The last generation of chatbots earned their reputation honestly. Rigid decision trees, no memory, and a dead end that told you to email support anyway. Language models removed that constraint — a modern assistant can understand a question phrased any way, hold the thread of a conversation, and answer from your actual documentation. The engineering that matters now is grounding, escalation and handover.

Who this is for

Businesses answering the same questions repeatedly across chat, WhatsApp and email. Companies whose support hours do not match their customers' time zones. Teams whose first-line queue is mostly questions already answered in writing somewhere.

Problems we solve

Answers that are plausible but wrong. A model with no grounding will invent a delivery window or a refund rule.
No exit to a human. Frustrated customers trapped in a loop with no visible escalation path.
Amnesia. A bot that asks for the order number it was given two messages ago.
Duplicate replies. Webhook retries that make the assistant answer the same message twice.
Handover with no context. An agent picking up a conversation and having to start again.

What we build

Assistants grounded in your documentation, policies and product data through retrieval
Deployment across WhatsApp Business, web chat widgets, Slack and Messenger from one core
Conversation memory in Redis, so context persists across a session and between sessions
Message deduplication, so a retried webhook never produces a second reply
Confidence-based escalation to a human, with the full transcript attached
Lead capture and ticket creation written straight into your CRM or helpdesk
Multilingual handling where your customers need it

How we work

We start from your real conversation history, because that is what tells us the actual distribution of questions rather than the ones you expect. The assistant is built to answer confidently within its grounded scope and hand over cleanly outside it — a bot that escalates well is more valuable than one that answers everything badly. Every conversation is logged and reviewable, so the knowledge base can be improved where the assistant struggled.

Technologies we use

WhatsApp Business Cloud API and embeddable web chat for channels; Claude, GPT, Gemini or self-hosted Ollama for generation; Qdrant or pgvector for grounding; Redis for memory and deduplication; n8n for orchestration where operations staff need to adjust routing themselves. We have a fully self-hosted reference build combining WhatsApp, web chat, Ollama, Qdrant and Redis — see our work for the architecture.

Business benefits

Common questions answered instantly, at any hour
Support staff freed for the cases that genuinely need judgement
Answers consistent with your documented policy every time
Conversation logs reveal exactly where your documentation is failing customers

Where it pays off

First-line customer support on WhatsApp and web chat
Order status, delivery and returns enquiries
Product and pricing questions during the buying decision
Appointment booking and qualification
Internal help desks for staff policy and IT questions

Common questions

Will it replace our support team?

Not in our experience, and we would be cautious of anyone promising it. It absorbs the repetitive share of the queue so your team handles the complex remainder. Escalation quality matters more than deflection rate.

How does it know our policies?

Through retrieval over your actual documents rather than instructions in a prompt. Update the document and the assistant's answer updates with it.

Can it run without sending data to a US cloud?

Yes. A self-hosted stack with Ollama and Qdrant keeps conversations entirely inside your infrastructure.

What about WhatsApp approval?

WhatsApp Business Cloud requires a verified business account and template approval for outbound messages. We handle the technical setup and tell you what Meta requires from you.

Want to see what your support queue would look like automated? Send us a week of anonymised transcripts.

AI Chatbots & Conversational AI