Five choices that set the cost of a customer support agent

agentes-ia

When a small business owner asks how much an AI support agent costs, I do not answer with a number. The cost depends on five decisions that do not show up in a demo.

Each decision can raise or lower the final cost. Several get attention only when the system breaks in production (during real use).

1. Channel and volume

An agent that answers through a website chat has one cost profile. An agent that answers through WhatsApp, with audio, during business hours has another.

Volume matters more than the channel. With 30 questions a day, a small local model can give acceptable answers at almost no cost. It runs on your own equipment. With 3,000 questions, it may not answer fast enough. Each question that moves to a cloud model also uses tokens (units of text that can be billed).

The first decision is to measure how many conversations arrive each day, through which channels, and at what times. Without those numbers, any quote is made up.

2. What goes into the knowledge base and who maintains it

RAG is a system that searches a knowledge base before writing an answer. Answer quality depends on the content it uses. The base may have an outdated manual, a generic FAQ, and a crooked scanned PDF. Then the agent answers confidently and gets it wrong with conviction.

Someone needs to decide what goes into the base, keep the files current, and test the answers. That is ongoing human work. It is not model training, or teaching the program with examples. It is looking after the content. In the systems I have built, document intake is automatic, but a person still checks quality. Without someone in that role, the real cost is the wrong answers the agent will give.

3. A model for each task and moving from local to cloud

Not every question needs the most expensive model. "What are your business hours?" can be answered by a small local model. "Does product X cover case Y under rule Z?" may need a larger cloud model.

The pattern I use is routing: the local model tries first. If confidence falls below a minimum score, the question moves to the cloud. The same applies when it needs reasoning the local model has been shown not to handle.

That means not every question pays the full price. Only the hard questions do. At high volume, that difference can make the bill fit the budget.

4. When the agent says "I don't know" and calls a person

RAG always returns a first result. Cosine similarity compares vectors (lists of numbers that represent the text chunks). It only sorts the results; it does not decide whether a chunk is relevant. As I wrote in another article, a good system knows when the result does not help.

In cost terms, an agent that never says "I don't know" gives wrong answers with a source. One that says it too often frustrates the customer. Between them is a product decision: below what similarity score does the agent call a person?

That choice sets how many human support staff you still need. Human support is the most expensive line on the spreadsheet.

5. Records and cost per conversation

Without activity records, you do not know what each conversation costs. Without the cost, you do not know whether the system pays for itself.

Minimum metrics:

• tokens used per conversation;
• share of conversations resolved without a human;
• p95 latency (the response time met by 95% of conversations);
• cost in your currency per conversation.

In week 1, the cost is low. By week 12, it may have tripled if nobody checks. That could happen because someone uploaded a 400-page PDF that the agent reads in full for every question.

Where it breaks

A support agent is not a project that ends. It is a product that needs maintenance. If you treat it as a project, ship it, test it, and close it, the agent works for a while. That lasts until the knowledge base goes stale, volume doubles, or the cloud model provider changes its price. That happens.

If you want to explore this for your business, tell me about your context.