• Artificial Intelligence

AI on WhatsApp: Why India’s Largest B2B Platforms Run Intelligence on a Chat App

Published On: 23 September 2026.By .

WhatsApp AI is not a fallback for companies that cannot build a mobile app. For B2B India it is often the correct primary channel, and treating it as second best is how good products end up with poor adoption. This blog draws on two live Auriga IT deployments: a self-service bot serving 800,000 FMCG retailers, and a hybrid agent inbox running across 400 retail stores.

800K+
Retailers served on WhatsApp
10,000+
Distributors supported
400+
Retail stores connected
Zero
App downloads required

The 60-second version

  • The app-first assumption breaks in B2B India. Distributors, field agents and small retailers use WhatsApp daily but rarely open a vendor app.
  • There are two distinct architectures. A self-service bot that resolves without a human, and an agent-assist inbox that routes to a human with context attached. They solve different problems.
  • The AI layer is the easy part. The hard part is the live data connection and the business rule enforcement underneath it.
  • Meta's 24-hour window is a real constraint. Any deployment that does not handle it natively will fail silently in production.
  • Both models are running at scale. Badho.in and Ferns N Petals, both built by Auriga IT, are linked in full below.

A distributor in a tier-three town needs to know whether an order shipped. He has a smartphone. He has WhatsApp open. He does not have your app installed, and if he did, he would not remember the password.

This is the situation that most B2B platforms in India are actually operating in, and it is the reason the channel decision matters more than the model decision. A well-built AI assistant behind a login wall reaches a fraction of the people a modest one on WhatsApp reaches.

01
Why the app-first assumption breaks in B2B India

Consumer product thinking assumes the app is the destination. Build a good enough experience and users will install it, open it, and return to it. In B2C India that assumption holds reasonably well. In B2B distribution it does not.

The people who need to interact with a distribution platform every day are not power users of software. They are retailers running a shop, distributors managing stock, and field agents moving between locations. Their phone is a working tool with limited storage and a handful of apps they actually use. WhatsApp is one of them. Your platform app, in most cases, is not.

App-first

  • Requires download, install and storage space
  • Login credentials to remember or reset
  • Training needed for first-time users
  • Updates require user action to install
  • Adoption drops sharply outside metro users
  • Support falls back to phone calls when the app is not opened

This is not an argument against building apps. Badho.in has an app, and it serves users who want a richer interface. The argument is narrower: for high-frequency, low-complexity queries from a distributed user base, the channel that requires no installation will always reach more people than the one that does.

02
What AI on WhatsApp actually means technically

A WhatsApp AI deployment is three connected layers, and confusion about which layer does what is the most common source of project failure.

User on WhatsApp Where is order 4521? Dispatched. ETA 2 hours. No app. No login. Phone number is the identity. Answer in seconds 1. Conversational layer LLM + WhatsApp Business API Interprets intent, drafts reply 2. Retrieval layer Hasura GraphQL Fetches live operational data 3. Business rule layer Custom mutations + SQL Enforces what the AI may write Order management Status, dispatch, ETA CRM / wallet Balance, credit, payouts Schemes / incidents Promotions, tickets Why the split matters The LLM never queries the database directly. Retrieval is a controlled set of reusable functions. Writes pass through rule checks before committing. Scale comes from this.

Three layers, clearly separated. The conversational layer never touches the database directly, which is what makes the system safe to scale.

Why the layer separation is not optional

The tempting shortcut is to let the language model generate queries against the database. It works in a demo and fails in production, because the model has no concept of business constraints. It will happily update a record that should have been locked, or return data a particular user should not see.

In the Badho.AI build, this was handled by putting every data operation behind custom GraphQL mutations backed by SQL functions. The AI calls a named function. The function enforces the rule. If the operation violates a constraint, it does not execute, regardless of what the model intended.

03
Two architectures, two different problems

Most teams evaluating WhatsApp AI assume there is one pattern. There are two, and choosing the wrong one produces either an expensive bot nobody trusts or an agent team drowning in queries a machine should have handled.

A

Self-service bot: resolve without a human

Suits high-volume, transactional, factually answerable queries. The user asks, the system retrieves, the conversation ends. No human is involved and none needs to be. This is the Badho.in model, serving 800,000 retailers and 10,000 distributors.

  • Query types: order status, wallet balance, payout status, active schemes, ticket creation, purchase order lists.
  • Identification: registered phone number, no OTP and no login screen.
  • Success measure: percentage of queries resolved without agent escalation.
  • Failure mode: the bot confidently answers a question it should have escalated. Guardrails matter more than coverage.
800K+
retailers served
10,000+
distributors
1,000+
brands connected
24/7
availability
Stack
LLM engine WhatsApp Business API Hasura GraphQL Custom GraphQL mutations SQL business rules
B

Agent-assist inbox: route to a human with context

Suits queries where the outcome matters emotionally or commercially, and where a wrong automated answer is worse than a slower human one. WhatsApp becomes one channel feeding a unified agent workspace. This is the Ferns N Petals model, running across 400 stores and 100 countries.

  • Query types: failed deliveries, damaged goods, refund disputes, address changes, loyalty complaints.
  • What the AI does: classifies intent and emotional weight, pulls order context into the conversation, suggests a response for the agent.
  • Routing tiers: automation for routine, AI-assisted agent for moderate, senior human for high-stakes.
  • Failure mode: treating every query as equal. A delivery time question and a failed anniversary gift should never take the same path.
5 to 1
channels unified
400+
stores connected
100+
countries served
3-tier
routing logic
Stack
CygnusAlpha Cygnus Reach WhatsApp self-service bot AI triage engine Order data sync 24-hour window handling

04
Which model fits which problem

If your situation isChooseBecause
High query volume, factual answersSelf-serviceHuman involvement adds cost without adding accuracy
Emotionally weighted outcomesAgent-assistA wrong automated reply damages the relationship permanently
Users spread across many tiers and regionsSelf-serviceConsistency matters more than nuance at that spread
High-value transactions per customerAgent-assistThe cost of a senior agent is small against the order value
Both patterns present in the same businessHybrid routingRoute by intent and complexity, not by channel

Most mature deployments end up hybrid. The decision is made per query type, not once for the whole system.

05
Five mistakes we have seen repeatedly

Letting the model query the database

It demos beautifully and breaks in production. The model has no concept of business constraints. Every data operation needs to sit behind a named function that enforces rules independently of what the model intended.

Ignoring the 24-hour window

Meta only permits free-form replies within 24 hours of the user's last message. Deployments that do not handle this natively produce agents drafting replies that silently fail to send, with no error surfaced.

Building coverage before guardrails

A bot that answers 95 percent of queries but confidently invents the other 5 percent is worse than one that answers 70 percent and escalates cleanly. Users stop trusting the channel after one bad answer.

Adding OTP verification

The registered phone number is already the identity on WhatsApp. Adding an OTP step reintroduces exactly the login friction the channel was chosen to avoid, and drop-off climbs immediately.

Treating it as a support-only channel

Once users trust the channel, they will try to transact on it. Order placement, scheme queries and ticket creation all belong there. Restricting it to FAQ answers leaves most of the value unclaimed.

No conversation tracking

Informal WhatsApp support loses conversations. If the deployment does not create a structured, trackable record per query, it reproduces the original problem with an AI layer painted on top.

A bot that answers 70 percent of queries and escalates cleanly beats one that answers 95 percent and invents the rest. Trust is the constraint, not coverage.

06
How to tell if your operation is ready

Before scoping a WhatsApp AI build, these four questions determine whether the project will move quickly or stall in the data layer.

Can your data be reached by API right now?

If order status and account balance live in systems without an API layer, that work comes first. The AI cannot answer questions about data it cannot retrieve, and this is where most timelines slip.

Do you know your top ten query types?

Pull the last month of support conversations and count. If the top ten account for most of the volume and are factually answerable, the self-service model will produce fast results.

Is your business logic written down anywhere?

Rules that live only in the heads of experienced staff cannot be enforced in code. The rule layer is only as good as the documented constraints you can hand to it.

Do you have a verified WhatsApp Business number?

Verification and template approval through Meta take time and sit on the critical path. Starting this in parallel with development rather than after it saves weeks.

07
Questions about WhatsApp AI

What is a WhatsApp AI chatbot and how does it work for B2B platforms?

A WhatsApp AI chatbot connects an LLM layer to the WhatsApp Business API and to a company's live operational data. When a user messages the verified business number, the AI interprets the query, retrieves live data through an API layer such as Hasura GraphQL, and returns an accurate answer in seconds. Badho.AI, built by Auriga IT for Badho.in, serves 800,000 retailers this way with no app required and no login step.

Why do Indian B2B companies choose WhatsApp over a mobile app?

Distributors, field agents and small retailers in India already use WhatsApp daily but often do not install or regularly open a vendor app. Deploying support on WhatsApp removes the download, login and training barrier entirely. Adoption is immediate because the channel is already familiar and already open on the device, which matters far more than interface richness for high-frequency transactional queries.

What is the difference between a self-service bot and an agent-assist inbox?

A self-service bot resolves the query entirely without a human, which suits high-volume transactional questions such as order status or wallet balance. An agent-assist inbox routes the conversation to a human with AI-suggested responses and full context attached, which suits complex or emotionally weighted cases. Badho.in uses the self-service model and Ferns N Petals uses a hybrid of both, with routing decided by intent and complexity rather than by channel.

How does WhatsApp AI access live order and account data safely?

The AI layer connects to the operational database through a controlled API layer rather than querying tables directly. In the Badho.AI build, custom GraphQL mutations backed by SQL functions enforce business rules on every create, update and delete operation, so the AI cannot write data that violates platform constraints even when acting autonomously. Reusable retrieval functions handle reads, which keeps behaviour predictable as volume grows.

What is the WhatsApp 24-hour messaging window and why does it matter?

Meta's policy allows a business to send free-form messages to a user only within 24 hours of that user's last message. Outside that window, only pre-approved template messages can be sent. Any production WhatsApp deployment must handle this natively in the agent workflow, otherwise agents unknowingly draft replies that cannot be delivered and the failure is invisible until a customer complains.

Can a WhatsApp AI system scale to hundreds of thousands of users?

Yes, if the architecture separates the conversational layer from the data retrieval layer. Badho.AI was designed for a base of 800,000 retailers and 10,000 distributors, using reusable data retrieval functions so that the AI fetches structured platform data through simple calls rather than bespoke queries per conversation. Scale problems in these systems almost always originate in the data layer, not the model.

Thinking about WhatsApp AI for your platform?

Auriga IT has built both patterns in production, for a B2B platform serving 800,000 retailers and for a retail brand operating across 400 stores and 100 countries.

Related content

Stay Close to What We’re Building

Get insights on product engineering, AI, and real-world technology decisions shaping modern businesses.

Suman Yubraj
Suman Yubraj
Suman Yubraj is a Technical Writer at Auriga IT with a background in computer science and content writing. He translates complex technical topics into clear, accessible content for developers and business audiences alike.
Go to Top