Every product, variant, price, minimum quantity and bulk rate. It can compare two products, recommend an alternative when something is out of stock, and calculate a total for a mixed order including GST.
An agent that knows yourcatalogue, your prices andyour rules
Not a decision tree with a friendly voice. An agent that reasons from what it knows about your business — and hands over to you when it should not decide alone.
Chatbot, AI chatbot, AI agent — the actual differences
| Question | Rule-based chatbot | AI chatbot | AI agent |
|---|---|---|---|
| How it answers | Matches keywords to pre-written replies | Generates a reply from a language model | Reasons from your business context, then acts |
| Unexpected question | Falls through or loops | Answers, sometimes invents an answer | Answers from your data, or says it does not know and escalates |
| Who builds the logic | You, on a flow canvas | You, plus a prompt | Nobody — it is derived from your business context |
| Can it take an action? | Only what you wired up | Usually not — it talks | Yes: sends payment links, creates invoices, books slots, moves segments |
| After the conversation | Nothing | Nothing | Segments the customer and starts the follow-up |
| Gets better over time? | Only if you edit it | Only if you edit the prompt | Yes, from your corrections and from what actually earned money |
What it knows about
your business
It reads your website and learns how you describe what you sell. The answers sound like your business rather than like generic software.
The discount ceiling. What it must never promise — delivery dates you cannot guarantee, claims you are not allowed to make. Which enquiries go straight to a person. These are the rules it will not cross.
What they bought before, what they paid, what they asked last time, which segment they are in. A returning wholesale buyer does not get treated like a first-time visitor.
When it stops and calls you
Knowing when not to answer matters more than knowing how to answer. The agent escalates in five situations, and you can add your own.
| Trigger | What happens |
|---|---|
| It is not confident | Below a confidence threshold it stops rather than guessing. The conversation goes to a person with a summary of what has been said. |
| The order is large | You set a value above which a human confirms. Most B2B businesses set this at their average order value. |
| The customer is upset | Frustration, a complaint or an escalation in tone routes straight to a person. It does not try to handle a complaint. |
| It touches a rule | A discount request beyond the ceiling, or anything on your never-promise list. |
| The customer asks for a person | Immediately, without an argument. It does not try to deflect. |
Can you do ₹2,400 if I take 200 sheets?
That's below what I'm able to confirm on my own — let me get Sandeep on this. He'll message you within the hour.
Handed to SandeepContext: 200 sheets 20mm marine ply, asked for ₹2,400 vs list ₹2,690
What stops it going wrong
It answers from your data, not from the internet
Prices, stock and policies come from what you gave it. It does not improvise a price or invent a delivery date.
Guardrails you write in plain English
A page of rules, not a configuration file. “Never promise same-day delivery.” “Never discuss competitors.” “Never go below ₹2,600.”
Everything is logged
Every message and every escalation decision, exportable. If a customer disputes what was said, you have the record.
Off, Suggest, Approve or Auto
Four autonomy levels. Start at Approve, where nothing goes out without you seeing it, and move to Auto when you trust it.
Questions about the AI agent
It answers from your catalogue and rules rather than generating facts, which removes most of the risk. Below a confidence threshold it escalates instead of guessing. If something does go wrong, you will see it in the inbox, you can correct it there, and the correction becomes part of what it knows. Every conversation is logged, so you always have the record.
Yes, for business-specific agents. From January 2026 Meta stopped allowing general-purpose AI assistants — the kind that answer any question about anything — on the WhatsApp Business Platform. An agent that represents one business and answers about that business's products remains fully permitted, and that is what this is.
It replies in the language the customer writes in, including Hinglish and the common Roman-script forms of Hindi, Tamil, Telugu, Marathi and Gujarati. Your catalogue and rules can stay in English.
You choose. Many businesses set a short disclosure in the first message, and some regulators are moving that way. We recommend disclosing — it costs almost nothing in conversion and it removes the risk of a customer feeling deceived later.
No. Your catalogue, conversations and customer data are used to serve your business only. They are not used to train models, they are not pooled with other customers' data, and this is written into the contract, not just the marketing.
A general assistant does not know your prices, cannot send a payment link, cannot issue an invoice, does not know which customer it is talking to, and does nothing after the conversation ends. It also is not permitted on the WhatsApp Business Platform. The value here is in the actions and the retention, not in the conversation quality alone.
See it run on
your own catalogue
Send us your product list and we will set the agent up on it before the call. You will watch it answer questions about your own products, not a generic demo account.Thirty minutes. No obligation.
A demo on your products, not ours
Pick a time that works. Share a catalogue, price list or website link when you book and we load it into the agent before we speak.
- Before the call
You send a catalogue, price list or website link. We set the agent up on it.
- On the call · 30 min
You ask it the questions your customers actually ask. We show you the segments and follow-ups it would run.
- After
A written summary of what it would do for you, and what it would cost.
Can't see the calendar? Book here
Or email sandeep@3xretention.com