An AI agent is not a chatbot with a polished answer. It is a program that takes a task, decides which tools to call, and works until the job is done: it finds data, updates the CRM, drafts a document. Here is what an agent is made of and where to start.
The problem
Companies "add ChatGPT", get generic answers, and give up. The model does not know your prices, customers, or policies, and it cannot do anything inside your systems.
Why it happens
A language model on its own can only generate text. To be useful it needs three things: access to your data, tools to take action, and clear limits on what it may and may not do.
What an agent is made of
- A model — GPT, Claude, or an open-source model on your own server when data cannot leave the company.
- Context — a knowledge base of documents, price lists, and policies, usually searched through a vector index (RAG).
- Tools — functions the agent can call: look up a customer in the CRM, create a task, calculate a price, send an email.
- Orchestration — step logic, retries on errors, and a cap on the number of actions.
- Oversight — a log of every step and human approval for risky operations.
How to start
- Pick one process with a clear outcome: triaging inbound requests, answering common questions, drafting proposals.
- List the data the agent needs and the actions it must be able to take.
- Build a prototype on real examples and measure quality: what share of tasks is completed without a human stepping in.
- Only then connect the agent to production systems and expand its toolset.
Technical details
Tools are described as functions with typed parameters. The model chooses which one to call; your code executes it and returns the result. Keep permissions tight: an agent that creates deals should not be able to delete customers.
const tools = {
findCustomer: { params: { phone: "string" }, run: crm.findByPhone },
createTask: { params: { customerId: "string", text: "string" }, run: crm.createTask },
};
Example
An agent for website inquiries reads the message, identifies the service, finds the customer in the CRM or creates a new record, estimates the budget from the price list, and assigns a task to a sales rep with a draft reply. The rep reviews and sends it — a couple of minutes per inquiry instead of a quarter of an hour.
Takeaway
AI becomes useful when the model can reach your company's data and tools. Start with one process, measure the result, and grow the agent step by step.