Table of Contents
- Key takeaways
- What is chatbot training?
- How are AI chatbots trained?
- Pretraining: how the language model learns
- Retrieval: how the chatbot learns your content
- Instructions and examples: how the chatbot learns behavior
- Intent training: how scripted and NLU chatbots learn
- Fine-tuning: when the model itself is retrained
- Which chatbot training method should you use?
- What data do you need to train a chatbot?
- How do you train a chatbot to give empathetic responses?
- Best practices for empathetic chatbot responses
- What the research says about humanlike chatbots and angry customers
- Example instructions and replies
- How do you test a trained chatbot?
- Worked example: scoring a 40-question test
- How do you keep a trained chatbot accurate after launch?
- Where SiteGPT fits
- How long does it take to train a chatbot?
- Worked example: hours for a small retrieval chatbot
- Common chatbot training mistakes
- Frequently asked questions
- Basics
- What is chatbot training?
- How are AI chatbots trained?
- What is a zero-training AI chatbot?
- What does a chatbot trainer do?
- Methods and data
- Do you need machine learning skills to train a chatbot?
- How much data do you need to train a chatbot?
- What is the best way to train a chatbot?
- Is chatbot training the same as fine-tuning?
- Tone and handoff
- What are the best practices for empathetic chatbot replies?
- Should a chatbot sound human?
- What is the best way to add human experts to an AI chatbot?
- Testing and upkeep
- How do you know when a chatbot is trained well enough to launch?
- How often should you retrain a chatbot?
- How long does it take to train a chatbot?
- Sources
Created by
Do not index
Created time
Sep 30, 2023 01:27 AM
Chatbot training is the work of giving a chatbot the knowledge, instructions, and examples it needs to answer correctly, then testing it and keeping it current. For most business chatbots in 2026 that means indexing your own content for a pretrained language model, and it does not mean training a model from scratch.
TL;DR
- "Training" covers five different jobs: pretraining a model, retrieval over your content, written instructions, intent training, and fine-tuning. Only the first and last change the model itself.
- Most business chatbots are trained by retrieval. The platform indexes your pages and files, and a pretrained model answers from them.
- The data that matters most is real customer questions, the pages that answer them, and past transcripts with personal data removed.
- Empathy is trained through instructions and example replies. The bot acknowledges the problem once, gives the fix, and offers a person.
- Training is finished only when the bot passes a fixed test set. It needs a rerun after every content or instruction change.
Key takeaways
Stage | What you do | What you need | How to check it |
1. Pick the method | Choose retrieval, intent training, or fine-tuning | A list of what the bot must do | Every task on the list has a method |
2. Prepare the data | Collect questions, source pages, and transcripts | 50 to 100 real customer questions | Every question maps to one source page |
3. Set behavior and tone | Write instructions and example replies | Scope, tone rules, handoff rules | Replies to upset test messages follow the rules |
4. Test before launch | Run a fixed question set and score it | Questions with known correct outcomes | Pass rate per question type |
5. Maintain after launch | Read transcripts, fill gaps, re-index | A weekly review slot and an owner | Days since the last re-index, unanswered questions per week |
What is chatbot training?
Chatbot training is every step that changes what a chatbot knows or how it behaves. It includes the data the bot learns from, the rules it follows, the tests it must pass, and the upkeep after launch.
The word "training" causes confusion because it describes two different things. One is machine learning, where a model's internal numbers change as it processes data. The other is setup work, where a team loads content and writes instructions for a model that is already trained.
Most teams only do the second kind. A company that adds an AI chatbot to its website does not build a language model. It supplies content, instructions, and tests to a model that a provider such as OpenAI has already trained.
This guide explains the methods, the data, the tone work, the testing, and the upkeep. The hands-on steps for loading your own content are in how to train an AI chatbot on your own data. If you have not yet picked a chatbot type or a build route, start with how to make a chatbot.
How are AI chatbots trained?
AI chatbots are trained in up to five ways, and a single chatbot often uses several of them at once. The model provider handles the first. The team that deploys the chatbot handles the rest.
Method | What changes | Data needed | Who does it | When it fits |
Pretraining | The model's weights | Very large text collections | The model provider | Never a task for the business deploying a chatbot |
Retrieval (RAG) | A searchable index of your content | Your pages, files, and help articles | You or a no-code platform | Answering questions from your own content |
Instructions and examples | The prompt sent with every question | Written rules and sample replies | You | Scope, tone, empathy, and handoff rules |
Intent training | A classifier that maps messages to intents | 10 to 20 example phrases per intent | A bot builder or developer | Fixed tasks such as booking or order lookup |
Fine-tuning | A copy of the model's weights | At least 10 example conversations, 50 recommended to start | Developers | A fixed style or output format that instructions cannot hold |
Pretraining: how the language model learns
Pretraining is the stage where a provider teaches a large language model to predict text. OpenAI describes the process for the models behind ChatGPT. The model reads a large volume of text and learns to predict the next most likely word, one word at a time.
OpenAI lists three sources of training information: content that is publicly available on the internet, content from third-party partners, and content that users, human trainers, and researchers provide. The model does not keep copies of that text. Its parameters, also called weights, shift slightly as it learns patterns.
A business never repeats this stage. It is the reason a new chatbot can already write fluent sentences in many languages before it has seen one page of your content.
Retrieval: how the chatbot learns your content
Retrieval is the method where the chatbot looks up your content at answer time. The platform splits your pages and files into passages and stores them in an index. For each question it finds the closest passages and passes them to the model, and the model writes the answer from them.
The technique is called retrieval-augmented generation, or RAG. A 2020 paper by Patrick Lewis and colleagues introduced the term. It describes models that pair a pretrained generator with a retrieval index.
The model's weights do not change in this method. Updating the chatbot means updating the content and rebuilding the index. The full explanation is in what is RAG.
Instructions and examples: how the chatbot learns behavior
Instructions are written rules sent to the model with every question. They set the bot's scope, answer length, tone, and what to do when it has no answer. Sample replies inside the instructions show the model the pattern to copy.
This is the fastest kind of training to change. Editing one sentence changes every later reply. It is also where empathy rules and handoff rules live.
Intent training: how scripted and NLU chatbots learn
Intent training teaches a classifier to recognize what a user wants from example phrases. Each intent, such as "track my order", gets a list of ways people might say it. The bot matches a new message to the closest intent and runs the flow attached to it.
Google's Dialogflow documentation recommends at least 10 to 20 training phrases per intent, depending on the intent's complexity. Its built-in machine learning expands the list with similar phrases, so nobody has to write every possible wording.
Intent training suits fixed tasks with a small number of outcomes. It struggles with open questions, because every new topic needs a new intent and new phrases. The language side of this method is covered in the guide to chatbot NLP.
Fine-tuning: when the model itself is retrained
Fine-tuning continues a pretrained model's training on your own examples, which produces a private copy with adjusted weights. OpenAI's documentation sets a minimum of 10 examples and recommends starting with 50 well-written ones. The examples must be in JSONL format, with one conversation per line.
Fine-tuning changes how a model writes. It is a poor way to teach facts, because a price or policy baked into the weights cannot be edited later without another training run.
Access is also shrinking. OpenAI's documentation states that it is winding down its fine-tuning platform and that new users can no longer access it. The remaining options for OpenAI models are compared in how to train ChatGPT on your own data.
Which chatbot training method should you use?
Use retrieval when the chatbot must answer questions from your content, intent training when it must run fixed tasks, and instructions for tone and rules. Fine-tuning comes last, and only for a style or format problem that instructions did not fix.
If you need | Pick | Why |
Answers drawn from a website, help center, or documents | Retrieval | The content stays editable, and an update is a re-index |
A fixed sequence such as booking, order lookup, or an eligibility check | Intent training with scripted flows | Every step is predictable and can be audited |
A consistent tone, shorter answers, or empathy rules | Instructions and example replies | One edit changes every reply, with no retraining |
A strict output format that instructions keep breaking | Fine-tuning, if you already have access | The format becomes part of the model |
Open questions and a few fixed tasks in one bot | Retrieval plus scripted flows for the fixed tasks | Each method covers what the other cannot |
A chatbot with no developers on the team | A no-code retrieval platform | The platform handles indexing and retrieval |
Match the fix to the cause. A bot that states an old price has a content problem. More example phrases or a fine-tuning run will not fix it.
What data do you need to train a chatbot?
You need three kinds of data to train a business chatbot: the questions customers ask, the content that answers them, and examples of good replies. The amount matters less than how current and consistent it is.
Data type | Used for | Where it comes from | Watch for |
Real customer questions | Deciding scope and building the test set | Email, chat logs, site search, the customer service team | Questions nobody has written an answer for |
Source content | Retrieval answers | Website pages, help articles, policy documents, product files | Duplicates, outdated pages, and two pages that disagree |
Past transcripts | Finding gaps and writing example replies | Live chat and helpdesk history | Names, emails, order numbers, and other personal data |
Example replies | Instructions and tone | The best answers your team has written | Replies that promise refunds or exceptions |
Example phrases per intent | Intent training | Real wording from transcripts | Phrases that fit two intents at once |
Start with the questions. A list of 50 to 100 real customer questions shows what the chatbot must know. Map each question to the one page that answers it. A question with no page is a content gap, and no training method fills it.
Transcripts are useful, and they need care. Raw transcripts contain personal data and one-off promises. Use them to find the common questions and the wording customers use. Rewrite the good answers as clean help content before loading anything.
A retrieval chatbot repeats what its sources say. If two pages give different return windows, the bot will state one of them with full confidence. The cleanup steps are in the sibling guide on training a chatbot on your own data, and the indexing side is explained in what is content ingestion.
How do you train a chatbot to give empathetic responses?
You train a chatbot to give empathetic responses by writing rules and example replies into its instructions, then testing those rules with upset messages. Empathy in a chatbot means the reply shows it read the problem and moves the customer toward a fix. It does not mean the bot claims to have feelings.
Best practices for empathetic chatbot responses
Seven practices cover most of the tone work.
- Acknowledge the specific problem once. Name what went wrong in the customer's own terms, such as a late order or a double charge. A generic line about understanding frustration reads as a script.
- Give the fix or the next step in the same reply. Acknowledgment with no information makes the customer ask again.
- Do not claim feelings. The bot can say the delay is a problem. It should not say it knows how the customer feels.
- Apologize once, and only when the company is at fault. Repeated apologies add length and no help.
- Match the tone to the stakes. Drop jokes, emoji, and exclamation marks for complaints, billing, health, and anything urgent.
- Never repeat the same answer. If the first reply did not help, the bot should try a different source or offer a person.
- Offer a person early when the customer is angry. A calm handoff with the transcript is more empathetic than a third attempt.
What the research says about humanlike chatbots and angry customers
A humanlike persona can backfire with angry customers. A 2022 Journal of Marketing paper by Crolic, Thomaz, Hadi, and Stephen ran five studies, including an analysis of real chats from an international telecommunications company. When customers arrived angry, a more humanlike chatbot lowered their satisfaction, their view of the company, and their intent to buy again.
The effect did not appear for customers who were not angry. The authors traced it to inflated expectations. A bot with a name and an avatar raises hopes that it can solve the problem, and an angry customer punishes it more when it cannot.
Nielsen Norman Group's usability research on chatbots points the same way. Participants in its 2018 study were annoyed by bots that repeated the same answer, and they rated an honest admission of not understanding above a wrong answer.
The practical rule is to put effort into solving the problem before adding personality. Persona and conversation flow have their own guide in chatbot design.
Example instructions and replies
The table shows three situations, a weak reply, and a reply that follows the practices above. Use pairs like these as examples inside the bot's instructions.
Situation | Weak reply | Better reply |
Late delivery | "Sorry for the inconvenience! Orders ship in 3 to 5 days." | "Your order is past the 5-day window, and that is a real delay. Here is how to check the tracking status. If it shows no movement, a team member can look into it now." |
Double charge | "I understand how you feel. Please see the billing policy." | "A double charge should not happen. Billing corrections need a person, so this conversation is going to the team with the details you gave." |
Question the bot cannot answer | "I didn't get that. Please rephrase." (three times) | "That is not covered in the help content. A team member can answer it. Would you like to be connected?" |
The same rules apply to people who answer chats and email. The human version is in how to respond to an angry customer.
How do you test a trained chatbot?
You test a trained chatbot by running a fixed set of questions with known correct outcomes and scoring each reply. The set must include questions the bot should answer, questions it should decline, and questions it should hand to a person.
Build the set from the real questions collected earlier. Write the expected outcome next to each one before running anything. Score each question type separately, because a single total hides the failures that cost the most.
Worked example: scoring a 40-question test
This example uses a hypothetical software company. The numbers are illustrative.
Question type | Questions | Expected outcome | Passed | Pass rate |
Answerable from content | 28 | Correct answer from the right source | 25 | 89% |
Not in the content | 6 | Bot declines and offers a person | 5 | 83% |
Should go to a person | 6 | Handoff with the transcript | 6 | 100% |
Total | 40 | All three outcomes | 36 | 90% |
The total is 36 out of 40, or 90%. The four failures are not equal. Three were answerable questions the bot got wrong, which points to missing or conflicting content. One was a question with no source that the bot answered anyway, and an invented answer is the most expensive failure.
The fixes follow from the failure type:
- For the three wrong answers, find the source page for each and correct or add it.
- For the invented answer, tighten the instruction to decline when no source matches.
- Rerun all 40 questions. Launch when the "not in the content" row reaches 6 out of 6.
The guardrails that reduce invented answers are covered in how to stop chatbot hallucinations. The numbers to track after launch are in chatbot best practices.
How do you keep a trained chatbot accurate after launch?
You keep a trained chatbot accurate by reviewing real conversations every week, fixing the content behind each failure, re-indexing when sources change, and rerunning the test set. Training does not end at launch, because prices, policies, and products keep changing.
Trigger | Action | Why |
A price, plan, or policy changes | Update the source page and re-index the same day | A retrieval bot keeps quoting the old page until the index is rebuilt |
Customers ask a question the bot declines | Write the missing page or answer, then re-index | Each declined question is a content gap |
The bot gives a wrong answer | Find the source passage and fix it | In a retrieval bot the cause is usually the content |
The instructions change | Rerun the full test set | A tone fix can break a refusal rule |
A week passes | Read 20 transcripts, starting with escalated ones | Problems show up in transcripts before they show up in totals |
Real transcripts are the best new training material. Each week, list the questions the bot declined or answered badly. Add the three most common to the content and to the test set. The test set then grows with the questions customers really ask.
Handoff is part of upkeep too. The best way to add human experts to an AI chatbot is to set clear escalation rules and pass the full transcript, so the expert does not start over. Tools differ on how much context transfers, and AI chatbots with human handoff compares them.
Where SiteGPT fits
SiteGPT is a no-code retrieval platform, so it covers the retrieval, instruction, and upkeep stages above. It has no visual flow canvas, so fixed multi-step flows need another tool.
- Content. It trains on website URLs, sitemaps, files, and raw text, plus Google Drive, Dropbox, OneDrive, Box, Notion, GitHub, and YouTube.
- Re-indexing. Refresh is manual on Starter. Growth refreshes monthly, Scale weekly, and Enterprise daily. Scale and Enterprise also scan the sitemap daily for new and removed URLs.
- Gap finding. Growth and above include a knowledge gaps report and a weekly email digest that lists the questions the chatbot could not answer.
- Handoff. Human support is available on any plan. By default the AI escalates when the visitor asks for a person, when the situation is urgent, or when the visitor is frustrated after an unhelpful answer. Custom escalation rules can be written in plain language.
Starter costs $59 a month billed monthly, or $39 a month billed yearly, with 4k GPT-4.1-mini messages a month and 1,000 pages. There is no free plan. The 7-day free trial requires a credit card.
How long does it take to train a chatbot?
The time to train a chatbot depends on the method and on the state of the content, and the indexing step is the shortest part. Preparing content and testing take most of the hours.
Five factors affect conversational AI training time:
- The method. Retrieval needs content and an index. Intent training needs phrases and flows for every task. Fine-tuning needs a reviewed set of examples and a training run.
- The amount of content. More pages take longer to gather and check.
- The condition of the content. Duplicates, outdated pages, and conflicts must be fixed before loading.
- The number of test rounds. Each failed round means a fix and a rerun.
- Review and sign-off. Legal or compliance review of the instructions adds calendar time.
Worked example: hours for a small retrieval chatbot
This estimate is for a hypothetical business with 60 customer questions. The per-task minutes are assumptions, so replace them with your own.
Task | Units | Minutes each | Total |
Map each question to a source page | 60 questions | 5 | 300 minutes (5 hours) |
Write or fix pages for the gaps found | 12 pages | 30 | 360 minutes (6 hours) |
Run and score the test set, two rounds | 80 replies | 2 | 160 minutes (2.7 hours) |
Total | 3 tasks | Varies | 820 minutes (13.7 hours) |
Under these assumptions the work is about 13.7 hours, or roughly two working days. Content work is 11 of those hours, which is 80% of the total. Cleaner content is the fastest way to shorten training.
Common chatbot training mistakes
The most common chatbot training mistakes come from treating training as a single upload. Five show up most often.
- Loading everything. Old blog posts, duplicate PDFs, and expired offers all become possible answers.
- Fixing facts with the wrong method. A wrong price is a content problem. Instructions and fine-tuning do not correct it.
- Loading raw transcripts. They carry personal data and promises made to one customer.
- Testing only the questions the bot can answer. The test must include questions it should decline and questions it should hand off.
- Stopping at launch. A bot nobody reviews drifts out of date as the business changes.
Frequently asked questions
Basics
What is chatbot training?
Chatbot training is the process of giving a chatbot knowledge, instructions, and examples, then testing it and keeping it current. For a business chatbot built on a pretrained model, the work is mostly content preparation, written rules, and testing.
How are AI chatbots trained?
AI chatbots are trained in stages. A provider pretrains a language model on large text collections. The business then adds its own content through retrieval, writes instructions for behavior, and tests the result. Some bots also use intent training or fine-tuning.
What is a zero-training AI chatbot?
A zero-training AI chatbot is a marketing term for a retrieval chatbot that needs no model training and no example phrases. It still needs accurate content, instructions, and testing. The model training is skipped, and the preparation is not.
What does a chatbot trainer do?
A chatbot trainer prepares the content and examples a chatbot learns from, writes its instructions, reviews transcripts, and fixes the causes of wrong answers. On a small team the role usually belongs to someone in customer service or documentation.
Methods and data
Do you need machine learning skills to train a chatbot?
No. Retrieval platforms handle the indexing and the model. The skills that matter are knowing the customers' questions, writing clear help content, and testing carefully. Machine learning skills are needed for fine-tuning and for custom retrieval code. If you plan to write that code yourself, the guide to machine learning in a chatbot explains each piece and prices the build.
How much data do you need to train a chatbot?
A retrieval chatbot needs enough content to answer the questions customers ask, which may be a few dozen pages. Intent training needs 10 to 20 example phrases per intent, according to Google's Dialogflow documentation. OpenAI recommends starting fine-tuning with 50 examples.
What is the best way to train a chatbot?
The best way to train a chatbot for customer questions is retrieval over clean, current content, with written instructions for tone and handoff, and a fixed test set. Start from real customer questions and add content only where a question has no answer.
Is chatbot training the same as fine-tuning?
No. Fine-tuning is one training method, and it changes the model's weights. Most business chatbot training changes no weights. It changes the content the bot retrieves and the instructions it follows.
Tone and handoff
What are the best practices for empathetic chatbot replies?
Acknowledge the specific problem once, give the fix or next step in the same reply, apologize only when the company is at fault, avoid claiming feelings, keep the tone plain for serious topics, never repeat an answer, and offer a person early when the customer is angry.
Should a chatbot sound human?
A chatbot should sound clear and polite, and it should say it is a bot. Research in the Journal of Marketing found that humanlike chatbots lowered satisfaction among customers who arrived angry. Solving the problem matters more than personality.
What is the best way to add human experts to an AI chatbot?
Set written escalation rules, show a visible option to reach a person, and pass the full transcript at handoff. The expert should see what the customer asked and what the bot answered.
Testing and upkeep
How do you know when a chatbot is trained well enough to launch?
A chatbot is ready when it passes a fixed test set at the level the team agreed before testing. The share of unanswerable questions it declines should be 100%, because an invented answer costs more than a refusal.
How often should you retrain a chatbot?
Re-index a retrieval chatbot whenever a price, plan, policy, or product page changes, and on a regular schedule in between. Rerun the test set after every re-index and every instruction change.
How long does it take to train a chatbot?
It depends on the content. The indexing step is short. Mapping questions to pages, fixing gaps, and testing take most of the time. A hypothetical 60-question project in this guide came to about 13.7 hours.
Sources
All sources were checked on October 11, 2026.
Written by
