Conversational data for AI training
Multi-turn text conversations with speaker roles and timestamps: support chats, sales chats, community threads, in-app messaging and agent transcripts.
- Measured in
- Conversations, turns or tokens. Buyers usually ask for conversation count plus average turns per conversation.
- Who buys
- Foundation model labs, assistant and agent builders, and vertical AI companies that need domain dialogue.
- Where it comes from
- Customer support platforms, sales and success teams, online communities, marketplaces and closed startups with chat history.
- Rights to clear
- Platform terms (some chat tools restrict use of exported data for model training), customer contracts and the privacy notice in force when the messages were collected.
- Privacy
- Names, emails, phone numbers, account IDs and free-text personal details must be removed or replaced before sampling.
- Pricing signal
- Priced deal by deal on volume, domain and exclusivity. There is no public rate card for business chat data.
- Conversational data is multi-turn dialogue between people, or between people and software, recorded with enough context to show how a conversation unfolds.
- AI teams license it to teach models turn-taking, clarification, tone and task completion, which single-turn text does not show.
- The main risk is personal information inside the messages, so de-identification and a clear consent basis come before any sample is shared.
- We source conversational data from companies that hold it in their own systems, and the buyer licenses it directly from them.
What buyers look for in conversational data
A useful conversational dataset preserves the structure of the exchange. Buyers want speaker roles (customer, agent, bot), turn order, timestamps and the outcome where it exists, such as whether a ticket was solved or a sale closed. Without that structure, chat logs are just text, and much of their training value disappears.
Domain matters more than raw size. A few hundred thousand conversations from a specialist field, such as insurance claims or B2B software support, can be worth more to a vertical AI team than millions of generic social messages, because the specialist data teaches vocabulary and procedures the model has not seen.
Typical sources
- Customer support history from tools such as Zendesk, Intercom, Freshdesk and Salesforce Service Cloud
- Live chat and messaging on ecommerce and SaaS sites
- Community forums and member networks the company operates
- Sales conversations and onboarding chats
- Workplace messaging, covered separately on our workplace chat page
Rights you need to clear first
Three questions decide whether chat data can be licensed:
- Who owns the words? Customers wrote half of every conversation. Your terms of service and privacy notice at the time of collection decide what you can do with them.
- What do the platform terms say? Some messaging platforms restrict using exported data to train large language models. Check the API and export terms of the tool the data came from.
- Is there a contractual limit? B2B customer agreements often contain confidentiality clauses that cover support conversations.
The Federal Trade Commission has warned that quietly changing terms to allow AI training could be unfair or deceptive, so rights are judged against the terms in place when the data was collected (FTC).
De-identification
Conversational data almost always contains personal information in free text. A typical workflow removes direct identifiers with automated detection, replaces them with consistent placeholders so the dialogue still reads naturally, then has people review a sample for misses. Removal reduces risk but does not make data anonymous by itself, which is why the license should also forbid re-identification. See our guide to de-identifying data for AI training.
How a deal usually works
The holder prepares a de-identified sample of a few hundred conversations. The buyer evaluates it under NDA, then both sides agree volume, exclusivity, permitted use and price. Delivery is a prepared export, never live access to the support system. We introduce the parties and help structure the terms; the license is signed between buyer and seller.
Questions
Can we sell customer support chats for AI training?
Often yes, if your privacy notice and customer contracts allowed that use when the chats were collected, the platform's terms permit it, and personal information is removed first. A rights review comes before any sample leaves your systems.
How many conversations does a buyer need?
It depends on the use. Fine-tuning a support model can start with tens of thousands of high-quality conversations. Labs building general assistants want far more, but they also pay attention to domain coverage and outcome labels.
Do buyers want the agent's replies or only the customer messages?
Both. The value is in the exchange: how a skilled agent asks clarifying questions and resolves a problem. Removing one side destroys most of the training value.
Is chatbot output useful training data?
Less so. Most labs avoid training on another model's output because it adds little new information and can carry that model's errors. Human-written turns carry the value.
Related guides
Related data types
Sources
Need conversational data, or hold it?
Tell us what you need or what you hold. We check rights and privacy first, then introduce vetted counterparties. You license directly with each other.
- We reply to every enquiry within two working days.
- No upfront fees for sellers
- We never hold your raw data