What an AI agent is (and why it works as a funnel)
An AI agent is software that talks in natural language, understands what the customer wants and takes action: it answers, fills in data, books meetings and brings in a person when needed. For the full concept, read what AI agents are and how they change sales. This guide has a different goal: building your own.
At Nuvia, an AI agent is not a standalone bot. It works inside a Smart Funnel: each type of conversation has its own funnel, with stages, rules and its own agent. When the topic changes, the conversation moves to another funnel without changing place, and the history, fields and summary go with it. That is why the first step is not writing a prompt. It is mapping your operation.
Before you build: define your funnels and stages
Before opening any tool, list the types of conversation that reach your WhatsApp and what needs to happen in each one. One funnel per conversation type covers most operations:
| Funnel | Purpose | Example stages |
|---|---|---|
| Triage | Understand the topic and route the conversation to the right funnel | Reception, topic identification, routing |
| Support | Answer questions from the knowledge base and solve problems | Service, data collection, support team, resolved |
| Sales | Qualify the lead and book the meeting or close the order | Qualification, proposal, scheduling, negotiation |
| Onboarding | Get new customers to their first use | Welcome, sign-up, first use |
| Post-sale | Follow the delivery, ask for a review and offer a repeat purchase | Follow-up, review, repeat purchase |
Each stage has a prompt. You write in plain language what the agent does there and when to move to the next one. It answers from the Knowledge Hub, fills in the fields and moves the conversation through the funnel. There is no decision tree to draw, which is the big difference from a flow-based chatbot.
A stage instruction can be as simple as: “In Reception, greet the customer, find out whether they need support, want to buy or have a billing question, and send the conversation to the right funnel with a summary. If they ask for a person, hand over right away.”
Step by step: how to build an AI agent with the Agent Builder
The Agent Builder is how you create and improve your Smart Funnel by talking to an AI. If you can explain your process, you can build your agent. It takes five steps:
Connect your AI
Add the Nuvia MCP server to Claude, ChatGPT, Gemini or Cursor and authorize it with your login.
Explain your operation
Describe the process in plain language or send your manual and playbook.
Review the proposal
Check funnels, stages, knowledge and follow-ups, explained in plain language.
Simulate the conversation
Test with fictional data, without talking to any customer.
Activate on the platform
Everything starts as a draft. You decide when it goes live.
Step 1: connect the AI you already use over MCP
The Agent Builder works in the AI assistant your team already uses: Claude, ChatGPT, Cursor, Gemini (via Google Antigravity) or your own agent. It connects over MCP, an open standard that lets an AI use tools from other systems:
- Add Nuvia as a connector, using the address
https://api.nuvia.ai/mcp. - Sign in with your Nuvia email and password and click Authorize.
- Ask for what you need in plain language, such as “create a support funnel”.
The connection uses your login and is limited to your company’s data. Any user can read and simulate, but creating and editing agents is reserved for admins. The technical details are on the API, MCP and webhooks page.
Step 2: explain your operation or send your playbook
You have two options. The first is to explain it in your own words, as you would to a new hire:
“I serve customers on WhatsApp. I want the AI to answer questions from our knowledge base, send people who want to buy to sales and bring in support when someone needs account access.”
The second is to send the documents you already have: a customer service manual in PDF, a sales playbook in DOCX, a qualification script. The Agent Builder reads the material, understands the process and builds the funnels, the instructions for each stage, the knowledge base and the follow-ups with you.
Step 3: review funnels, stages, knowledge and follow-ups
The proposal arrives explained step by step, ready for your review. Check four things:
| What to review | Question to ask |
|---|---|
| Funnels | Does each conversation type have its own funnel? Is the handoff between them clear? |
| Stages and instructions | Does the agent know what to do at each stage and when to move on? |
| Knowledge Hub | Are the right documents linked? Are prices, deadlines or policies missing? |
| Follow-ups | At which stages does the agent re-engage people who went quiet, after how long and with which message? |
Don’t like something? Ask for the change in one sentence, such as “only talk about pricing after you understand the case”. A follow-up can be plain text, a WhatsApp template or a message written by the AI, sent at any time or only during business hours.
Step 4: simulate with fictional data
Ask your assistant for a simulation, for example: “simulate a customer asking about credit cards in support”. The agent replies as it would in real life, without talking to any customer. A simulation runs up to 10 messages with fictional data, and external actions, such as booking a meeting or updating the CRM, are only flagged, not executed.
At the end, the quality check points out blockers and warnings, such as a tool that has not been connected yet. Adjust and simulate again until the conversation goes the way your best rep would handle it.
Step 5: publish as a draft and activate on the platform
Everything the AI creates over MCP starts as a draft. You activate it on the Nuvia platform. Three safeguards protect your operation:
- Lock on live agents: changes to an agent that is already live require your confirmation.
- Quality check: blockers and warnings show up before activation.
- MCP audit log: every access and action is recorded, with the user and the result.
Gets better with every rated conversation
Mark conversations as good or bad. Agent Builder finds the pattern and proposes instruction changes that you approve before they apply.
Learns from your team
It compares the conversations your team handled well with the AI's and proposes bringing over the questions, the order and the tone.
Simulate before the customer
Up to 10 messages with sample data. The agent replies as it would for real, without talking to any customer.
Google Calendar
HubSpot
MCPIntegrations you ask for in writing
Calendar with availability checks, CRM, webhooks and calls to your system, set up in plain language and reviewed before publishing.
- Always a draftEverything published via MCP is a draft. You activate it in the platform.
- Lock on live agentsChanges to a live agent require your confirmation.
- Quality checkBlockers and warnings flagged before activation.
- MCP audit logEvery access and action is recorded, with user and result.
Nothing goes live without you
Any user can read and simulate. Creating and editing agents is limited to admins.
How to build an AI agent with ChatGPT, Claude or Gemini
You build your customer-facing agent using ChatGPT, Claude or Gemini as the builder, not as the place where the agent lives. An assistant set up only inside one of these AIs talks to whoever uses that account. To serve your company’s WhatsApp, the agent must be connected to the number, the team inbox, the CRM and the calendar.
| What you want | Where to do it |
|---|---|
| An assistant that helps you write, summarize or research | Directly in ChatGPT, Claude or Gemini |
| An agent that serves customers on your company’s WhatsApp | In the Smart Funnel, built by the Agent Builder from your AI |
| Changing an agent that is already live | In the same chat with your AI, with your confirmation before it applies |
In practice, you stay in the assistant you already know. The difference is that, connected to Nuvia over MCP, it builds and tunes an agent that talks to your real customers.
What the agent does during the conversation
An AI agent acts, it doesn’t just reply. While it talks on WhatsApp, it:
- Books the meeting: checks availability in Google Calendar and schedules it on the spot.
- Updates the CRM both ways: with HubSpot, RD Station or Salesforce, stage, fields and owner go to the CRM, and changes your team makes there update Nuvia.
- Sends the contract: sends it for signature through ZapSign or Certisign without leaving the conversation.
- Qualifies the lead: in sales conversations, it uses BANT, GPCT or MEDDIC and gives a Lead Score from 0 to 100 based on your ICP.
- Re-engages people who went quiet: sends follow-ups on the cadence you set, with a message tied to the context.
- Checks your system: with an HTTP skill, it fetches real data (such as an order status) and answers the customer.
You also ask for these integrations in writing in the Agent Builder, such as “book on the specialist’s calendar, move the deal in HubSpot and notify my system by webhook”, and they go into the draft for your review.
WhatsApp
RD Station
Salesforce
ZapSign
CertisignAI and your team in the same conversation
You choose where the AI replies and where only your team talks. The AI can be turned on or off per stage or for the whole funnel. A common setup is to keep the AI on triage and qualification and hand negotiation to a sales rep.
Even with a person in the conversation, the Smart Funnel keeps working: it updates the stage, the summary and the customer analysis. Whoever takes over gets a ready summary, and the customer doesn’t have to repeat the story. Whenever you want, the team hands the conversation back to the AI.
How the agent improves after it goes live
The agent improves with the conversations you rate. Mark conversations as good or bad, and the Agent Builder finds the pattern and suggests changes to the instructions. For example, if the agent brought up pricing too early, it suggests “only discuss investment after you understand the case”.
It also compares the conversations your team handled well with the AI’s and suggests bringing what works into the instructions: the questions, the order and the tone. Suggestions come explained in plain language, and nothing changes without your approval.
How long it takes and what to expect
The first complete draft is ready in minutes, built from your documents. The time until it goes live depends on your review and simulations. To see what messages cost on the official WhatsApp API, check the guide WhatsApp Business API pricing.
With the agent live, the AI works alongside your team. At AbacatePay, whose team builds and tunes its own Smart Funnel through the Nuvia MCP, 92% of conversations have the AI at work, on its own or alongside the team, and 82% of switches between support and sales are made by the AI itself, with no customer repeating themselves.
Support and sales in the same agent: within a week most conversations were already handled by AI, without losing the human touch with customers.
Read the storyFrequently asked questions
How do you build an AI agent?
Define the funnels and stages of your operation, connect the AI you already use to the Agent Builder over MCP and explain your process or send your playbook. It sets up funnels, stages, knowledge and follow-ups as a draft, and you review, simulate and activate.
How do I create an AI agent with ChatGPT?
Connect ChatGPT to Nuvia over MCP and ask for the agent in plain language, such as “create a support funnel for my WhatsApp”. The Agent Builder builds the agent as a draft, and you simulate and activate it on the platform.
Can I build an AI agent with Gemini?
Yes. The Agent Builder works with Gemini via Google Antigravity and with other MCP-compatible assistants, such as Claude, ChatGPT and Cursor.
Do I need to code to build an AI agent?
No. You explain the process in plain language or send the documents you already have, and each stage of the agent is a written prompt. Code only comes in if you want to connect your own system through the API or webhooks.
How long does it take to build an AI agent?
The first complete draft is ready in minutes, built from your documents. The time until it goes live depends on your review and simulations.
What is the difference between an AI agent and a chatbot?
A chatbot follows hand-drawn paths and gets stuck when the customer goes off script. An AI agent understands natural language, steers the conversation back to the goal and takes action, as the chatbot vs AI agent comparison shows.
Does the AI agent go live on its own?
No. Everything the AI creates over MCP starts as a draft and goes through the quality check. You activate it on the platform, and changes to a live agent require your confirmation.
