How to sell customer support ticket data to AI companies

Why AI companies want resolved support tickets

A resolved ticket is a complete worked example. It starts with a problem in the customer’s own words, often vague or wrong about the cause. Then come the questions an agent asked, the account checks, the workaround, sometimes an escalation, and finally a fix. The status, tags and satisfaction score tell you whether it worked.

That shape is exactly what AI labs need to train agents that do customer service. Public tests show the gap. In τ-bench, a benchmark where AI agents handle airline and retail customer service tasks under written policies, even state-of-the-art agents such as GPT-4o succeeded on fewer than half the tasks in its 2024 paper, and were inconsistent across repeat tries (τ-bench paper). The benchmark came from Sierra, which builds customer service agents. Its head of research explained why it built one:

Buyers use real records to build training environments: simulated workplaces where an agent practices the job and gets graded (Gizmodo). An RL environment founder told Epoch AI that a single training task typically costs $200 to $2,000 to build (Epoch AI). A real ticket with a known good outcome is raw material for that kind of task. Data vendors now recruit businesses directly. micro1, for example, runs a data partnership program for companies’ everyday workflows and says those deals are typically worth $100K to $2M+ (micro1).

Tickets also sell as part of broader archives. The former CEO of Cielo24 sold its Slack messages, internal email and Jira tickets for “hundreds of thousands of dollars” after the company shut down (Gizmodo). Our guide to selling company data to AI covers the wider market and the ways to sell.

What makes a ticket archive worth more

Two helpdesks with the same ticket count can be worth very different amounts. Buyers pay for context and outcomes, not volume alone. These signals raise the value:

A specialized product helps too. Tickets about payroll rules, lab equipment or freight customs are harder to find than password resets. What lowers the value: spam, auto-closed tickets, threads with no reply, and tickets so full of health, payment or personal detail that most of the text has to come out.

There’s no public price list for ticket data. SimpleClosure reported about 100 deals in a year for whole startup archives at $10,000 to $100,000 each (SimpleClosure). For operating companies, licenses for company data run from about $20K to $5M, depending on history, specialization and what labs are buying right now.

How to export tickets from Zendesk, Intercom and Freshdesk

The usual trap: the quick CSV export in each tool leaves out the conversation itself, which is the part buyers want.

Zendesk

• The full JSON export includes ticket comments. It arrives as newline-delimited JSON in a zip file, with a download link sent by email (Zendesk).
• The CSV export leaves out ticket comments, descriptions and deleted tickets.
• Data exports are off by default. The account owner has to ask Zendesk support to turn them on. The export tools aren’t on Team plans, though every plan can use the REST API.
• Accounts with more than a million tickets export in 31-day chunks, and tickets over 1 MB can come through without their comments, so check the error file that comes with the export.

Intercom

• The CSV export of conversation data contains metadata, not message content (Intercom).
• For full conversations, Intercom can export to Amazon S3 or Google Cloud Storage as JSON or JSON Lines, optionally with attachments. A historical export covers two years of conversations, and you can add hourly or daily exports of new ones. Setting it up needs full administrator access (Intercom).
• Check how far back your exports reach before you promise a buyer a date range.

Freshdesk

• An account admin can export tickets with their content from Admin, then Account, then Account Details, then Export. You pick a date range, and a zip of XML files for tickets, users, groups, companies and solution articles is emailed to the admin (Freshdesk).
• Solution articles are your knowledge base, so this one export gives you tickets and the docs agents followed.

Whichever tool you use, run the export yourself with your own admin account. You shouldn’t have to hand a buyer a login or API key to your live helpdesk.

Customer privacy and contract checks

Your customers wrote these tickets, and many of them are businesses with their own confidentiality terms. Owning the helpdesk account doesn’t settle whether you can license what’s in it. Before you agree to anything, check:

1. Your privacy policy and terms. What did you tell customers about how support conversations are used and shared?
2. B2B customer contracts. Business customers often share confidential details in tickets, covered by NDAs or confidentiality clauses. Removing names doesn’t undo a contract that bans sharing (Frankfurt Kurnit).
3. Data processing agreements. If you handle tickets on behalf of clients, for example as an outsourced support team, the data may be theirs, not yours.
4. Sensitive categories. Health, payment, children’s and legal details call for extra care. Many sellers leave those queues out entirely.
5. Where your customers live. Privacy law depends on the customer’s location, not just yours.

Reputation matters as much as the letter of the law. Daniel Goldberg, a partner at Frankfurt Kurnit, notes that an operating company’s customers and partners may be upset to learn their communications were sold, and that reaction can turn into regulator complaints, contract disputes or lawsuits (Frankfurt Kurnit).

This isn’t legal advice. Have a lawyer who knows your markets review the scope and the license before any data moves.

How to redact support tickets

Support tickets are harder to clean than most company data, because customers type anything into a support form. Expect to find:

• Names, emails, phone numbers and postal addresses, in the body and in signatures.
• Order numbers, account IDs, invoice numbers and tracking codes.
• Card numbers, passwords, API keys and one-time codes pasted in by customers.
• Screenshots, invoices and photos as attachments.
• Requester, organization and custom fields in the ticket metadata.
• Agent names, which identify your staff.

Don’t assume your helpdesk already cleaned old tickets. Zendesk’s automatic credit card redaction, for example, applies to new comments and doesn’t touch tickets already in the account (eesel).

Replace identities with consistent labels, like “Customer 14” and “Agent 3”, so a thread still reads naturally and a buyer can follow who said what. Then review whole tickets, not just fields. The UK regulator’s guidance says pseudonymised data is still personal data (SCL), and it tests anonymity by asking whether a motivated intruder could work out who someone is (Boyes Turner). A one-off outage at a named hospital can identify a customer without a single name in the text. If a ticket can’t be cleaned, it stays out.

How a ticket dataset is packaged

Buyers judge a dataset from its documentation and a sample. A well-packaged one includes:

• One record per ticket, with every message in order and labeled by role: customer, agent or internal note.
• Timestamps, status history, tags, channel and priority, so a buyer can see how the case moved.
• Outcome fields: solved or not, satisfaction score, reopen count.
• Macros and help center articles as their own files, linked from the tickets that used them.
• A data sheet: date range, ticket counts, product areas, languages, and how redaction was done.
• A redacted sample buyers can review before they commit.

With Smelted, you pick the helpdesk, folders and date ranges, and you can leave out specific customers or sensitive work. We remove identities, package the tickets, and buyers see redacted previews only. You approve the buyer and the terms, and once the license is signed you’re paid in as little as 7 days, with nothing upfront. Get a 60-second estimate of what your ticket history could be worth.

Questions founders ask