How to Train a Sales Chatbot on Your Past Conversations

You train a sales chatbot on past conversations in four steps. First, collect and clean your best sales exchanges. Second, remove personal data. Third, use those conversations to teach the model your tone and qualifying questions, through examples in its instructions, retrieval, fine-tuning or a mix of these. Fourth, test it against conversations it has never seen. Most of the work is in choosing and preparing the data and checking that you are allowed to use it. Running the model itself is the smaller part.

What "training" actually means for a sales chatbot

When vendors say a chatbot is "trained on your conversations", they can mean three different techniques. The difference matters for cost, accuracy and how easily you can change the bot later.

Examples in the prompt. A general model gets a written brief about your business plus a handful of real exchanges that show how you sell. Nothing about the model changes. This is the quickest option to set up and the easiest to adjust, but every example takes up space in every request.
Retrieval (RAG). Your past conversations, price lists and product notes go into a searchable store. For each message, the system looks up the most relevant material and passes it to the model. AWS guidance says retrieval can add the latest documents in a few minutes and can point to the source of its information. It also says the risk of made-up answers is lower.
Fine-tuning. The model is trained further on a set of your example exchanges, so your phrasing and structure become part of the model. AWS notes that this can take a few hours to days and may need a data scientist. It also notes that fine-tuned models do not point to their sources and may be more likely to invent answers to factual questions.

In practice, use fine-tuning or examples to teach how to talk, and retrieval to supply what is true. Your conversations show tone, the order of questions and how objections are handled. They are a poor source of facts, because prices, stock and offers in old chats go out of date. A bot fine-tuned on last year's emails will confidently quote last year's prices. AWS describes combining retrieval and fine-tuning in one system, and for a sales bot that combination is usually where you end up.

Step 1: Choose the right conversations

If you train on every conversation you have ever had, the bot learns your average performance, weak moments included. Choose your examples deliberately.

Collect from every channel where selling happens: website chat logs, email threads, WhatsApp or SMS exports, CRM notes and call transcripts. Written chat is closest to what the bot will do. Call transcripts are useful for objection handling but need more cleaning.
Label outcomes. Link each conversation to what happened next: booked a call, bought, went quiet or was disqualified. Conversations that led to a good next step are your main training material.
Keep some good disqualifications. A sales bot that never says "we are not the right fit" wastes your team's time. Include exchanges where a person politely and correctly turned away an unsuitable lead.
Favour your best people. If one person's conversations convert far better than everyone else's, weight the dataset towards them. That is the voice you want to scale.
Exclude what you do not want copied. Remove one-off discounts, promises your current terms no longer support, and anything written in anger.

Step 2: Clean and de-identify the data

Raw sales conversations are full of names, email addresses, phone numbers, company details and sometimes payment information. Remove this before the data goes anywhere near a model, and especially before it goes to a third-party API.

Open-source tools such as Presidio work in two stages. An analyser finds personal data using named-entity recognition, regular expressions and rules. An anonymiser then masks, replaces or redacts what it finds. You can add custom recognisers for things like internal account numbers. Replace details with consistent placeholders such as [CUSTOMER_NAME] or [COMPANY] rather than deleting them. That way the conversation still reads naturally and the model learns where personalisation belongs.

Presidio's own documentation warns that automated detection gives no guarantee of finding all sensitive information. A customer's name in an email signature or a postcode in casual text can slip through. Have a person review a sample of the cleaned data before you use it.

Step 3: Turn conversations into training material

For prompting or fine-tuning, each conversation becomes a structured example: the prospect's messages and the ideal reply from your side. OpenAI's fine-tuning documentation requires a JSONL file with one complete conversation on each line. The minimum is 10 examples. OpenAI suggests starting with about 50 well-crafted examples and says clear improvements usually appear somewhere between 50 and 100. It recommends realistic examples drawn from historical or logged data, which is exactly what past sales conversations are.

While you build the examples, write down the rules your best salespeople follow without thinking about them:

The qualifying questions, and the order they are asked in (budget, timing, decision-maker, need).
Which answers mean a lead should go to a person straight away.
What the bot must never do: give legal or financial advice, agree custom pricing or promise delivery dates.
How it closes each conversation: book a meeting, send a link or hand over.

Put these rules in the system instructions rather than leaving the model to infer them from examples. Written rules are easier to check and change.

Step 4: Set up evaluation before you train

OpenAI advises setting up evaluations before investing in fine-tuning. It recommends keeping some data back as a control group, with roughly the same range of inputs and responses as the training set. For a sales bot, hold back a share of your selected conversations. Run the bot on the opening messages from those conversations and compare its replies with what your team actually said.

Score replies on criteria a sales manager would recognise:

Does it sound like us? Ask someone who knows the business to review replies without telling them which ones came from the bot.
Did it ask the right qualifying questions in a sensible order?
Were any facts wrong? Count even one wrong price or invented feature as a failure.
Did it hand over or disqualify at the right point?
Did it stay within the rules from Step 3?

Run the same test set again after every change to prompts, retrieval content or training data. Otherwise you cannot tell whether a change made the bot better or just different.

Step 5: Connect it to your sales process

A bot that qualifies leads well but does not pass them anywhere has only moved the bottleneck. Plan the integrations early:

Write qualified leads and conversation summaries into your CRM.
Book meetings directly into salespeople's calendars.
Hand over to a person mid-conversation, with the full transcript, when a prospect asks for one or a rule is triggered.
Log every live conversation, with consent and a retention policy, so the strongest ones can feed the next round of improvement.

The legal questions to settle first

Using customer conversations to build an AI system is likely to be a new purpose for that data, which your existing privacy notices may not cover. Settle these points before you start.

UK and EU: lawful basis and transparency

The UK Information Commissioner's Office's work on generative AI focused on web-scraped training data. For that data, it considered legitimate interests the likely lawful basis. If you rely on legitimate interests for your own customer conversations, the same three-part test applies. You need a clear and specific purpose, the processing must be necessary for it, and individuals' interests must not override yours. The ICO also stressed transparency, because people cannot exercise their rights if they do not know how their data is used. Thorough de-identification reduces the impact on individuals, which should help with the balancing part of the test. Record your assessment, and update your privacy notice before you start, not afterwards.

US: do not quietly change your privacy policy

In February 2024, the US Federal Trade Commission warned that adopting more permissive data practices, including using consumer data for AI training, may be unfair or deceptive if the change is announced only through a quiet, retroactive change to the privacy policy. If your existing policy did not cover this use, tell customers before you rely on it.

Disclosure that prospects are talking to AI

Article 50 of the EU AI Act has applied since 2 August 2026. Chatbots that interact with people must tell users they are dealing with an AI system unless that is already obvious from the context. Guidance on the Act advises documenting your reasoning before relying on that exception. The FTC has also warned businesses not to use chatbots to mislead people. A bot that sounds like your owner is fine. A bot that claims to be your owner is not. Put a clear disclosure at the start of every conversation.

What drives the effort and cost

We will not quote figures here, because the effort depends on a few specific factors:

State of your data. Conversations held in one chat tool, with outcomes recorded in a CRM, need far less preparation than threads spread across personal inboxes and phone recordings.
Method. Prompting and retrieval avoid the training runs, and the specialist skills, that AWS notes fine-tuning can require. Fine-tuning becomes worthwhile when prompts with many examples get long and slow, or when the bot keeps drifting from your style.
Integrations. CRM, calendar and handover connections often take more work than the conversational model itself.
Compliance work. De-identification, assessments and policy updates take real time, especially across the UK, US and EU at once.

What we learned building one

In a sales chatbot we built, the business owner's problem was that their best sales conversations did not scale, so every new prospect had to wait for a person. A generic chatbot did not sound like the business it spoke for. Prospects can tell when they are talking to a script. That shaped the build: the chatbot learns from real conversations the business has already had, not from a template, and it handles first conversations and qualification around the clock. For anyone planning a similar project, the conversations you select are the raw material for everything the bot says, so Step 1 deserves the most care.

A short checklist

Export conversations from every sales channel and link each one to its outcome.
Select your best conversations and some good disqualifications. Drop one-off exceptions.
De-identify the data, then have a person check a sample.
Keep facts in a retrieval store. Use conversations to teach style and process.
Set aside test conversations and score the bot against them.
Connect it to your CRM, calendar and human handover.
Complete your lawful basis assessment, privacy notice update and AI disclosure before launch.

If you want a sales chatbot that sounds like your business and that you fully own, see our LLM chatbot work. Then send us a brief describing where your past conversations are stored and what a qualified lead means for you.