Building an AI agent that writes into your CRM: architecture and costs
The architecture behind our agents: knowledge base, guardrails, CRM writing, model usage accounting — and what it costs to run per month.
The agent on this site answers questions about services and prices, collects a contact when the visitor asks for one, and creates a lead in the CRM with the conversation attached. The same architecture runs for our clients. Here is what is inside and what it costs to operate.
Architecture
Browser to server, never browser to model. The page sends the message to our server endpoint. The server holds the model key, applies limits, records the exchange, calls the model and streams the reply back. A key that reaches the browser is a key anyone can spend.
Knowledge base, not memory. The agent answers from a set of facts generated from the same data files as the site: services, prices, packages, statuses, process, published cases. Before each reply the server selects the most relevant chunks within a token budget. When a change is made on the site, the agent knows it after the next deployment.
Guardrails in the prompt and in code. The prompt lists what the agent may state and what it may not: no vendor licence prices, no discounts, no promises of results, no client names, no legal or tax advice. In code, prompt-injection patterns are detected and flagged, and a request for a human is handed off immediately.
One tool. The model can call a single function, register_lead, once the visitor has given a name and a phone. The server validates the phone, stores the lead, notifies the team in Telegram and creates the CRM lead with the dialogue in the timeline.
Limits
Per address: 30 messages an hour, 80 a day, a two-second pause, 1,500 characters per message. Per day: a budget of one dollar. When the budget is spent the agent does not go offline — it answers from prepared blocks built from the same data and offers the calculator or a contact. The director gets a notification.
What it costs to run
A message costs roughly 6,000 input tokens and 300 output tokens. At the prices we use for planning that is about half a cent per message, or under six dollars for a hundred ten-message conversations. Implementation of an agent for a client starts at $750 plus model usage; the calculator on this site prices it from the expected number of conversations.
What we do not automate
Negotiation, discounts, contract questions and complaints. The agent’s job is to answer well and hand over cleanly; the sale is still made by a person.