AI · 6 min read · Savina Bansal
AI agents inside WhatsApp flows: qualify a lead without a script
A chatbot follows the path you drew. An agent is given a job, the tools for it and hard limits. Where each belongs in a WhatsApp flow, and how to keep an agent safe.
Building the platform this article is about
A WhatsApp flow is a set of steps you draw: when this happens, send that, ask this, branch on the answer. It is predictable, which is its strength. It is also rigid. A flow cannot read a rambling first message and decide whether this person is a serious buyer, because nobody can draw every way a person might say it.
That judgement is where an AI agent belongs. Not instead of the flow, but as one step inside it.
A reply is not an agent
An AI reply answers a question from your knowledge. An agent is given a job and does it: it reads, decides and acts, using tools. In Ohanvi an agent is built from five things.
Instructions: the job, in plain words. “Read the lead’s first message and their answers. Decide whether they are ready to buy, comparing, or just curious. Add a note saying which, and why.”
Knowledge: the sources it may answer from, picked from your knowledge base.
Tools: what it may do. Find leads and contacts, add notes and tasks, tag a contact, send an approved WhatsApp template. It can only be given tools the person building it is allowed to use, and tools from your other systems’ MCP servers once an admin has approved each one.
Skills: ready-made know-how the agent can lean on. Built-in skills are read-only, so nobody edits one by accident.
Limits: the most steps it may take, the most changes it may make to your data, and the most AI cost a single run may spend.
Put it where judgement is needed
Picture a home-interiors studio that runs click-to-WhatsApp ads. Every lead lands in the CRM, and most are not ready: some want a price for one cupboard, some are planning a whole flat next year, a few want a site visit this week. The sales team calls them in the order they arrived.
The flow starts on “New CRM lead”. Its first steps ask two questions on WhatsApp: what are you planning, and when. Then a “Run an AI agent” step reads the answers and the original message, tags the lead, and adds a note for the sales team. The flow branches on the tag: the ready ones get a site-visit template and a task for a person; the rest get a brochure and a follow-up date.
The agent did one thing no drawn path could: it read free text and made a call. Everything around it is still a flow you can see and test.
Draft, test, publish
An agent has a draft and a published version. Test runs the draft, so you can try it on real-looking input without touching the version your flows use. Publish makes the draft the version flows run.
Every run is kept, step by step: what it read, which tools it called, what it changed. When a lead was tagged wrongly, you can see why, and change the instructions rather than guess.
Keep it on a short lead
The limits are the safety net. A step cap stops an agent going round in circles. A cap on changes means a bad instruction can tag a handful of contacts, not your whole list. A cost cap per run means a strange input cannot run up a bill. On top of that, monthly and per-task spending caps in AI Settings stop AI spend on our key altogether once reached.
And because tools are limited to what the builder may do, an agent made by a sales executive cannot reach data a sales executive cannot.
What it costs
One agent run is charged like one AI reply: ₹2 on our key, or 50 paise on your own model key. Calls to your MCP servers cost nothing extra. No monthly AI fee on any plan.
Keep reading
Questions this answers
- What is the difference between a WhatsApp chatbot and an AI agent?
- A chatbot follows a path you drew. An AI agent is given a job, the tools it may use and limits, and decides how to do it — for example reading a lead’s messages and tagging them. In Ohanvi an agent runs as one step inside a flow.
- Can an AI agent change my CRM data?
- Only with the tools you give it, only within what the person who built it may do, and only up to the change limit you set per run. Every run is kept step by step.
- How much does an AI agent run cost?
- One run is charged like one AI reply: ₹2 on our key or 50 paise on your own key, with no monthly AI fee.
