In this article (19 sections)
In short: for an AI agent to answer with your company's information, you connect your documents (price lists, policies, manuals, FAQs). When it replies, it searches those documents first and answers only with what it finds, citing where it came from. That way it doesn't make up details, and if it can't find an answer, it says so or passes the question to a person. The quality of the agent depends more on your documents being organized and up to date than on which AI model you use.
Why a general-purpose assistant isn't enough
A general AI assistant knows a lot about the world but nothing about your company. It doesn't know your prices, your delivery times, your return policy, or how to troubleshoot the problem your customers ask about most. If you ask it, it can give an answer that sounds reasonable and is wrong.
For a business, that's the main risk: a wrong answer delivered with confidence. So what matters is not that the agent "knows," but that it checks your information and sticks to it.
How it works, without the jargon
The mechanism has four steps:
- You load your material. Documents, spreadsheets, web pages or FAQs. This is the agent's "library."
- The customer asks. Through WhatsApp, the website or email.
- The agent searches the library. It finds the passages that deal with the topic, the same way a person would search a folder.
- It answers with what it found. If the information isn't there, it says so and offers to connect the customer with a person.
This technique is known as searching your own documents. AI platforms offer it as a feature. For example, the OpenAI documentation on file search describes a tool that lets the model search files you upload and return answers with references to the source files. The Anthropic documentation on citations describes a similar approach: the model's answers can be grounded in your documents and return the exact passages that support each claim, so they can be verified.
What material to load
Type of information | Examples | Watch out for |
|---|---|---|
Products and services | Catalog, spec sheets, features | Old versions sitting alongside current ones |
Prices and terms | Price lists, current promotions | Deciding who updates them and how often |
Policies | Returns, warranties, delivery times | Writing them in plain language |
Processes | How a request works, what documents are required | One step per line, no ambiguity |
Frequently asked questions | The 30 questions that come up most | Answering the way your best salesperson would |
Support | Fixes for common problems | Stating when to escalate to a person |
Information that shouldn't be there
Don't load customers' personal data, contracts with confidential clauses, or internal information you wouldn't want any customer to receive. The practical rule: if you wouldn't say it to any customer who asks, it doesn't belong in the agent's public library. Access rules and backups belong in the system's design from day one.
What makes an agent reliable
It answers only from your documents
This is the central instruction. If a price isn't in the material, the agent doesn't estimate it. It says it doesn't have that information and offers to hand off.
It cites its source
When the agent names the document it took the answer from, your team can verify it and fix the document if it's out of date. It's a control tool, not just a transparency feature.
It knows when to hand off
Some questions shouldn't be answered with a text: a complaint, a negotiation, a case that doesn't fit any document. The agent recognizes those situations and passes the conversation to a person with a summary. A broader guide to how these limits are designed is in AI agents for business: what they actually automate.
It gets updated when the business changes
An agent is only as good as the latest update to its documents. If a price changed and nobody updated the library, the agent quotes the old price. Assign one person to own the material and a routine for keeping it current.
Where the agent talks
The most used channels are WhatsApp, the chat on your website, and email. WhatsApp has its own platform rules: according to the official Meta documentation, within the 24-hour window after the customer's last message you can send free-form replies, and outside it only template messages. For an agent that answers questions, nearly all replies happen inside the window. For choosing a WhatsApp provider, look for a comparison of WhatsApp Business API providers before you commit.
A concrete example
A company that sells HVAC equipment gets the same questions every day: which unit fits a room of about 215 square feet (20 square meters), what installation costs, what the warranty covers, and how long delivery takes. If those answers are in clear documents, the agent replies at any hour. When the customer asks for a technician visit or disputes a quote, the agent collects the details and alerts a salesperson. The team stops repeating the same answers and focuses on sales and complex cases.
What it costs and how long it takes
Atomix's price for an agent connected to your company's information, with its contact channel and handoff rules, starts at USD 3,000 for the agent, plus monthly maintenance from USD 300. It starts with a free working demo in 24 to 48 hours, built on a sample of your material, and most projects reach production in under a month. The AI model usage and WhatsApp messages are variable costs paid separately, based on volume. Our page about AI agents for business shows more use cases.
How to prepare your information before you start
- Gather the documents your team uses today to answer questions.
- Remove old and duplicate versions.
- List the 20 most frequent questions with the correct answer.
- Define which questions the agent shouldn't answer and who they go to.
- Name one person responsible for keeping the material up to date.
Frequently asked questions
Can the agent make up answers?
Every AI model can make mistakes. The risk drops when you require the agent to answer only from your material, cite its source, and hand off when it can't find the answer. It's also worth reviewing conversations during the first few weeks to fix documents and rules.
Are my documents exposed or used to train models?
It depends on the AI provider and the plan you choose. Before choosing, review their data usage terms. It's a decision made when designing the agent and it should be documented.
What document formats are accepted?
Usually text files, PDFs, spreadsheets and web pages. Scanned documents that are images need an extra step to extract the text. This gets checked in the demo with your real material.
How much work is it to maintain?
Little, if there's a routine: updating documents when something changes and reviewing the conversations where the agent handed off or didn't know the answer. Those conversations show what's missing from the library.
Can it serve customers in several languages?
Yes. Current models answer in several languages from documents written in one. If your customers come from different countries, test real questions in each language before launching.
If you'd like to try it with a sample of your own material, you can book a call.