Industrias Qatar is a Colombian manufacturer and installer of tow hitches and bike racks, with two locations in Bogotá and Chía, and its sales happen on WhatsApp. In March 2026 we started building a WhatsApp AI sales agent for a manufacturer whose catalog has a different product for nearly every vehicle make and model. The assistant answers every customer in seconds, at any hour, in natural Colombian Spanish. It understands text, voice notes and photos, sends the exact product photo for the customer's vehicle, captures the lead, books the installation and hands the chat to a person when it matters. It went live five weeks after kickoff. This article covers the problem, the architecture and the decisions behind it. For the visual walkthrough, read the full case study.
We are WebMinds, an AI and software development company based in Atlanta, Georgia, that builds AI agents, custom software and business automation for companies in the United States and Latin America. Industrias Qatar is one of several WhatsApp AI agents we have built for Colombian businesses.
Key facts
- Client: Industrias Qatar, a manufacturer and installer of tow hitches and bike racks in Bogotá and Chía, Colombia.
- Time to live operation: 5 weeks, from kickoff on March 11, 2026 to live operation on April 14, 2026, with continuous improvements since.
- Availability: 24/7 sales in natural Spanish across text, voice notes and customer photos, with Colombian holidays and business hours built in.
- AI per message: 6 AI sub-agents (reply generation, lead extraction and four independent escalation detectors) and 4 parallel vector searches over documents, catalog images, PDFs and videos.
- Pipeline: a single WhatsApp webhook runs a 16-step pipeline on every inbound message.
- Security and scale: 13/13 tables with row-level security, 44 security policies, 31 versioned migrations, 47 API endpoints and 24 screens in the operations panel.
The problem: a vehicle-by-vehicle catalog, sold one chat at a time
Industrias Qatar sells a product where fit is everything. Almost every vehicle make and model has its own tow hitch or bike rack reference, so the right answer depends on knowing exactly which car the customer drives. Before the project, every WhatsApp enquiry was handled by hand. An advisor had to identify the customer's vehicle, find the right reference, send the photo, quote, handle objections and book the installation, one message at a time and only during office hours.
That created four problems:
- Volume and repetition. Answering "does it fit my 2020 Hilux?" with the right photo and price is repetitive work, but it requires precision. A wrong answer means a wrong product.
- After hours. Messages arrive at any time. Every late reply is a lead that cools down or goes to a competitor.
- Consistency. The sales pitch (lifetime warranty, Colombian manufacturing, installation included) depended on which advisor answered.
- Lost information. Vehicle, product, value and appointment stayed buried in the chat and never fed a sales pipeline.
In Colombia and across Latin America, WhatsApp is where customers already write to businesses. Moving the conversation to a web form or an email address adds friction. So the goal was not to replace the channel but to make it work at any hour and at scale, without losing the precision that this catalog demands.
Why an off-the-shelf chatbot did not fit
Most WhatsApp chatbot tools are built around text FAQs and decision trees. This case needed several things those tools handle poorly:
- Visual answers tied to a real catalog. The right answer is often a photo of the exact hitch or rack for the customer's car, plus a PDF or video. It has to be the right file, every time.
- Inputs beyond text. Customers send voice notes and photos of their vehicle. The agent has to transcribe and see, not just read.
- Commitments the AI must not make. Installation and delivery dates are operational promises. A bot that invents them creates real problems for the team.
- Data that flows somewhere. Each conversation contains a lead (name, vehicle, product, estimated value, location, appointment) that belongs in a pipeline, not in a chat history.
- A knowledge base that lasts. Files uploaded to a hosted model API expire after 48 hours, which is unworkable for a catalog that must always be searchable.
That is why we built a dedicated platform instead of configuring a generic bot. It is part of our AI & Automation work, combined with the WhatsApp and email integrations that connect it to the team.
What we built: a WhatsApp AI sales agent for manufacturers
We delivered five systems as one operating platform at ai.industriasqatar.com. This is not a website with a chatbot attached. A single WhatsApp webhook orchestrates a 16-step pipeline on every inbound message: deduplication, transcription, vision, four parallel vector searches, generation, media resolution, logging, analytics and four escalation detectors. Roughly 60% of the work lives in the backend.
The WhatsApp AI agent
The agent answers 24/7 in natural Colombian Spanish. It understands text, voice notes (transcribed with Google Cloud speech-to-text) and the photos customers send, and it respects business hours and Colombian holidays. Generation, vision and embeddings run on Gemini, and messages travel over the official WhatsApp Cloud API.
A visual RAG knowledge base that does not expire
RAG (retrieval-augmented generation) means the model answers from documents it retrieves at query time instead of relying only on what it memorized during training. Here the knowledge base holds documents and a catalog of product images, embedded at 768 dimensions in pgvector with an HNSW index. Because hosted model files expire after 48 hours, we moved to our own embeddings: searchable indefinitely and refreshed whenever the catalog changes.
The bot sends the photo, not a promise
The model replies with media tags tied to real catalog IDs. The pipeline then resolves those tags into WhatsApp images, PDFs and videos, and strips any tag that points to an ID that does not exist. The customer gets the exact photo of the hitch or rack for their vehicle, and the model cannot send a file that is not in the catalog.
Automatic lead extraction into a forward-only pipeline
A separate AI pass extracts name, vehicle, product, estimated value, location and appointment from every conversation. Leads flow into a four-stage pipeline that only moves forward, with email alerts and appointment confirmations sent through Resend.
Human escalation through four independent signals
Four detectors decide when a person should step in: frustration or an explicit request for a human, the model's own flag, operational promises the bot must not make, and payment proofs. Alerts go to the team by WhatsApp and email. One detector is dedicated to installation or delivery dates. If the bot is about to commit to one, the team is alerted. The AI handles volume and people handle commitments.
Built for serverless and Meta retries
A full turn can take 15 to 30 seconds. Meta resends webhooks it considers unanswered, which on serverless hosting can produce duplicate replies, or none at all. Extended execution time and a three-state deduplication make sure retries never produce either outcome.
Never silent
If the model hits a rate limit, a safety filter or a timeout, the customer still gets a polite fallback that is stored and auditable. A monthly token ceiling cuts off gracefully instead of failing mid-conversation.
An operations panel the client runs
The panel has 24 screens: conversations in a WhatsApp Web–style view, a lead board, self-service training with a phone simulator, analytics, and editable personality, prompt, hours and branding. The team can test catalog changes in the simulator before any customer sees them.
How we shipped in five weeks: Hybrid AI Coding
Senior WebMinds engineering owns the architecture and every decision with the client, while AI agents implement under direction, with QA on every change. That delivery model is how the platform went live five weeks after kickoff. The stack is Next.js and TypeScript end to end, Supabase (Postgres, Auth, Storage) with pgvector, Gemini, the WhatsApp Cloud API, Google Cloud STT, Resend and Vercel.
Selling on WhatsApp by hand? We can map your catalog, your questions and your handoff rules in one call. Get your free assessment.
Timeline
- March 11, 2026: kickoff, technical specification and first presentation.
- March 17, 2026: RAG with pgvector, lead extraction, email and branding.
- March 27, 2026: visual catalog, image replies on WhatsApp and the phone simulator.
- April 14, 2026: live operation and signed delivery.
- April–May 2026: human escalation, WhatsApp profile, web widget, Click-to-WhatsApp campaign attribution and QR links.
- Ongoing: continuous improvements and catalog expansion.
Results: sales without office hours
These are the outcomes documented in the Industrias Qatar case study:
- Sales without office hours. The assistant answers, quotes, shows the exact product and books installations around the clock in natural Spanish, including nights, weekends and Colombian holidays.
- A complex catalog, mastered. A vehicle-by-vehicle catalog lives in a permanent vector knowledge base, and the bot sends the right photo of the hitch or rack for the customer's car. That used to mean an advisor searching by hand.
- Nothing lost, humans when it counts. Every conversation is logged, leads land in a sales pipeline automatically, and four AI detectors bring in a person for frustration, explicit requests, payment proofs or sensitive commitments.
We have built other WhatsApp AI sales desks too: for an IT consultancy in the Netthplus case study and for a cosmetic packaging manufacturer in the Esval case study. If your goal is sales automation more broadly, our sales automation use case shows where this fits.
FAQ
What is a WhatsApp AI sales agent for manufacturers?
It is an AI assistant connected to the official WhatsApp Cloud API that answers product questions, quotes, sends catalog media and captures leads for a manufacturer, grounded in that company's own catalog and documents. For Industrias Qatar, it identifies the customer's vehicle and sends the matching tow hitch or bike rack photo.
Can a WhatsApp AI agent hand a conversation to a human?
Yes. In this project four independent detectors trigger a handoff: frustration or an explicit request, the model's own flag, operational promises the bot must not make, and payment proofs. The team is alerted by WhatsApp and email.
Can the agent understand voice notes and photos?
Yes. Voice notes are transcribed with Google Cloud speech-to-text, and photos customers send are read with Gemini vision before the agent replies.
Will the bot promise installation dates it cannot keep?
It is designed not to. A dedicated detector catches any installation or delivery date the bot should not commit to and alerts the team, so people make the commitments.
How long does it take to launch a WhatsApp AI sales agent?
For Industrias Qatar, live operation came five weeks after kickoff (March 11 to April 14, 2026). Timelines depend on catalog size, integrations and how much of the platform already exists.
Does it work at night, on weekends and on holidays?
Yes. The agent answers 24/7 and has business hours and Colombian holidays built in, so it knows when the human team is available.
Talk to us
If your sales depend on WhatsApp and your catalog is too detailed for a generic bot, we can help you design an agent that answers precisely and knows when to call a person. Get your free assessment, or book a call with our team. WebMinds is based in Atlanta, Georgia, and works with companies across the US and Latin America, in English and Spanish.