Can you license transcripts from your AI support agent?

What is actually in an AI support agent transcript?#

An AI support agent transcript is more than a chat log. Behind the visible messages sit the help center passages the bot retrieved, the actions it took such as looking up an order or resetting a password, its routing decision, the point where a human took over, the human's reply, the resolution and whether the customer came back.

Where those records sit depends on how the agent was built. An AI add-on inside a support platform such as Zendesk, Intercom or Salesforce stores conversations in that platform, under its terms. An agent your team built on a model provider's API keeps logs wherever your engineers put them, under the provider's terms and your own.

Platform terms and export limits shape the answer before any rights review starts. Intercom's Additional Product Terms, for example, treat content submitted to its AI products and the output they generate as Customer Data. Zendesk's export documentation says AI agent tickets cannot be exported through its account export tools, so check early whether the bot side of the record can be retrieved at all.

Either way, at least two contracts sit behind every bot reply, and both need reading.

Who holds rights in each turn of the conversation?#

Rights in a bot conversation differ by turn type, so each turn type is its own question.

In B2B software, customer messages are often defined as customer data in your master agreement and data processing addendum, which can limit your role to providing the service. In consumer-facing products, privacy notices and state privacy laws may apply. Either way, the analysis is done deal by deal with counsel.

Why do bot replies need a separate check?#

Bot replies need a separate check because they come from someone else's model under someone else's terms. Many AI vendor terms assign outputs to the customer but restrict using them to develop competing models, and a training license to an AI developer can fall squarely inside that restriction.

Support platforms add a second layer. Some platform terms let the platform use customer conversations to improve its own AI features, often with an admin opt-out. If that applies, the platform may already be learning from the same transcripts, which affects any exclusivity you could offer and what you can honestly say about the data.

The decision rule: if bot outputs are restricted, scope customer and human turns plus resolution labels, and either drop bot turns or replace each with a short neutral marker, such as a note that the bot suggested an article and the customer asked for a person.

Why are handoffs the strongest records?#

Handoffs are the strongest records because they show exactly where automation fell short and how a person fixed it. An escalated conversation pairs a customer's real problem with a failed automated attempt and an expert resolution, which is useful for evaluating support AI as well as training it.

Keep the metadata that makes handoffs readable: the reason for escalation, the queue it went to, the time to resolution, whether a refund or credit was issued, and whether the ticket was reopened. Without those fields a transcript is just text; with them it is a labeled outcome.

Handoffs also show your support process at its most specific. The human reply usually cites internal policy, product limits or workarounds that never made it into the help center, which is why these records are worth the extra preparation.

How should bot transcripts be prepared?#

Bot transcripts are prepared like any chat log, with extra attention to the structured actions that expose account data.

Automated tools help but are not enough on their own. The open-source Presidio project, for example, warns in its own documentation that because it uses automated detection, there is no guarantee it will find all sensitive information, and that additional protections should be used.

• 1Export conversations with channel, timestamps, handoff point, resolution and reopen flag.
• 2Remove names, email addresses, phone numbers, street addresses, order and account numbers, and payment details.
• 3Remove or generalize customer company names in B2B conversations.
• 4Check free text and attachments for pasted screenshots, logs and documents.
• 5Strip identifiers from bot action payloads, such as lookup results and API responses.
• 6Run automated detection, then human review of a sample from every batch.
• 7Record what was removed, how and by whom in the privacy record.

Which conversations should be left out entirely?#

Some conversations should be left out of any scope no matter how well they are de-identified, because the content itself is the risk. Filtering them out early saves review time and keeps the remaining set easier to defend.

Build the exclusion filter from fields you already have, such as ticket category, queue, tags and customer account, and check a sample by hand. Categories drift over time, so a security conversation can hide under a billing tag.

• 1Account security: identity verification, password resets with security answers and suspected account takeovers.
• 2Payment disputes, chargebacks and anything quoting card or bank details.
• 3Legal complaints, threats, harassment reports and requests from regulators or law enforcement.
• 4Conversations where customers volunteered health, family or other sensitive personal details.
• 5Conversations from customers whose contracts prohibit reuse, or from regions you have chosen to exclude.
• 6Any conversation under a legal hold.

Illustrative: an accounting software company reviews its bot transcripts#

Illustrative: a fictional accounting software company for small firms runs an AI agent inside its support platform. Customers ask about bank feeds, invoice templates and payroll errors, and conversations the bot cannot resolve escalate to human specialists.

The general counsel finds two issues. The platform's terms allow it to use conversations to improve its AI features unless the account opts out, and the underlying model terms restrict using outputs for competing models. Customer messages are also covered by the company's data processing addendum.

The company opts out of platform model improvement, scopes only escalated conversations, keeps de-identified customer turns, human replies and resolution data, and replaces bot turns with neutral markers. Customer contracts are reviewed before anything moves forward, and the remaining scope is small but well documented.

How SourceX reviews AI agent transcripts#

SourceX reviews AI agent transcripts turn type by turn type within the SourceX five-step transaction. In the Rights step, support platform terms and model provider terms are read alongside customer contracts and privacy notices; in Preparation, personal and confidential details are removed and the work is recorded.

The SourceX Evidence Packet states which turn types are included, under what permitted use and with what privacy record, and the company approves each step before anything is released. As with any rights question, this is general information rather than legal advice, and each deal is assessed with counsel.