with who, when, and which fields
- Card numbers, OTPs, and similar values are replaced before the message is saved
- A human can take the conversation back at any point in the flow
See it answer your questions
Bring a few real threads to the walkthrough and we will point Copilot and an agent at them.