AI training data from support tickets
Disclaimer: The author is not a lawyer, accountant, or financial advisor. This article is for general information only and does not constitute legal, tax, or financial advice. Laws and market terms change, so consult a qualified professional before making any decision.
If you want to sell AI training data and you run a support team, start with your ticket history. Every solved ticket pairs a real customer question with the answer that fixed it, and labs building support models need that pairing at volume. This guide covers which archives are worth licensing, how to scrub names, account numbers and card data, when an outsourcing firm (BPO) needs client permission, and a prep checklist for your support lead.
Why support tickets work as AI training data
A ticket thread records a full piece of work: what the customer asked, what the agent checked, which policy applied, whether it was escalated, and what closed it. That ask-plus-resolution pair, repeated across years of real customers, is hard for a lab to get any other way. In an October 2026 note, law firm Frankfurt Kurnit Klein & Selz wrote that AI developers are hunting for internal operational data, tickets included, and that operating companies now get the same calls as startups that are shutting down.
Public alternatives look different. The Bitext customer support dataset on Hugging Face describes itself as hybrid synthetic: 26,872 question and answer pairs across 27 intents such as cancelling an order or tracking a refund. Your archive has what generated pairs lack: typos, the customer who replies three days later, the ticket that reopens after a bad fix. Those turns are why labs want real customer support data for AI training.
Contact centre, SaaS or BPO: where your prep work lands
| If you run | You likely hold | Biggest prep job | Who signs off |
|---|---|---|---|
| In-house support at a SaaS company | Email, chat and in-app tickets about your own product | Personal details in free text and screenshots | You, after checking customer terms |
| A contact centre for your own brand | Tickets, chat logs and call transcripts | Card data and account numbers | You, plus whoever owns PCI compliance |
| A BPO serving clients | Client tickets handled under contract | Written permission from each client | Each client, in writing |
What makes a ticket archive worth licensing
Beyond the general value drivers for company data, a strong ticket archive usually has:
- Tens of thousands of solved tickets over several years, so the same problem shows up handled by different agents.
- Mostly English threads from customers in Canada, the US or another Western market.
- Complete threads: the first message, every public reply in order, and the final status.
- Resolution signals such as solved status, satisfaction ratings or a reopen flag, so a buyer can tell a real fix from an auto-closed ticket.
- Specialized subject matter, such as insurance claims, billing disputes or technical troubleshooting.
Queues full of password resets and macro-only replies add volume but teach a model little. Before you sell AI training data from your helpdesk, weight the inventory toward queues where agents had to think.
Our partner publicly states that companies have been paid $100K to $2M+ per approved data package. That is not a guarantee, and many archives will not qualify. The free data value calculator gives a rough sense of where yours sits. If you have years of English tickets with real resolutions, run the two-minute eligibility check before spending anything on cleanup.
Catch the personal information hiding in ticket threads
Start with the structured fields. Requester name, email, phone, account ID, order numbers, mailing address and IP address sit in known columns, so you can drop or replace them in bulk.
Free text is harder. Customers paste account numbers and home addresses into messages, name family members, and describe health or money problems in their own words. Agents add signatures and direct lines, and internal notes can hold candid comments about the customer. Replace each identifier with a typed placeholder such as [CUSTOMER_NAME] or [ACCOUNT_ID] so the conversation still reads naturally. The Bitext dataset uses the same approach, writing order numbers as {{Order Number}}.
Leave attachments, screenshots and call audio out unless your partner asks for them and you can clean them separately.
Automated detection handles the first pass. The Presidio toolkit finds names, phone numbers, credit card numbers, locations and other entities in text, and its documentation warns that automated detection cannot guarantee it finds all sensitive information. Pair it with people: sample scrubbed tickets from every queue, log each miss, fix the rule and rerun.
Then test what is left. A rare product, a small town and an exact date can identify someone together, even after the names are gone. Our PIPEDA anonymization guide explains the serious possibility test this has to pass, and Quebec data faces the stricter Law 25 anonymization rules.
Stop card data from reaching a training set
If a customer ever typed a card number into chat or email, it is still in your archive. PCI DSS Requirement 3.3.1 bars storing sensitive authentication data, including card verification codes, after authorization even when encrypted, and the PCI Security Standards Council applies that rule to call recordings too. If verification codes turn up in your chat logs or transcripts, raise it with whoever owns PCI compliance before any licensing conversation.
For the package itself, run a card-number detector over every message. Presidio's credit card recognizer uses a pattern match plus a checksum. Replace hits with [CARD_NUMBER], exclude any ticket where a verification code appears, and scan the surrounding text for expiry dates and last-four digits.
BPOs: avoid licensing client tickets without written permission
A BPO answering tickets for a bank or a retailer holds that data on the client's behalf. Under clause 4.1.3 of PIPEDA's Schedule 1, an organization stays responsible for personal information it transfers to a third party for processing. It must use contractual or other means to give that information comparable protection while the third party processes it. Your client remains accountable for those tickets, so read your master services agreement for how it defines client data.
Permission comes first. A BPO that wants to sell AI training data built from client tickets should ask each client in writing. Explain what would be licensed and who anonymizes it, and expect the conversation to cover whether the client shares in the fee. Expect clients to check from their side: the same Frankfurt Kurnit Klein & Selz note advises companies to review whether vendor contracts let a vendor sell or license their data.
Your own QA scorecards, agent training programs and workforce processes are more likely to be yours, though SOPs written for a client's program can belong to the client. Our licensing rights checklist lists the other questions to clear.
A prep checklist for selling support data as AI training data
Before you apply:
- List each helpdesk, chat tool and call platform, with ticket counts by year and language.
- Pick the queues with real resolutions, and set aside spam, auto-replies and merged duplicates.
- Check your customer terms and privacy policy and, for client work, each client contract.
- Test the export. Zendesk's CSV export leaves out ticket comments and descriptions, which are the conversation itself. Its full XML export includes comments but caps at 500 MB (roughly 200,000 tickets), and the account owner has to ask Zendesk support to switch exports on.
- Run the eligibility checker and apply through our partner link.
After the partner confirms fit and counsel has reviewed the terms:
- Decide whether internal notes go in. They explain why an agent chose a fix, but they hold the most candid comments about customers, so scrub them as hard as customer text.
- Run automated detection, replace hits with typed placeholders, and sweep for card data.
- Sample scrubbed tickets by hand, log misses, fix and rerun.
- Generalize exact dates and small locations, then reassess re-identification risk.
- Document the method, file the record with your privacy officer, and deliver under the signed licence.
FAQ
Can a Canadian company sell AI training data from support tickets?
Yes, if it has the rights to the data and removes personal information first. Companies with 30 or more employees, mostly English tickets and customers in Canada, the US or another Western market are the clearest fit. Most deals are licences: you keep the original and grant defined rights to an anonymized copy.
Do customers need to consent before their tickets are used?
PIPEDA governs personal information, so properly anonymized tickets are generally outside its scope. The work is in proving the anonymization holds, and Quebec data has its own standard. Ask counsel to confirm for your situation.
What about call recordings?
Transcripts can be licensed like chat logs once scrubbed. Audio is harder to anonymize because a voice can identify a person, and recordings may hold card verification codes, which PCI DSS bars storing after authorization.
How much are support tickets worth?
There is no public price list. Our partner publicly states that companies have been paid $100K to $2M+ per approved data package, which is not a guarantee. Volume, years of history, complete threads and clean rights move the number.
Should we anonymize before applying?
Size the archive and check rights first. Full anonymization comes after the partner confirms fit and your team and counsel have reviewed the terms, so you avoid scrubbing data that will never be used.
Sources
- Frankfurt Kurnit Klein & Selz: Selling Company Data for AI Training (October 2026)
- Bitext customer support training dataset, Hugging Face
- Presidio documentation: PII identification and anonymization
- Presidio: PII entities supported
- PCI SSC FAQ 1210: sensitive authentication data in recordings
- PIPEDA, Schedule 1 (Justice Laws Website)
- Zendesk: exporting ticket data
This guide is general information, not legal advice. Have Canadian counsel review your rights, client contracts and anonymization method before you sell AI training data built from tickets.
Could your data qualify? Answer six questions in about two minutes. Nothing is stored.
Check if your company qualifies
Ready now? Apply to license your data (partner link).