How to sell customer service transcripts for AI training.

Who buys support tickets, chat logs and call transcripts, how your team gets them out of Zendesk or Intercom, what your customers were promised, and how a license pays. Checked 16 November 2026. If your company has 30 or more office employees, Allied arranges a direct license with an AI lab and gets you a written offer.

I

The short answer

Yes, a company can license its customer service transcripts to an AI lab, if the transcripts are its own and the customers in them cannot be identified. What changes hands is a copy with a written license. Your help desk keeps running and the originals stay where they are.

The help desk vendor does not own the conversations. Zendesk’s Customer Agreement says: “As between the parties, Customer retains ownership of all Service Data.” Service Data is defined there as all data, text, messages, communications or other information submitted to and stored within the services.1

There are buyers. micro1, which supplies training data to AI labs, lists customer support as one of six categories it licenses from companies: support documentation, ticket workflows, QA processes, customer operations and knowledge management. Its page says payments typically range from $100K to over $3M, and that it looks for established companies with 30+ full-time employees.6

The hard part is the customers. A transcript is full of names, order numbers, addresses and sometimes card details, typed by people who were told their data would be used to help them. Most of this page is about that: what your contracts and privacy policy allow, and what has to come out before a copy leaves.

If you only need the checklist for tickets, see selling support ticket data. For recorded audio, see selling call recordings. Every system Allied covers is listed on the sell page.

RecordWhere it usually livesWhat a lab learns from itWhat comes out first
Ticket threadsZendesk, Freshdesk, Help Scout, Salesforce Service CloudA problem stated in the customer’s words, the replies, and how it was closedNames, email addresses, order and account numbers, signatures
Live chat and messagingIntercom, Zendesk messaging, GorgiasShort turns, clarifying questions, the hand-off from a bot to a personThe same, plus anything a customer pasted into the chat box
Call transcriptsGong, Aircall, Dialpad, contact center platformsSpoken troubleshooting, interruptions, how an agent calms a callerSpoken names and numbers; calls recorded without the required notice
Internal notes and escalationsThe same help desk, on the private side of the ticketWhat the agent was thinking and when a case moved up a tierColleague names and other customers mentioned in passing
Macros and help center articlesHelp desk admin, knowledge baseThe approved answer, to compare with what agents actually wroteUsually nothing; much of it is already public
Satisfaction scores, tags, resolution timesTicket fieldsWhich conversations went well, so a grader can score an AI agent’s attemptUsually nothing
II

Why labs want support conversations

Support is one of the first jobs companies hand to AI agents. To train one, a lab needs many real examples of the job done by people: a question, the back-and-forth, the fix, and whether the customer left satisfied. A resolved ticket holds all four.

After Google bought Spirit Airlines’ records in bankruptcy, CRM magazine asked contact center executives what the data was good for. Adi Tahiri, chief technology officer of the outsourcer Konecta, called customer service data “undoubtedly a treasure trove for training LLMs,” and said airline data is especially useful because it varies across languages, demographics and regions.9

Internal notes are the part outsiders never see. A public reply shows what the customer was told. The note behind it shows why: the log the agent checked, the policy they weighed, the reason they escalated. Labs building agents that have to reason, not just reply, ask about notes early.

Outcome fields matter for a practical reason. A training environment needs to score an attempt. A ticket marked solved with a good satisfaction rating is a scored example. A thread with no status and no rating is harder to use.

III

Who buys customer service transcripts

Three kinds of buyer show up on the public record. The first is the lab supplier. micro1 names support tickets among the records it wants and says it has 30 days to review and approve data after transfer, then pays within 30 days of invoicing.6

Not every supplier asks for help desk data by name. Mercor’s company data page lists nine tools it connects, among them Slack, Gmail, Salesforce and GitHub, and none of the nine is a help desk. The page says more than 100 tools are supported, so ask before assuming.7

The second is an AI lab buying directly. In August 2026 Google outbid Mercor for Spirit Airlines’ records in bankruptcy, including 100 million emails and 500 million Teams messages. The price was $10 million.8

CRM magazine reported that the Spirit data also held 7.5 billion passenger transaction records, and quoted David Fischer of Luware describing the purchase as a large dataset of deidentified airline customer-service interactions.9

The third is the dataset marketplace, which sells call center data off the shelf. One Datarade listing, from the complaints site PissedConsumer, offers 25,000 hours of U.S. support calls with transcripts at $5,000 per purchase. Its listed fields are company name, date, transcription, summary, duration and location. There is no internal note, no resolution status and no link to the systems the agent used.10

That gap is why one company’s own archive is licensed on different terms from pooled calls. For the wider list of buyers, see who buys AI training data. Allied’s role is narrower: it arranges a direct license between your company and a lab. How Allied works.

IV

Getting transcripts out of Zendesk and Intercom

Your own admin runs the export. Plan for it early, because the default export in both tools leaves the conversation out.

In Zendesk, data exports are not enabled by default. The help center says the account owner must contact Zendesk Customer Support to enable them, and the export tools are not available on Team plans. Once they are on, an admin can export tickets to JSON, CSV or XML.2

Pick JSON. Zendesk says the CSV export does not include deleted tickets, ticket comments or ticket descriptions, so it carries no transcript at all. Full XML is not available for accounts with more than 200,000 tickets, and XML ticket data leaves out custom ticket fields.2

JSON has two limits of its own. If a single ticket has more than 1 MB of data, its comments are not included in the file. Accounts with more than one million tickets are downloaded in 31-day increments, which turns a large archive into a long series of files.2

Intercom works differently. Its help center says conversation and message content is not included in the CSV export, which holds reporting metadata. For full conversations it points to the Conversations API: list the conversation IDs, then make a GET request for each one. The Export API allows a maximum 90-day timeframe per request.4

The same Intercom article says the inbox can export selected conversations as PDF or text, but internal notes are excluded from those transcript exports. It also says data is available for up to two years for both CSV and JSON exports. Check how far back yours reaches before you describe the archive to a buyer.4

Intercom’s API reference adds a trap. Without paging, a retrieved conversation returns only the 500 most recent conversation parts. Long threads need paging, or they arrive with the opening messages missing.5

Freshdesk, Help Scout, Gorgias and Salesforce Service Cloud each have their own export route. The test is the same in all of them: open the file and confirm that the full message text and the private notes are there, with their timestamps and author roles.

V

Three documents decide whether the transcripts are yours to license

The first is the help desk agreement. Zendesk’s gives the customer ownership of Service Data and has the customer instruct Zendesk to use that data to provide, secure and improve Zendesk’s own products and services. It does not restrict what the customer does with its own copy. It also sets a clock: the customer may export Service Data during the subscription term and for 30 days after it.1

The second is your privacy policy as it read when the tickets were written. That is covered in the next section.

The third is your contracts with business customers. If you sell to companies, their master agreements often call anything they send your support team confidential, and a data processing agreement may limit you to using their users’ data for providing the service. Those accounts are normally left out of scope. Your counsel reads these; nobody else can.

Outsourcers have a harder version of this. A contact center that answers tickets for client brands usually holds the transcripts under the client’s contract and on the client’s instructions. The client decides whether they are licensed, not the outsourcer.

Clause-by-clause detail on the license itself is in the licensing agreement guide. The general legal questions are in is it legal to sell company data.

VI

The customers in the transcripts

Start with what you told them. The Federal Trade Commission wrote in February 2024 that it may be unfair or deceptive for a company to adopt more permissive data practices, such as sharing consumers’ data with third parties or using it for AI training, and tell consumers only through a “surreptitious, retroactive amendment to its terms of service or privacy policy.” Tickets collected under an old policy are licensed under that policy’s promises.11

California’s privacy law draws the line that most licenses are built on. Making a consumer’s personal information available to a third party for money or other valuable consideration is a sale. Deidentified information is treated differently: it must not reasonably be usable to infer information about, or be linked to, a particular consumer, and the business has to commit publicly not to reidentify it and bind recipients by contract to the same rule.12

Call transcripts inherit the rules of the recording. California Penal Code section 632 makes it an offense to record a confidential communication without the consent of all parties, with a fine of up to $2,500 per violation. A transcript made from a call that was recorded without the notice the caller’s state requires carries the same defect. Leave those calls out.13

Card data is a separate problem. The PCI Security Standards Council’s guidance is that sensitive authentication data, meaning the card verification code and the PIN, must not be stored after authorization even if encrypted. Customers paste card details into chats anyway.14

Help desk redaction does not settle it. Zendesk’s automatic redaction keeps the first six and last four digits of a card number, does not support redacting numbers already in the system, and comes with a warning that its checks do not guarantee every card number will be identified. Scan the export itself.15

Health details need more care than this page gives. The HIPAA safe harbor method lists 18 categories of identifiers to remove, including names, telephone numbers, email addresses and account numbers. If your support desk serves patients or health plan members, keep those queues out unless your counsel says otherwise.16

Europe is stricter. Tahiri told CRM magazine the Spirit deal would not pass in Europe, because selling recordings to a third party for AI training changes how the data is used, which “would create significant legal and regulatory barriers” under the GDPR. Many sellers exclude EU and UK customers from the first license.9

Half of every transcript was written by your own staff. Agent names come out with customer names. Whether employees need notice or consent depends on where they work: see employee consent for AI training data.

VII

Removing identifiers from a transcript

Your own team does this, inside your own systems, before any copy leaves. Allied never sees, reads or receives the transcripts. It sees only a description of the archive that you approve.

Structured fields are the easy part. Drop the requester name, email, phone, organization and any custom field holding an account or order number. Keep the fields that carry no identity: status, priority, tags, channel, timestamps, satisfaction rating.

The message body is where identifiers hide. Customers sign their emails. Agents greet them by name. Order numbers, tracking numbers, street addresses and serial numbers sit in the middle of sentences, and quoted reply chains repeat them. Detection tools that find names and numbers in free text do most of the work; a person reading a sample catches what they miss.

Replace, do not just delete. A reply that reads “Hi [NAME], your order [ORDER_ID] shipped” still teaches the pattern. A reply with holes in it teaches less. Use the same placeholder for the same person through a thread so the conversation still makes sense.

Attachments are usually excluded. Screenshots, invoices and photos of ID documents carry identifiers that text tools cannot see. For calls, the transcript and the audio have to match: a name removed from the text and left in the recording is not removed.

The lab reviews a sample before the full delivery, and so should you. The method for each system is in how to anonymize company data for AI training.

VIII

What a support archive is worth

No lab has published a price for a company’s support tickets alone. What is public is how buyers say they pay. micro1 bases compensation on the quality, uniqueness and value of the data, shows two tiers on its page ($100K-500K and $500K-3M), and says companies with several years of records across multiple business systems are generally a stronger fit.6

Allied’s own figure: old operational business data could be worth $100K–⁠$2M. Where a support archive lands depends on the points below. More on pricing is in what company data is worth.

  1. Years of resolved history

    A long run of closed tickets shows a product changing and the same problems being solved in new ways. An archive that starts last spring shows much less.

  2. Notes and escalations kept

    Private notes and tier hand-offs are the reasoning. If your export drops them, as some default exports do, the archive loses the part buyers ask about first.

  3. Outcome fields filled in

    Solved status, reopen counts and satisfaction ratings let a grader tell a good answer from a poor one.

  4. Technical depth

    Troubleshooting a product over several turns is rarer than “where is my order”. Software, hardware, finance operations and logistics desks tend to have more of it.

  5. More than one language

    Tahiri singled out variety across languages and regions as what made an airline’s service data useful.9

  6. Links to other systems

    A ticket that points to the bug it caused, the order it concerned or the call that followed lets a buyer rebuild the whole task. See what data AI companies buy.

IX

How a license happens, step by step

  1. Count the archive

    Tickets and conversations by year and by channel. Note how many carry private notes and a satisfaction rating, and which languages appear.

  2. Read the three documents

    Your help desk agreement, your privacy policy as it read over the years in scope, and your contracts with business customers. List the accounts and queues that have to stay out.

  3. Switch on the export

    In Zendesk the account owner has to ask Zendesk Customer Support to enable data exports, so start that request before you need the file.2

  4. Request an offer

    Answer a few questions on the request form. If the records fit, a discovery call follows. After you sign with Allied, a written offer naming the lab follows in one to two weeks.

  5. Export and clean a sample

    Your admin pulls a scoped sample and your team removes identifiers. The lab reviews the sample, not the full archive.

  6. Sign, deliver, get paid

    You sign the license with the lab and send the full scoped copy straight to it. Payment arrives by wire 30 to 60 days after the lab accepts delivery.

The license is non-exclusive, so the same archive can be licensed again. You keep ownership throughout. The process for every record type is in how companies license data to AI labs.

X

If the company is closing

A support archive is often the largest written record a closing company has, and the easiest to lose. The loss usually happens when someone cancels the help desk to stop the bill.

Zendesk’s agreement lets a customer export Service Data during the subscription term and for 30 days after.1

Its deletion policy says the deletion process starts 90 days after an account is canceled or terminated. Ticket data is then deleted within 40 days and live chat content within 120 days. In the policy’s words: “Once commenced, this process cannot be reversed and Service Data will be permanently deleted.”3

So export before you cancel, and keep the export somewhere the company still controls. The sale of records is now a normal part of a wind-down: SimpleClosure told MIT Technology Review that 47% of its deals now include code or data sales.8

A closed company’s promises still count. Tahiri pointed out that Spirit, being defunct, no longer had service obligations to its customers, and that the deal was about recovering value for creditors. The privacy policy those customers agreed to did not end with the airline.9

Someone with authority has to sign: the board, or the assignee or trustee once one is appointed. See selling startup data when shutting down, what happens to company data when a business closes and the shutdown checklist.

XI

The risks, plainly

  1. A customer is identified

    The usual cause is free text: a signature, an address in a quoted reply, a rare product combined with a town. Sample review before delivery is the control.

  2. A promise is broken

    Licensing transcripts that your privacy policy never covered is what the FTC’s 2024 guidance warns about.11

  3. A business customer’s contract is breached

    Their confidentiality clause can cover what their staff told your support team, whatever privacy law says. Exclude those accounts.

  4. Card or health data slips through

    Old tickets predate whatever redaction you run today. Scan the export, not the help desk.15

  5. Your product’s weak points are on show

    A support archive is a list of what broke. The license should limit use to training and evaluating models, forbid reidentification and resale, and set a deletion date.

  6. Exclusivity

    An exclusive grant stops you licensing the same history a second time. Non-exclusive terms keep that open.

XII

Questions about selling support transcripts

Can we sell our customer support tickets to AI companies?

Yes, if the tickets are your company’s own, your privacy policy and customer contracts allow it, and names and other identifiers are removed first. The usual form is a license of a de-identified copy, not a sale of the help desk. You keep the originals.

Who buys customer service transcripts for AI training?

Lab suppliers such as micro1, which lists customer support records among the company data it licenses; AI labs buying directly, as Google did with Spirit Airlines’ records; and dataset marketplaces that resell pooled call center data. Allied arranges a direct license between your company and a lab.

How much are support transcripts worth?

There is no public price for support tickets alone. Old operational business data could be worth $100K–$2M. Years of history, private notes, outcome fields, technical depth and languages decide where an archive lands.

Do our customers have to consent?

It depends on what your privacy policy said when the tickets were written, where the customers live and whether they can be identified in the copy. Properly de-identified records are treated differently from personal information under California law. Business customers’ contracts can restrict use even then. Ask your counsel.

Do support agents need to consent before their messages are licensed?

Agent names are removed along with customer names. Whether staff need notice or consent depends on the state or country they work in. European rules are the strictest.

How is the data anonymized before it goes to a lab?

Your own team removes identity fields, replaces names, emails, phone numbers, addresses and order numbers in the message text with placeholders, scans for card numbers, and drops attachments. You approve a sample before the full copy is sent. Allied never sees the transcripts.

Does Zendesk or Intercom own our tickets?

Zendesk’s Customer Agreement says the customer retains ownership of all Service Data. Read your own signed agreement and order forms, and Intercom’s terms if you use it, because your version controls.

How long does it take, and how are we paid?

After you sign with Allied, a written offer naming the lab follows in one to two weeks. Export, sample review and signing depend on your team’s pace. Payment arrives by wire 30 to 60 days after the lab accepts delivery.

Is selling transcripts straight to OpenAI or Anthropic possible?

Labs do buy company records, and Google bought Spirit Airlines’ directly in a bankruptcy auction. Most purchases on the public record run through suppliers that prepare training data for the labs. A direct license names the lab in the written offer.

We are a call center that works for other brands. Can we license the transcripts?

Usually not on your own. The transcripts are normally held under each client’s contract, so the client has to agree, and many client contracts forbid any other use.

Our company is shutting down. Can it still license its support history?

Yes, while someone has authority to sign and the archive still exists. Export from the help desk before cancelling the subscription, because the vendor deletes the data after cancellation and the deletion cannot be reversed.

What are the alternatives to selling to an AI lab?

Keep the archive and use it to improve your own support tools, license through a supplier’s intake program, or list with a dataset marketplace. Each has different terms on who sees the data and whether the license is exclusive.

undefined

Sources

  1. Zendesk, Zendesk Customer Agreement (effective August 1, 2025 for new customers), sections 3.1, 3.4 and 12
  2. Zendesk Help, “Exporting ticket, user, or organization data from your account”
  3. Zendesk Help, “Zendesk Service Data Deletion Policy” (effective June 3, 2026 for new customers)
  4. Intercom Help, “Export your conversation data”
  5. Intercom Developers, REST API reference, “Retrieve a conversation”
  6. micro1, “Monetize your company’s operational data” (company data partnerships page), read 16 November 2026
  7. Mercor, “Mercor Data” (company data page), read 16 November 2026
  8. Matthew King, “AI labs are buying data from dead startups to train their models,” MIT Technology Review, 13 October 2026 (reporting)
  9. Leonard Klie, “Spirit Sells Off Customer Data to Google,” CRM magazine (destinationCRM), 21 August 2026 (reporting)
  10. Datarade, PissedConsumer customer service call dataset listing, read 16 November 2026
  11. Federal Trade Commission, “AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive,” 13 February 2024
  12. California Civil Code § 1798.140, definitions of “deidentified” and “sell” (CCPA), via California Public Law
  13. California Penal Code § 632, recording confidential communications, via California Public Law
  14. PCI Security Standards Council, FAQ 1533, storage of sensitive authentication data after authorization
  15. Zendesk Help, “Automatically redacting credit card numbers from tickets”
  16. 45 CFR § 164.514, de-identification of protected health information (Cornell LII)

This page is general information, not legal advice. Whether your company can license its support records depends on your contracts, your privacy policy and the law where your customers and employees are. Vendor terms and export limits change; read the current versions. Ask your own counsel.

Find out what your support history is worth.

A few questions. About two minutes. How Allied works.

Request an offer