AI Agent Builder lets your customers put a conversational agent in front of their own customers. One agent, one set of instructions, one knowledge base, deployed across whichever channels that account has active.
It is a paid module, so treat it as something you sell rather than something every account gets.
What an agent is made of
- A system prompt. The instructions that set what the agent does and how it speaks.
- A knowledge base. Grounding content, which can come from crawling the customer’s website or from documents they upload.
- A ruleset and guardrails. Topic filtering so the agent stays on subject, PII redaction, and escalation when a conversation needs a human.
- Tools. Actions the agent can take rather than just answering.
- Channel bindings. Where the agent is reachable.
- Analytics. What it handled and how.
Setup is no code, and it supports models from the major providers, including OpenAI, Anthropic, Azure OpenAI and Google Gemini. Your customer picks the model, so you are not locked into one vendor’s pricing or policies.
Where an agent can be deployed
WhatsApp, RCS, SMS, Telegram, the web chat widget, and the REST API. Which of those appear depends on which messaging products are active on the account. An account without RCS cannot bind an agent to RCS, which is worth knowing before a sales call turns into a support ticket.
The distinction that confuses everyone
Your platform has two AI related products and they do opposite things:
| AI Agent Builder | MCP Server Integration | |
|---|---|---|
| Who talks to it | Your customer’s end customers | Your customer’s own AI tools and developers |
| What it is | A managed bot you deploy on messaging channels | An interface that lets an external AI host call your platform’s functions |
| Typical use | Answering questions, taking orders, qualifying leads | An AI assistant checking balance, sending a message, pulling a DLR |
| Direction | Inbound conversations from the public | Outbound tool calls into the platform |
Keep these separate in every proposal and every support reply. A customer who asks for “AI on our messaging” could mean either, and building the wrong one wastes a week.
Selling it without creating problems
- Guardrails are the selling point, not the model. Anyone can connect a model to WhatsApp. Topic filtering, PII redaction and a clean escalation path to a human are what makes it safe for a bank or a hospital to use.
- Grounding beats prompting. An agent that answers from the customer’s own site and documents is far more useful than one running on a clever prompt. Push customers to supply real content during onboarding.
- Set escalation up on day one. The fastest way to lose trust in a bot is for a frustrated person to have no way to reach a human.
- Watch the analytics in the first fortnight. That is where you find the questions the agent is failing, which is usually a gap in the knowledge base rather than a model problem.
One thing to check per deployment
Model choices and any provider credentials belong to the account that uses them. Make it explicit in your commercials who pays for model usage, because that cost sits outside your per message pricing.
Related articles
- Web Chat Widget: Putting an Agent on Your Customer’s Website
- Journey Builder: Cross-Channel Journeys
- Telegram Chat Plugin and Telegram Notifications: Two Different Things
- Does your WhatsApp business software provide live chat agent?