80% of incoming messages handled automatically. Deployed in 2 weeks. Near-zero operating cost.
That’s the summary. Now the story.
The Problem Nobody Talks About
If your team spends hours every day answering the same WhatsApp messages, you already know this pain.
A student asks about course pricing. You answer. Another student asks the same question ten minutes later. You answer again. Enrollment steps. Certification timelines. Schedule changes. The same twenty questions, hundreds of times per week.
That was life at CEOSH, an accredited awarding body for professional qualifications in the UK. Their small team was buried under WhatsApp messages from prospective and current students. Hundreds per week. The same questions, over and over. Every hour spent answering “What are the enrollment requirements?” was an hour stolen from student support, curriculum development, or growing the business.
They needed a solution. But not just any chatbot.
The Requirement That Changes Everything
Most chatbot projects go wrong right here.
The typical WhatsApp chatbot solution forces you onto a separate phone number. So now your students have two numbers: one for the bot, one for the real team. Conversations split across two threads. Students get confused. Some flat-out refuse to message the “bot number.”
CEOSH didn’t want that.
They wanted the bot living on their existing WhatsApp Business number. Same number, same thread. The student texts like they always have. Sometimes the AI answers. Sometimes a human does. The student can’t tell.
That’s the killer feature. It’s also the thing most chatbot vendors can’t pull off.
We built it in two weeks.
How It Actually Works
A student texts the CEOSH WhatsApp number. Same number they’ve always used.
The bot reads the message and checks it against CEOSH’s knowledge base: course catalogs, pricing, enrollment procedures, FAQs. If it can answer, the student gets an instant, accurate reply. No waiting hours for a team member to get back to them.
If the bot can’t answer, maybe the question is too specific or it falls outside the knowledge base, the bot tells the student a team member will follow up shortly. At the same time, it fires off a notification to a staff member on their personal WhatsApp with the student’s phone number, their original message, and full context.
The team member opens their WhatsApp Business app, picks up the conversation, and replies directly. They can pause the bot for that specific student with a simple command, handle things personally, then re-enable the bot when they’re done.
One continuous thread. One number. No confusion.
The Architecture That Makes It Work
Engineers love to overcomplicate this stuff.
When most people hear “AI chatbot,” they think vector databases, RAG pipelines, embedding models, retrieval chains. Months of development. Thousands in infrastructure costs.
We did none of that.
CEOSH’s knowledge base is roughly 5,000 words. Course information, pricing, enrollment steps, FAQs. That entire knowledge base gets injected directly into the system prompt. Every single message the bot processes has the full context of everything it needs to know. No database. No vector store. No retrieval pipeline. No embeddings. Just a well-written prompt with clear constraints.
The model is Google Gemini 2.0 Flash, one of the cheapest LLMs available. We’re talking fractions of a cent per message. At hundreds of messages per week, the AI cost is essentially a rounding error.
Stateless by design. Each message is processed independently. No conversation history to manage, no session state to lose if the server restarts. For an FAQ-style bot, statefulness adds complexity without meaningful benefit.
The entire stack: Python, FastAPI, PydanticAI for structured outputs, and the Meta WhatsApp Cloud API. Hosted on Railway. One-click deploy.
And because the LLM is configured as a single string, swapping from Gemini Flash to Claude or GPT-4 is a one-line change. No refactoring. The model is a configuration detail, not an architectural decision.
The Constraints That Prevent Disaster
An AI chatbot that makes things up is worse than no chatbot at all.
The system prompt explicitly instructs the model to only use information from the provided knowledge base. It can’t fabricate answers. It can’t respond to questions outside its scope. If a student asks something unrelated to CEOSH, the bot doesn’t engage. This is constraint-based AI. The model is powerful, but the prompt keeps it on a tight leash. It answers what it knows. It escalates what it doesn’t. It never guesses.
Zero hallucination risk on topics outside the knowledge base, because the bot simply won’t go there.
The Admin Controls
The team manages the bot with simple WhatsApp commands from their own phones:
/pause 923001234567pauses the bot for a specific student so the team can take over/resume 923001234567re-enables the bot for that student/statusshows which conversations are currently paused
No dashboard to log into. No admin panel to build. The control plane is WhatsApp itself. The team was already living in WhatsApp all day, so we met them where they were.
The Results
After deployment:
- 80% of incoming messages handled automatically, without human intervention
- Instant responses to student inquiries, no more waiting hours for a reply
- Zero missed queries. Every unanswered question triggers a notification with full context
- Smooth experience for students. One number, completely unaware of the bot/human handoff
- Near-zero operating cost. No database, no vector store, and Gemini Flash is dirt cheap
- Live in 2 weeks, not months
The team got their time back. Students got faster answers.
What This Means for Your Business
If you’re running a business where your team answers the same WhatsApp questions over and over, whether that’s an education company, a clinic, an e-commerce brand, a real estate agency, this exact architecture works.
You don’t need complex AI infrastructure. You don’t need a six-month project. You don’t need a separate bot number that confuses your customers.
You need a focused knowledge base, the right model, a clean handoff system, and someone who knows how to put it together. That’s what we do.
If your team is drowning in WhatsApp messages and you want to see what automation looks like for your specific situation, start a conversation with us. We’ll tell you exactly what’s possible, what it costs, and how fast we can ship it.