CEA Circular is a driving school in Bogotá, Colombia, that sells driver's license courses, and almost all of its demand arrives on WhatsApp. In 2026 we deployed a dedicated WhatsApp AI advisor for a driving school on the WebMinds WhatsApp AI platform. It answers prices, requirements, schedules, payment options and the new CALE theory exam in about eight seconds, sends the right flyer or reel, turns every chat into a structured lead tagged by source, and hands the conversation to a human advisor the moment a student is ready to enroll. A full admin panel sits behind it. It went live with real students in October 2026. This article explains how we built it; for the screens and gallery, 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. CEA Circular is one of several WhatsApp AI assistants we have built for Colombian companies.
Key facts
- Client: CEA Circular, a driving school selling driver's license courses in Bogotá, Colombia.
- Speed: ~8 s median AI reply time, measured on real production traffic in October 2026.
- Accuracy test: 40/40 on a real-question test battery built from 24 exported sales chats and 32 transcribed voice notes.
- Scope: 8 systems in production: agent, RAG, leads, human handoff, alerts, attribution, automations and admin panel.
- Security: 14/14 tables under row-level security, with secrets readable only by the server, column by column.
- Handover: a 49-page user manual, 45 versioned migrations and 53 API endpoints.
The problem: one advisor, one channel, contradictory information
Almost all of CEA Circular's demand arrives on WhatsApp, and it was answered by hand by a single sales advisor. The questions repeat: price per license category, required documents, course length, schedules, payment options and the new CALE theory exam. Worse, the official information was scattered and contradictory. There were three versions of the opening hours, different phone numbers on the flyers, and prices that only existed inside images.
In short:
- WhatsApp is the main sales channel, handled one chat at a time by one advisor.
- Most inquiries repeat the same questions about price, documents, duration, schedules and payment.
- Leads arriving at night, on Sundays and on holidays went cold or went to a competitor.
- Business knowledge lived in flyers, voice notes and the advisor's head, with conflicting versions.
Prospective students expect to resolve these questions on WhatsApp, the same app where they already talk to everyone else. Someone comparing schools on a Sunday evening should not have to wait until Monday for a price.
Why a generic chatbot was not the answer
The fastest path would have been a generic bot loaded with a price list. That would have failed here, for three reasons:
- There was no single source of truth. If the bot were trained on the flyers, it would repeat the contradictions. The knowledge had to be rebuilt from the real operation and confirmed by the owner.
- Honesty is a hard requirement. A driving school must not invent prices, schedules or license promises, and must never advise a student to misstate information in any official step of the process.
- The handoff matters more than the bot. The moment a student is ready to enroll, a person needs to step in. If the AI says "an advisor will contact you," a human must actually be notified.
Our AI & Automation work is built around those constraints, and the full deployment, from admin panel to website widget, is delivered by our Web & App Development team.
What we built: a 24/7 WhatsApp AI advisor
The chat window is the tip. Underneath is a conversational sales backend: a webhook with more than fifteen decision branches, row-level security on every table, vector search, scheduled jobs on two schedulers, five external integrations and up to five AI calls per message. By effort, roughly two thirds of the work is backend. It is a dedicated deployment with CEA Circular's brand, and the client owns the code.
We read the operation, not a brochure
We started from 24 real sales chats and 32 voice notes. They surfaced the conflicting hours, prices and phone numbers. We turned them into 39 questions for the owner, seven knowledge documents and a 40-question test the agent passes in full. Personal data never entered the index.
The AI advisor
A Gemini agent on the official WhatsApp Cloud API answers prices, requirements, payment options and the CALE exam, and sends the right flyer or reel from the catalog. The knowledge base uses retrieval-augmented generation (RAG) on pgvector, with 768-dimension embeddings and an HNSW index, so every answer is grounded in the owner-confirmed documents rather than in the model's general knowledge.
Honest by rule
Domain guardrails stop the agent from inventing prices, schedules or license promises. An explicit honesty rule keeps it from ever advising a student to misstate information in any official step of the process.
Lead engine and source attribution
Every chat becomes a structured lead with license category and stage, tagged by channel: website, social, QR, Google or Meta ads. wa.me links cannot carry hidden parameters, so each channel injects a short source tag into the prefilled text. The webhook reads it on the first message and always strips it before the AI or the history sees it.
If the AI promises a human, a human gets notified
One-click takeover pauses the AI until the advisor hands the conversation back. Escalation goes out by WhatsApp template and email. A future-tense promise detector covers the case where the model writes "an advisor will contact you" but forgets the escalation flag: it fires the escalation anyway, with per-conversation cooldowns and an audit trail.
Secrets the API can never return
Credentials live in a single-row settings table. Column-level privileges in Postgres mean neither admins nor viewers can read tokens through the public API; only the server role can.
Admin panel and automations
The panel is role-based, with analytics, wa.me links and QR codes per channel, a website widget with live preview, hourly close-out of inactive chats with an approved template (running on GitHub Actions), five branded transactional emails and a 49-page user manual.
How we deliver: Hybrid AI Coding
A senior engineer owns the architecture, the decisions with the client and every trade-off, while AI agents handle implementation under direction. An independent end-to-end QA pass gates every change before it ships. The stack is strict TypeScript and Next.js, Supabase with pgvector, Gemini, the WhatsApp Cloud API, Resend, Vercel and GitHub Actions.
Do your students and customers write at night and on Sundays? We can show you what a 24/7 advisor would look like for your business. Get your free assessment.
Timeline
- April 2026: proposal, and a dedicated deployment created on the WebMinds platform.
- August 20, 2026: human takeover, inactivity close-out, source attribution, branded email, analytics and legal pages; domain live.
- September 24, 2026: official WhatsApp number connected; training from 24 real chats (40/40); functional QA 16/16.
- September 25–28, 2026: website widget with live preview; owner-confirmed prices, hours and payments; five reels.
- October 5, 2026: delivery QA, security package, approved Meta templates and the 49-page manual.
- October 2026: live with real students, tuning with production traffic, ongoing.
Results: every inquiry answered, at any hour
The CEA Circular case study documents three outcomes:
- Every inquiry answered, at any hour. Prices, requirements, payment options and the new CALE exam are answered in about eight seconds, including nights, Sundays and holidays, when the advisor is off.
- WhatsApp becomes a pipeline, not a pile. Each chat lands as a structured lead with license category and source channel, so the team knows what students want and where they come from.
- AI does the first mile, people close. One-click takeover and guaranteed escalation mean the advisor steps in exactly when a student is ready to enroll, never left waiting on a promise the bot made.
CEA Circular runs on the WebMinds WhatsApp AI platform, the same platform behind the B2B IT assistant in the Netthplus case study. For a WhatsApp sales assistant built for a manufacturer, see the Industrias Qatar case study. For schools and academies, our education industry page covers more of what we build.
FAQ
What can a WhatsApp AI advisor answer for a driving school?
For CEA Circular, it answers prices per license category, required documents, course length, schedules, payment options and questions about the CALE theory exam, and it sends the matching flyer or reel.
Can a WhatsApp AI agent hand a conversation to a human?
Yes. The advisor can take over with one click, which pauses the AI until they hand it back. Escalation also fires automatically by WhatsApp template and email, including when the model promises that an advisor will make contact.
How do you stop the AI from inventing prices or schedules?
The agent answers from a knowledge base confirmed by the owner, and domain guardrails stop it from inventing prices, schedules or license promises. Before launch it passed 40 of 40 questions built from real chats and voice notes.
How fast does it answer?
The median AI reply time is about eight seconds, measured on real production traffic in October 2026.
Can I see which channel each lead came from?
Yes. Each channel (website, social, QR, Google or Meta ads) gets its own wa.me link with a short source tag in the prefilled text. The webhook reads it on the first message, tags the lead and strips the tag before the AI sees it.
Is it available 24/7?
Yes. The assistant answers at night, on Sundays and on holidays, and inactive chats are closed out every hour with an approved template.
Talk to us
If your school, academy or service business lives on WhatsApp and one person cannot answer everything, we can help you build an advisor that never clocks out and always knows when to call a human. Get your free assessment, or book a call. WebMinds is an AI and software development company based in Atlanta, Georgia, working across the US and Latin America, in English and Spanish.