The Indian SMB sales stack is WhatsApp plus AI now

If you sell anything in India and you're under 100 people, your sales happen on WhatsApp. Not on the website. Not in the CRM. WhatsApp. The website is a brochure, the CRM is where you copy notes after the fact, and the actual conversation, the actual order, the actual payment confirmation, all of it happens in a green chat bubble on someone's phone.

This is the stack we wire on top of that reality. Real names, real numbers, real shape.

The five things that have to happen, every time

A typical SMB order, from first message to delivered:

1. Customer messages on WhatsApp. Sometimes a clear order. Sometimes a photo of a sample. Sometimes "kya rate hai bhai".
2. Somebody identifies the customer, looks up their history, their pricing, their preferences.
3. Confirms the order, quantity, delivery date.
4. Generates a proforma or invoice (GST-compliant) and shares it.
5. Shares a payment link, confirms receipt, dispatches.

If you have a salesperson, they do all five. If you're the owner, you do all five while running everything else. The work isn't hard. The volume is, and the context-switching cost is what burns people out.

A working AI stack absorbs steps 1 to 4 for the bulk of routine cases. Step 5 stays with a human. The roles that used to do the work shift to oversight and exceptions.

The pieces

What we wire, almost every time:

• WAHA or a similar WhatsApp gateway. Self-hostable, sits between WhatsApp's protocol and our backend. We built WhatSender on top of WAHA for bulk messaging; the same gateway powers inbound flows.
• An AI agent (Claude Sonnet 4.6 is the default in 2026; Haiku for triage). The agent reads inbound messages, classifies intent, looks up data, drafts responses, prepares actions.
• A workflow orchestrator (n8n self-hosted, almost always). Handles the glue, the schedules, the retries.
• Tally integration for invoice generation and GST.
• Razorpay (or PayU) for payment links and webhooks.
• A Postgres database with customer, product, order, and conversation tables.
• An audit log table for every agent decision.

Total monthly run cost for an SMB doing 500-2000 orders a month: ₹16,000 to ₹40,000 all-in, depending on volume and how much of the LLM budget is on cheap-tier triage.

What the agent does end-to-end

Here's the shape of a real intake flow:

[Inbound WhatsApp] → WAHA → n8n trigger
  → Agent (Haiku) classifies: order? inquiry? complaint? other?
  → If order:
    → Agent (Sonnet) extracts items, quantities, delivery date
    → Look up customer by phone number
    → Look up products by fuzzy match
    → Check inventory
    → Draft invoice in Tally
    → Draft payment link in Razorpay
    → Compose reply for human review
  → Office manager sees the draft, approves or edits
  → On approval: send WhatsApp message, push invoice, share link

The whole flow runs in seconds. The human approval at the end is what makes the system safe. Without it, an LLM misread on quantity could ship 10x what was wanted.

Two things matter operationally.

One: the agent's draft is the artefact. The human edits the draft, doesn't write from scratch. That's a 5-second action, not a 5-minute action. Volume becomes manageable.

Two: the audit log is real. Every classification, every extraction, every tool call is logged with the inputs. When an order is wrong, we can replay exactly what the agent saw and decide whether to tighten the prompt, add an example, or escalate that case type sooner.

The Hindi-and-English problem, solved

A specific thing nobody outside India seems to mention. Indian SMB customers don't message in English. They message in Hindi, in Hinglish, in their local language. They send voice notes. They send photos of paper chits. The agent has to handle all of it.

In 2026 this is fine. Claude and GPT both handle Indian languages well, and voice notes flow through Whisper or the vendor's voice transcription before hitting the agent. Hinglish ("2 piece dena bhai by Friday") is handled cleanly when there are examples in the prompt.

The trick is to put a few Hindi and Hinglish examples in the prompt. Without them, the model defaults to English-mode and misclassifies. With them, accuracy on a 50-case test goes from ~70% to ~95%.

TRAI rules, which you should not ignore

WhatsApp business messaging in India is regulated. TRAI has specific rules around marketing, opt-in/opt-out, and unsolicited messaging. WhatsApp's own Business API enforces a stricter version. Practical implications:

• Outbound marketing requires template messages that are pre-approved by WhatsApp. You can't blast a free-form ad.
• Two-way conversations (initiated by the customer) are much freer.
• Opt-out has to work. If the customer says "STOP" or equivalent, you stop. Encode this as a hard rule, not a prompt suggestion.
• The 24-hour window matters. After a customer messages you, you have 24 hours to respond outside the template rules. After that, you're back to templates.

These rules favour the inbound-first, conversational-first stack we've described. They make pure outbound spam expensive and slow, which is broadly a good thing.

The Indian SMB that wins on WhatsApp in 2026 isn't the one with the slickest outbound funnel. It's the one whose inbound reply comes back in 90 seconds with the right invoice attached.

What we built, what we learned

WhatSender is our bulk-messaging platform on top of WAHA. It runs in production for several clients, handles campaign scheduling, contact-list management, and multi-worker delivery. We learned three things building it.

WAHA itself is robust if you respect rate limits. We've had sessions running for months without intervention. The failure mode when it does fail is usually a phone-side issue, not a WAHA issue.

Delivery tracking is a real product. Customers want to see who got the message, who read it, who replied. Without that view, the campaign feels like a black box. We built it.

Multi-worker architecture matters at volume. A single Node process choked at around 200 messages/hour. Splitting into workers with PM2 and a shared queue scaled cleanly to several thousand per hour.

That's not the same as the inbound-AI stack described above, but the gateway is the same and the operational lessons transfer.

The replacing-three-roles claim

Specific: at a typical 30-person textile or trading SMB, before automation:

• One office manager handling WhatsApp inbound full-time.
• One junior accountant generating invoices in Tally and sharing them.
• One follow-up person chasing payments and dispatches.

After:

• The office manager reviews and approves agent drafts. Maybe 2 hours a day of focused review instead of 8 hours of context-switching.
• The accountant focuses on month-end and exceptions.
• The follow-up person handles the agent's escalations and the genuinely awkward conversations.

The headcount doesn't necessarily go down. The output capacity per head goes up 2-3x, the stress goes down, and the owner stops being on WhatsApp at 11 p.m. That last one is, in our experience, the metric the owner cares about most.

The honest summary

WhatsApp is not going away as the Indian SMB sales surface. The right move isn't to replace it. The right move is to put a competent AI underneath it, route everything through there, and free your human team to do the parts only humans can.

The stack is well-known, the pieces are cheap, and the build is a few weeks. The hard part isn't the technology. It's the discipline to keep the agent's actions visible and the audit logs honest.

The customer is on WhatsApp. The question is whether your sales operation is on WhatsApp with them, or three apps away.