Learning how to train AI on company data has become a common question for businesses that want to use generative AI in customer service. In practice, almost no business needs to train a model: it needs to give the AI the right documents and clear instructions.
The word “train” blends three different techniques. OpenAI’s documentation on accuracy separates what context solves, such as proprietary, up-to-date facts, from what requires changing the model’s behavior, such as format and tone.
This guide explains the three approaches, when to use each one, how to prepare the data, what data protection law requires, how to test with real questions, how to correct, how to measure, what it costs in tokens, and a step-by-step plan to put the AI on WhatsApp.
Key Takeaways
- “Training” an AI can mean fine-tuning, a knowledge base checked at answer time (RAG), or instructions with examples and tools. In customer service, the last two solve almost everything.
- Quality depends on clean data, written in the customer’s words, with no personal data the task does not require.
- The AI does not learn on its own: it improves through tests with real questions and corrections made by people on the team.
What It Means to Train AI on Company Data
When a manager asks how to train an AI, they almost always want it to answer with the company’s prices, timelines, and rules. That goal has very different paths in cost, time, and upkeep, and picking the wrong one tends to delay the project.
Three Meanings for the Same Word
AI models for chatbots come ready-made, trained on large volumes of text. From there, what businesses call training is usually one of these three things, with very different effort and results:
- Fine-tuning: continuing the model’s training with your own examples to change how it answers.
- Knowledge base with RAG: storing documents the AI checks before answering, without changing the model.
- Instructions, examples, and tools: telling it how to act, showing model answers, and allowing actions, such as checking a calendar.
Anyone researching how to train an AI to serve customers finds the three approaches mixed together in tutorials. Separating them avoids spending months on fine-tuning when the problem was just an outdated document or a poorly written instruction.
Why You Almost Never Need to Touch the Model
OpenAI’s model optimization guide states that prompt work may be all a use case needs. Fine-tuning appears as an option for some cases, not as a starting point.
The market backs this up. Anthropic’s glossary says the Claude API does not offer fine-tuning today. In October 2026, OpenAI’s documentation said its fine-tuning platform was being wound down and no longer accepted new users.
Anyone comparing the best AI tools available notices that models change often. The company’s documents and rules stay valid, which is why they tend to be the most durable investment.
Fine-Tuning: When Changing the Model Makes Sense
Fine-tuning exists and has legitimate uses. The point is to understand what it solves, how much work it takes, and why it is rarely the answer for customer service.
What Fine-Tuning Changes
Anthropic’s glossary defines fine-tuning as continuing to train a pretrained model with additional data, so it starts to mimic the patterns of that dataset, such as a domain, a task, or a writing style.
OpenAI suggests changing the model when it produces inconsistent results, with the wrong format, an unsuitable tone, or reasoning it fails to follow. When proprietary or up-to-date knowledge is missing, the recommendation is to optimize the context.
Cost, Number of Examples, and Upkeep
OpenAI’s supervised fine-tuning guide requires at least 10 examples and reports improvements with 50 to 100. Google Cloud’s documentation suggests thinking in terms of 100 examples or more and recommends starting with prompts.
Each example is a question paired with the ideal answer, reviewed by someone. If a price or policy changes, the tuned model keeps the old version until a new tuning cycle, which makes upkeep more expensive.
Google adds that tuning data needs to reflect the distribution, format, and context the model will see in production. In other words, a few examples made up by the team do not replace real conversations reviewed with care.
Knowledge Base and RAG: Context Instead of Training
The second approach leaves the model alone. The business organizes its documents, and the AI retrieves the relevant excerpts before each answer. This is what most people mean when they talk about training AI on their own data.
How the AI Checks the Documents
According to Anthropic, RAG combines information retrieval with text generation: data is retrieved at question time and added to the model’s context. The guide to an AI knowledge base covers what to put in it.
Google Cloud calls linking an answer to verifiable sources grounding and says it reduces the chances of the model making up content. When the search finds nothing, the agent should admit it does not know.
Why a Knowledge Base Wins for Prices, Timelines, and Policies
Prices, timelines, and policies change. In a knowledge base, you edit the document and the next answer already uses the new version, with no new training cycle to wait for or pay for.
In ConverZAP, the AI agent for customer support uses the consultar_base (search knowledge base) tool to look things up before answering. The search matches words, not meaning, and is built for Portuguese text, which changes how documents should be written.
In practice, it pays to repeat synonyms in the same excerpt: “price,” “cost,” and “how much” near the price table, “exchange” and “return” next to the policy. The AI reads the excerpt it finds and writes the answer in a conversational tone.
Instructions, Examples, and Tools
The third approach defines behavior without touching the model. Instructions set tone and limits, examples show the expected answer, and tools let the AI carry out tasks, such as saving data or booking a slot.
Instructions and Examples in the Prompt
Instructions cover what the knowledge base does not: tone of voice, what never to promise, and when to hand off to a person. They also need to hold up against manipulation attempts, such as customer messages that try to override the agent’s rules.
Anthropic’s prompting guides recommend 3 to 5 examples, similar to real use and varied enough to cover hard cases. The idea, also described by OpenAI, is to show the desired tone instead of only explaining it.
Tools That Make the AI Act
Anthropic describes tool use as the model’s ability to call functions the business defines, deciding when to use them based on the request. That is how an AI goes from just answering to actually solving.
An AI SDR agent on WhatsApp can save lead data in the WhatsApp CRM, and a WhatsApp appointment scheduling agent can book a slot during the conversation. The guide to an AI agent integrated with a CRM shows this kind of action in detail.
When to Use Each Way to Train AI
The choice starts with the right question: is the AI missing knowledge, behavior, or the ability to act? Each answer points to an approach, and the table below sums up the most common customer service cases.
Decision Table
OpenAI frames the decision on two axes: context, when the model does not know the information, and behavior, when it knows but answers inconsistently. The table applies that logic to common business situations:
| Need | Recommended approach | Reason |
|---|---|---|
| Answer prices, timelines, and policies | Knowledge base (RAG) | Changes often and needs a source |
| Speak in the brand’s tone | Instructions and examples | Quick adjustment, no training |
| Book appointments or save data | Tools | Requires an action in another system |
| Very rigid format at high volume | Fine-tuning | Repetitive, stable behavior |
The Most Common Combination in Customer Service
In most customer service cases, the combination is a knowledge base, instructions with examples, and tools. OpenAI notes that the techniques add up: fine-tuning can come later if tests reveal a behavior problem.
An AI agent platform for WhatsApp brings these pieces together in setup screens. The business writes, tests, and publishes without coding, and technical help is only needed for specific integrations with in-house systems.
How to Prepare Company Data
Good data matters more than a better model. A confusing document produces confusing answers, and an outdated table produces promises the team later has to walk back with the customer, often with a discount or refund.
FAQs, Policies, and Product Tables
In most businesses, three sources cover a large share of the questions that arrive on WhatsApp. Each one should live in a separate document, with a clear title and the date of the last review:
- FAQs: hours, address, payment, delivery, and other short, repeated questions.
- Policies: exchanges, cancellations, warranties, and refunds, with deadlines and exceptions written out in full.
- Product tables: name, description, price, and terms, with a header on every column.
In ConverZAP, this content lives in Brand & Knowledge, which holds documents in PDF, TXT, Markdown, or CSV, plus products, policies, and FAQs. Clean text works better than a scanned image or a spreadsheet with no headers.
Cleanup and the Customer’s Words
Before uploading any file, delete old versions, expired promotions, and duplicate passages. Two tables with different prices make the AI pick one of them, and not always the right one.
Because ConverZAP’s search matches words, documents should use the customer’s terms. Someone who asks “what’s the total” or “can I pay in installments” needs to find those expressions in the right excerpt.
Privacy and Data Protection: What Not to Send to the AI
Anything that goes into the knowledge base can show up in an answer. So the question is not only what helps the AI answer, but also what no customer should ever receive.
Data That Stays Out
Brazil’s data protection law (LGPD) (in Portuguese) defines sensitive data as information about health, racial origin, religion, political opinion, sex life, biometrics, and genetics linked to a person. Similar rules apply under GDPR in the EU. This data has no place in customer service documents.
Beyond sensitive data, other information should also stay out of the documents, because it does not help answer customers and greatly increases the damage if an excerpt shows up in the wrong answer:
- Customer data: lists with names, phone numbers, tax IDs (such as Brazil’s CPF), or purchase history.
- Credentials: passwords, access tokens, and the company’s bank details.
- Internal documents: employment contracts, salaries, and financial figures that are of no use to the customer.
Minimization and Accountability
The LGPD requires limiting processing to the minimum necessary and adopting security measures against unauthorized access. If the task is answering product questions, no personal data needs to go into the knowledge base.
The technology radar on generative AI from ANPD, Brazil’s data protection authority (in Portuguese), warns that attachments sent to these systems may contain personal and sensitive data. According to the study, anyone who shares third-party data with the AI may, depending on the context, be considered a processing agent.
Testing, Corrections, and Metrics
Training an AI does not end when the documents are uploaded. The improvement loop has three parts: test with real questions, correct what went wrong, and measure to find out whether the change helped.
Tests With Real Questions
The best test set comes from the inbox: 30 to 50 real questions, with the expected answer next to each. Anthropic advises that tests mirror real use and include hard, ambiguous, or off-topic cases.
The set should also include questions the AI must not answer, such as requests for discounts outside the price list. In those cases, the right move is a handoff, and the guide to human and AI customer service shows how to split that work.
The set should run before going live and after every change to a document, an instruction, or the model. OpenAI describes the same loop: measure a baseline, adjust, test again with data similar to real data, and repeat.
Iterating From Corrections
Every error has a source. A wrong price is fixed in the document, a wrong tone in the instruction, and a missing answer with a new FAQ item. Fixing it in the right place avoids patches that break other answers.
In ConverZAP, correction is manual: in the simulator, the team fixes the wrong answer, and it goes into that agent’s FAQ, with a limit of 40 items. There is no automatic learning: the AI only changes when someone corrects it.
Metrics to Track
Three simple indicators show, week by week, whether the work of training the AI is paying off. They do not replace the customer service metrics the team already tracks; they add to them:
- Accuracy on the test set: how many fixed questions get the expected answer.
- “Let me check” answers: they reveal gaps in the knowledge base that need a document.
- Handoffs to people: the reason for each one shows what still needs to be taught.
Costs: Tokens Explained Simply
AI models are billed by tokens, not by message. Understanding this unit helps forecast cost and explains why sending only the right excerpt is cheaper than sending the whole document.
What a Token Is
A token is a piece of text: a short word, part of a long word, or a symbol. According to the Gemini documentation, a token is about four characters, and 100 tokens equal roughly 60 to 80 English words.
Anthropic estimates about 3.5 English characters per token in Claude and notes that the ratio varies by language. In Portuguese or other languages, the ratio may differ, so it is worth measuring with the company’s real texts.
What Drives the Bill Up
Google says the cost of a call depends, in part, on input and output tokens. Input includes instructions, the chat history, and knowledge base excerpts; output is the answer the AI writes.
That is why lean instructions and short excerpts lower the cost of each answer. Fine-tuning adds another bill: OpenAI charges for training and also for using the tuned model, on top of the team’s time to build the examples.
To estimate monthly spend, a business can start from its conversation volume and the average length of each one, always checking the model provider’s official price list, because per-token prices change often.
How to Train AI on Company Data: Step by Step
The plan below works with any tool and requires no coding. The examples use ConverZAP, but the logic is the same on any AI agent platform for WhatsApp.
From Research to the First Version
- List the 30 to 50 most common questions in the inbox.
- Gather FAQs, policies, and tables, one topic per document.
- Clean out old versions and remove personal data.
- Write instructions on tone, limits, and handoff.
In ConverZAP, the website is imported once, at sign-up, as a starting point, and then the team reviews and fills in the gaps. The free Test the AI with your website tool shows how an agent would answer using only what is written on the site.
From the Simulator to WhatsApp
- Run the test questions in the simulator.
- Fix each error at the right source.
- Activate the agent on the official number.
- Review the gaps every week.
Activation uses the official WhatsApp Business API, and the guide on how to create AI agents for WhatsApp details the setup. The help center walks through every agent screen.
Common Mistakes When Training AI
The costliest mistakes are not technical. They come from wrong expectations about how to train an AI and from skipping simple steps, such as cleaning up documents and testing before turning the agent on.
Strategy Mistakes
The first mistake is starting with fine-tuning to teach prices and policies, which change. The second is expecting the AI to learn on its own from conversations, when improvement depends on corrections made by the team.
Another mistake is treating the AI as an isolated project. The same clear texts that answer on WhatsApp can also help the business show up in ChatGPT, since consistent information serves both people and search engines.
Operational Mistakes
In operations, the mistakes repeat: duplicate documents with different prices, personal data forgotten in spreadsheets, an agent turned on without tests, and no one responsible for reviewing the gaps every week.
ConverZAP, a Tech Provider verified by Meta, organizes this work in the company’s AI business brain: documents, products, policies, and FAQs that the agents check before answering on WhatsApp. You can try it with your own content in a free trial.
Frequently Asked Questions
In most cases, you do not need to train the model: gather FAQs, policies, and tables into clean documents, write instructions, and test with real questions. The Test the AI with your website tool shows a first result based on your own site.
No. Fine-tuning changes the model with dozens or hundreds of examples and suits stable behavior, such as format. A knowledge base and instructions provide context without changing the model, and OpenAI points to context when proprietary or up-to-date information is missing.
Yes. Agent platforms let you upload documents, write instructions, and test in a simulator with no code. In ConverZAP, activation runs on the official WhatsApp Business API, and integrations with the company’s in-house systems may need technical help.
Not in ConverZAP. There is no automatic learning: the team corrects the answer in the simulator, and the correction goes into that agent’s FAQ, with a limit of 40 items. That way, the business controls what changes in the answers and when.
Customers’ personal data, sensitive data, passwords, and internal documents that do not help answer questions. Brazil’s data protection law (LGPD) calls for using the minimum necessary, and the ANPD study on generative AI warns about personal data in attachments sent to these systems.
With a knowledge base and instructions, the main cost is model usage, billed by input and output tokens, plus the team’s time. Fine-tuning adds the cost of training and of building many reviewed examples.
