One WhatsApp number, two businesses, nobody manning the phone — "Arjun"
How Yukti AI built and runs a WhatsApp AI sales agent on its own sister business — an agent that answers every inbound message instantly, works out which of two businesses the sender is asking about, retrieves the right answer at a controlled cost, hands off to a human when it should, and moves a new lead to "prospect" within about two hours.
Yukti AI's sister business shares one WhatsApp number between an AI automation agency and a trading-indicator subscription. Rather than staff it, Yukti built "Arjun": an n8n and WhatsApp Cloud API agent that classifies each message to the right business, answers from a Supabase pgvector knowledge base using a tiered set of models to keep cost low, captures the sender as a lead, advances the pipeline, and flags anything it shouldn't answer for a human. It runs in production every day, with zero manual work on inbound.
Quick facts
| Client | Yukti AI's own sister business (IndiabizlistFX) — an internal build, run in production |
|---|---|
| Setup | One WhatsApp number serving two businesses — an AI automation agency and a trading-indicator subscription |
| What Arjun does | Answers every inbound message, classifies the business, retrieves the answer, captures the lead, advances the pipeline, hands off to a human when needed |
| Response time | Instant, any hour |
| Lead handling | A new lead is typically moved to the "prospect" stage within ~2 hours of first contact, by the agent |
| Manual work on inbound | None — humans only touch conversations the agent hands off |
| Cost control | Tiered models — tiny model for classification/translation, mid model for FAQ rephrasing, top model only for knowledge-base answers that need it |
| Stack | n8n, WhatsApp Business Cloud API, Supabase + pgvector, OpenAI models; part of Yukti AI's multi-tenant CRM platform |
| Status | Live — running daily in production on Yukti's own operation |
Inbound WhatsApp is a full-time job that nobody has time for
Yukti AI's sister business has the same problem every small business has: enquiries arrive on WhatsApp, all day, and the first reply has to be fast or the lead cools. The twist is that one number carries two very different businesses — an AI automation agency, and a paid trading-indicator subscription. A message could be "do you build WhatsApp bots" or "my indicator alert isn't showing", and the right answer, the right knowledge base and the right sales pipeline are completely different.
Staffing that means someone who knows both businesses, watching the phone from morning to night, replying fast, logging every lead, and never confusing the two. That person is expensive, and on an off day the whole funnel leaks.
Yukti AI sells WhatsApp AI agents. Running one on our own business — the hard, two-in-one version — is how we know it works, keep it sharp, and can show a prospect a system we operate, not just a demo.
How inbound is handled
Before (staffing it)
- Someone watches the phone morning to night
- First reply waits for that person to be free
- They must know both businesses and never confuse them
- Every lead logged into the CRM by hand, or not at all
- An off day means the whole funnel leaks
- Cost scales with message volume
After (Arjun)
- Every message answered instantly, any hour
- The right business identified from the message content
- Answers grounded in that business's knowledge base
- Lead captured and pipeline advanced automatically
- Humans only touch conversations Arjun hands off
- Tiered models keep cost per conversation low at volume
Why build an agent, not use a chatbot tool
There are WhatsApp autoresponders and no-code chatbot builders. They send canned replies, run decision-tree menus, and don't know your product beyond what you scripted. Arjun had to do something harder: correctly serve two different businesses on one number, answer from a real knowledge base, move a CRM pipeline, and know when to step back. That is an agent, not a flow.
| What we needed | WhatsApp autoresponder / chatbot builder | Arjun |
|---|---|---|
| Route one number to the right business | Menu ("press 1 / press 2") | Classified from the message content, no menu |
| Answer from real product knowledge | Only what you scripted | Semantic search over a pgvector knowledge base |
| Reply in the customer's language | Usually one language | Translate step in the pipeline |
| Capture the lead & move the pipeline | Export a contact, maybe | Creates the contact, advances the stage |
| Know its own limits | Loops or gives a wrong answer | Flags for human handoff |
| Cost at volume | Per-conversation fee | Tiered models, most messages on the cheap tier |
Every message, start to finish
How a message becomes a moved lead
The agent, piece by piece
Business classification
Every message first goes to a small, cheap model that decides which of the two businesses the sender is asking about — from the content of what they wrote, not a menu. Everything after this point uses that business's knowledge base and that business's pipeline. Benefit: the sender gets the right answer without ever being asked "press 1 for…".
Tiered, cost-controlled retrieval
Not every message needs the most capable (and most expensive) model. Arjun uses a tiny model for classification and translation, a mid-sized model to rephrase answers to frequently asked questions, and the top model only for knowledge-base answers that genuinely require reasoning. Most conversations resolve on the cheaper tiers. Benefit: cost per conversation stays low, so the agent is economic to run at volume.
Knowledge-base answers (pgvector)
Each business's FAQs, product details and policies are embedded in Supabase with pgvector. When a message needs a real answer, the agent does a semantic search over that base and answers from what it finds — grounded, not improvised. Benefit: the agent talks about the real offering, and updating the knowledge base updates the agent.
Lead capture & pipeline movement
A new sender becomes a contact; the conversation sets and advances their pipeline stage; a genuine enquiry is moved toward "prospect". On Yukti's own operation, this happens within about two hours of first contact, entirely by the agent. Benefit: the CRM stays current without anyone updating it.
Human handoff
When a message is outside the knowledge base, or is a negotiation or a complaint, the agent flags the conversation for a human rather than guessing. A person takes it over from the CRM inbox, and Arjun stays out of that thread until it is handed back. Benefit: the agent knows the edge of its competence.
What it does for us
Observed on Yukti AI's own sister business — stated plainly.
Technology
- n8n — the workflow that orchestrates classification, retrieval, reply and lead updates.
- WhatsApp Business Cloud API — inbound and outbound messaging on the shared number.
- Supabase + pgvector — each business's knowledge base, embedded for semantic retrieval, plus the contact and pipeline data.
- OpenAI models across tiers — a tiny model for classification and translation, a mid model for FAQ rephrasing, the top model for knowledge-base answers.
- Yukti AI CRM — Arjun is part of the same multi-tenant CRM platform, so handed-off conversations land in a real inbox.
The same agent, grounded in your business
Arjun is Yukti AI's own build, running in production — a system we operate every day, not a demo. The same agent is built for other businesses, grounded in their catalogue, their FAQs and their pipeline:
- Answer every WhatsApp enquiry instantly, in the customer's language
- Ground replies in your real product data, not generic text
- Capture the lead and advance your pipeline automatically
- Hand off cleanly to your team when a conversation needs a person
- Run at a controlled, predictable AI cost
Start with a free workflow audit. See WhatsApp AI Chatbot, AI Agent Development or Yukti AI CRM.
Questions about a WhatsApp AI sales agent
Explore more
WhatsApp AI Chatbot
An AI that answers your WhatsApp enquiries grounded in your real data.
AI Agent Development
Custom AI agents that do a job end to end — not just chat.
Yukti AI CRM
The multi-tenant CRM platform Arjun is part of — WhatsApp, pipeline and AI in one.
Case study: Surati Jewellery Factory
A two-way WhatsApp inbox with an AI first-responder inside a jewellery CRM.
Case study: Sureja
An outbound voice AI for payment reminders — the calling counterpart to Arjun.
All case studies
Every Yukti AI build — jewellery, manufacturing, trading and services.
Summary
Yukti AI built and runs "Arjun", a WhatsApp AI sales agent, on its own sister business, where one WhatsApp number serves two businesses — an AI automation agency and a trading-indicator subscription. Arjun answers every inbound message instantly: a small cheap model classifies which business the message is for and translates if needed; a semantic search over that business's Supabase pgvector knowledge base retrieves the answer; a mid-sized model rephrases FAQ answers while the top model is reserved for knowledge-base answers that need reasoning, keeping cost per conversation low; the sender is captured as a contact and advanced through that business's pipeline — typically reaching the "prospect" stage within about two hours of first contact; and anything outside the knowledge base, or a negotiation or complaint, is flagged for human handoff into the CRM inbox. Built on n8n, the WhatsApp Business Cloud API, Supabase with pgvector and OpenAI models, as part of Yukti AI's multi-tenant CRM platform. It runs in production daily with no manual work on inbound. Yukti AI builds the same agent for other businesses, grounded in their own data.