AI Chatbot Development
An assistant that knows your business, not the internet's vague impression of it. I build RAG chatbots grounded in your own documents, with an honest answer when they do not know.
The failure mode of a support chatbot is not being unable to answer. It is answering confidently and wrongly, at scale, in writing, to your customers. Everything below is arranged around preventing that — which is also why every one of these is built on retrieval-augmented generation rather than on a model fine-tuned once and hoped over.
What you get
- Retrieval-augmented generation over your own contentYour docs, product data, policies, pricing PDFs and past conversations are embedded into a vector database and searched at question time, so each answer comes from your material and can be traced back to the passage it came from.
- "I don't know" is a valid answerThe bot is built to decline and escalate outside its knowledge rather than improvise. Off-topic and adversarial prompts are handled too.
- Tested against real questionsAn evaluation set built from your actual inbox, re-run on every change, so an improvement to one answer cannot quietly break twenty others.
- Handoff that keeps the contextWhen a human takes over they get the full transcript and the customer's intent, so nobody is asked to repeat themselves.
- English-language, wherever your customers already areBuilt, tuned and evaluated for English-speaking audiences, and deployable to a website widget, WhatsApp, Messenger, Instagram, Slack, or your own app through an API.
My AI Chatbot Development process
Assemble the knowledge
Docs, FAQs, policies, product data, past tickets — cleaned and chunked, because retrieval quality is mostly a data-preparation problem.
Set the boundaries
Tone, what it must never say, what it must always escalate, and how it behaves when unsure.
Wire in the actions
Order lookup, booking, ticket creation, CRM writes — the things that make it useful rather than merely conversational.
Evaluate, then adjust
Score against the real-question set, fix the failures, re-run. Repeat until the numbers are good enough to put in front of customers.
Launch behind a safety net
Start on a share of traffic, watch resolution and escalation rates, then widen once it has earned it.
Tools I use for this
Chosen per project against your constraints and budget — never one stack forced onto every problem.
Questions I get asked about this
How do you stop it from making things up?
Three things, and none of them alone is enough. The model answers from retrieved passages of your own content rather than from memory, so there is a source behind each answer. It is instructed and tested to say it does not know and hand off, instead of filling a gap. And it is scored against a fixed set of real questions with known-correct answers before launch and after every change. Hallucination is not eliminated by prompt wording; it is held down by grounding and measured by evaluation.
Which model do you use, and am I locked into it?
Usually Claude or GPT for quality-sensitive conversation, and a smaller or open-weights model for high-volume classification and routing where the task is easy and the volume is what costs money. You are not locked in: the model sits behind an interface, so switching is a configuration change plus a re-run of the evaluation set, not a rebuild. That matters, because pricing and capability in this space change roughly every few months.
What does it cost to run each month?
Two numbers, and most quotes online conflate them. The token cost is what the model provider charges: a grounded support conversation runs from a fraction of a cent to a few cents depending on the model and how much context each answer needs, so a few thousand conversations a month is often single or low double-digit dollars. The all-in running cost adds hosting, a vector database if the bot uses retrieval, and whoever reviews escalations — which is why published figures for a small support bot land between roughly $300 and $800 a month once everything is counted. Output tokens dominate the model bill, so I cut it with caching, tight retrieval windows and routing easy questions to a cheaper model. You get a projection from your real volume before we build, billed to your own provider account so you can see it directly.
Can it hand over to a person?
Yes, and this is deliberately easy to trigger. Escalation fires on explicit request, on detected frustration, on any topic you have marked as human-only such as refunds or complaints, and whenever confidence is low. The human receives the full transcript, what the customer appears to be trying to do, and any records already looked up. A handoff that drops context is worse than no bot, because the customer has now explained themselves twice.
Build an assistant worth putting in front of customers
Send me the twenty questions your team answers most often. That is the whole starting point.
Start with a free discovery call