What does an AI chatbot cost to run?

Running cost is driven by conversation volume and how much context each answer needs, since both feed the per-token model price. Retrieval keeps it lower than sending whole documents on every request, and caching common questions removes repeat calls entirely. Build cost is separate and depends on how many sources the bot has to read from and how many channels it serves.

Can it work with our existing website?

Yes. The chatbot is added to an existing site or application rather than requiring a rebuild — the retrieval layer and API sit alongside what you already run, and the interface is embedded into the current front end.

What stops it giving wrong answers?

Answers are assembled from passages the retrieval step actually returned from your content, and the system prompt constrains the model to those passages. When retrieval finds nothing relevant, the bot says so and offers escalation instead of generating a plausible-sounding guess.

AI chatbot development servicesService — AI Chatbot Development

Chatbots that answer from your own content rather than guessing — retrieval-backed, with a defined escalation path and guardrails around what they are allowed to say.

What a production chatbot actually needs

A demo chatbot is a prompt and an API key. A production one has to answer from your material, admit when it does not know, hand over to a person cleanly, and be measurable afterwards.

Retrieval over your own documents, product data or knowledge base
A defined answer when the retrieval finds nothing — not an invented one
Escalation to a human, with the conversation history attached
Guardrails on topic, tone and what the bot may commit you to
Conversation logging, so you can see what people actually ask

Grounding it in your content

The bot's answers come from a retrieval layer built over your own material — documents, product catalogue, help centre, or whatever the source of truth already is. Content is chunked, embedded and stored in a vector database, and every answer is assembled from the passages the retrieval step actually returned. That is what stops the model inventing policy, pricing or features that do not exist.

Escalation, guardrails and tone

Every deployment defines what the bot may not do. Topics outside scope get a fixed response rather than an improvised one. Anything that touches pricing, contractual commitments or a frustrated customer routes to a person, with the transcript attached so they are not starting cold. Tone is configured against your existing writing, not left to the model's defaults.

Where it runs

The same retrieval layer serves whichever channel your customers are already in.

Embedded in your website or web app
WhatsApp, for markets where that is the default support channel
Slack or Microsoft Teams, for internal knowledge assistants
A REST API, where the chat interface is already built

Where I’ve done this

Shipped projects that use the work described above.

Common questions

What does an AI chatbot cost to run?: Running cost is driven by conversation volume and how much context each answer needs, since both feed the per-token model price. Retrieval keeps it lower than sending whole documents on every request, and caching common questions removes repeat calls entirely. Build cost is separate and depends on how many sources the bot has to read from and how many channels it serves.
Can it work with our existing website?: Yes. The chatbot is added to an existing site or application rather than requiring a rebuild — the retrieval layer and API sit alongside what you already run, and the interface is embedded into the current front end.
What stops it giving wrong answers?: Answers are assembled from passages the retrieval step actually returned from your content, and the system prompt constrains the model to those passages. When retrieval finds nothing relevant, the bot says so and offers escalation instead of generating a plausible-sounding guess.