Key takeaways
- AI companies test models on support work: OpenAI's GDPval includes customer service representatives, and τ-bench grades AI agents on simulated support conversations.
- When we review support history, depth matters more than volume: years of history and tickets with clear outcomes.
- Data a BPO holds on behalf of clients is out of scope, because those records belong to the client.
Which tickets, chats and calls are AI companies after?
| Record | What it shows | What comes out first |
|---|---|---|
| Resolved tickets | The problem, the steps taken and the fix, with status, tags and timestamps | Customer names, email addresses, phone numbers, account and order numbers |
| Live chat and messaging threads | Multi-turn conversations: the question, the follow-ups and the answer that worked | Names, addresses and anything customers typed about themselves |
| Escalations and internal notes | When a case gets handed on, why, and how specialists work out harder problems | References to named customers, and passwords or access details pasted into notes |
| Refund and billing cases | How policy gets applied when money is involved | Card numbers, bank details and all other payment data |
| Call recordings and transcripts | How agents explain, calm a caller down and confirm a fix | Card details and security answers read out on calls, and anything that identifies the caller |
| Macros, policies and QA scores | The rules agents work to, how they changed, and which answers reviewers judged good or poor | Content licensed from someone else, internal security procedures, and scores tied to named staff |
Customer service is a job AI models are already tested on
Customer service is one of the jobs AI labs measure models against. OpenAI's GDPval (September 2025) tests models on tasks from 44 occupations, including customer service representatives, and names "a customer support conversation" as an example work product. τ-bench, released by Sierra in June 2024, simulates conversations between a customer and an AI agent with tools and written policies; in the original paper GPT-4o solved fewer than half the tasks. τ²-bench (June 2025) adds telecom support, and Anthropic reported TAU-bench results for Claude 3.7 Sonnet in February 2025.
These benchmarks use simulated customers. A support team's history holds real ones, policies that changed, and edge cases nobody wrote down.
Which helpdesk tools hold the history?
Support history usually lives in help-desk and messaging tools, with call recordings in a separate contact-centre platform.
- Zendesk
- Intercom
- Freshdesk
- Salesforce Service Cloud
For a BPO, client tickets stay with the client
For a BPO, ownership comes first: data you hold on behalf of clients is out of scope. Under the UK and EU GDPR, an organisation handling personal data on a client's behalf is usually a processor and may only use it on the client's documented instructions, so the decision belongs to the client.
- Tickets, chats and calls handled for clients, unless the client itself agrees
- Payment card data, bank details and full account numbers
- Health, financial or other sensitive details customers shared in a ticket or call
- Anything covered by a client's confidentiality terms or an NDA
- Recordings where you are unsure what callers were told about how calls would be used
- Passwords, access details and content that belongs to a software vendor
Why depth beats volume in a ticket history
When we review support history, depth matters more than volume: years of history, tickets that kept their full thread and internal notes, clear outcomes such as status, tags, satisfaction ratings or reopens, and specialist work such as technical, banking or telecom support. A BPO's own material, such as training courses, QA frameworks and process documents, may also count, subject to its client contracts. Whether data you hold for clients can be included is answered in the FAQ on how to sell company data to AI companies, and the free check tells you whether your support history could qualify.
Frequently asked questions
Can you sell customer support data to AI companies?
Usually it is licensed rather than sold outright. A company that owns its support history can license tickets, chats and call transcripts to AI companies once customer details are taken out. Buyers want records that show a problem being solved, with the outcome attached. A BPO can only do this with records it owns: data it holds on behalf of clients is out of scope.
Can a BPO license the tickets it handles for clients?
Not on its own. Under the UK and EU GDPR a BPO handling a client's customer records is usually a processor and may only use them on the client's documented instructions. A processor that decides for itself how the data is used is treated as a controller for that use. Any licence has to come from the client. See whether licensing company data is legal.
How much do AI companies pay for support data?
There is no public price list for support data. Value depends on years of history, whether tickets carry outcomes, how specialist the work is and how much has to be removed. That is what our review covers: we look at what you have and tell you whether it could qualify and what affects its value. The check is free, with no commitment.
Sources8
- OpenAI, Measuring the performance of our models on real-world tasks (GDPval), 25 September 2025
- Yao, Shinn, Razavi and Narasimhan, τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains, arXiv, 17 June 2024
- Sierra, τ-Bench: Benchmarking AI agents for the real-world, June 2024
- Barres et al., τ²-Bench: Evaluating Conversational Agents in a Dual-Control Environment, arXiv, 9 June 2025
- Sierra Research, tau2-bench repository (airline, retail, telecom and banking domains), GitHub, checked 20 October 2026
- Anthropic, Claude 3.7 Sonnet and Claude Code (TAU-bench results), 24 February 2025
- GDPR Article 28, Processor (paragraphs 3(a) and 10)
- ICO, What are controllers and processors? (UK GDPR guidance)



