Which ticket fields do AI data buyers want?

Ask a support lead what is valuable in their helpdesk and they will point at the macros, or the knowledge base. Ask an AI data buyer and the answer is narrower and stranger: the messy middle of a ticket, where a confused customer and a patient agent worked out what was actually wrong.

What do buyers’ own pages say they want?

Four buyers name support data on their pages, each in its own words.

• micro1 puts support records in with the rest of its operational wish list: bug tracking, SOPs, process docs, internal wikis.
• Scale AI names support records, then chat and messaging histories as a separate line.
• Labelbox Prism asks for longitudinal customer support and, for support logs, favors busy consumer-facing brands.
• Nyne lists support tickets next to service histories, CRM and dispatch records.

None of them publishes a field list. What follows is what those categories mean in practice, built from how support data is used to train and test AI support agents.

Which fields carry the value?

The thread, in order

The single most useful thing is the complete exchange: first message, every reply, every follow-up, in sequence. A model learns from the turn-taking: what the customer said, what the agent asked back, what changed the outcome. A thread with replies out of order is worth much less.

Who spoke, as a role

Each turn needs a speaker label: customer, agent, bot, or internal note. Not a name. The difference between a bot’s answer and a human hand-off is exactly what many buyers want to study.

Timestamps

Response gaps and resolution times show pace. Shift them by a fixed amount if exact dates could identify an incident; keep the intervals intact.

Shape fields

Channel, status, priority, ticket type, tags and the path of status changes give the conversation a frame. They let a buyer sort ten years of tickets into refunds, outages, how-to questions and complaints.

Reusable replies

Macros, canned responses or saved replies show what your team decided was the right answer to a common problem. Paired with the tickets where they were used, they are a small, well-labeled dataset on their own.

The outcome

Solved, escalated, refunded, churned, rated good or bad. A satisfaction rating attached to a thread is one of the clearest signals of what a good answer looked like.

Which fields add little, or harm?

Field	Value to a buyer	Risk to you	What to do
Thread text	High	Medium, until redacted	Keep, redact
Speaker role	High	Low	Keep
Timestamps	Medium	Low	Keep or shift
Tags, status, channel	Medium	Low	Keep
Macros and saved replies	Medium	Low	Keep
Satisfaction rating	Medium	Low	Keep
Internal notes	High	High	Decide with care
Customer name, email, phone	None	High	Replace with tokens
Order or account numbers	None	High	Replace with tokens
Card digits	None	Very high	Remove
Attachments	Low	High	Drop

Internal notes sit in the awkward corner. They show an agent’s reasoning, which buyers prize, and they are also where agents paste account details and write unguarded comments. Many teams leave them out of a first dataset.

Does volume or history matter more?

Both matter, differently. Volume gives coverage of rare problems. History gives change over time: a product launch, a pricing change, a new policy, and how support conversations shifted around it. That is one plain reading of what Labelbox Prism calls longitudinal customer support.

For entry, though, buyers look at the company first. The published rules start with headcount: 20 full-time US employees and three years in business at Nyne, 30+ employees at micro1, and 40+ as Scale AI’s preferred starting point. Our archive value check runs your numbers against those rules.

Do languages and channels change anything?

Language does at one buyer we track: micro1 says it wants data primarily in English. The others do not publish a language rule.

Channel mix changes what the data is good for. Email threads are long and structured. Chat sessions are short, fast and often start with a bot. Phone notes are an agent’s summary, not the customer’s words, so buyers treat them differently. Say which you have when describing the archive.

What should you prepare before talking to a buyer?

A one-page description, not a file:

1. Helpdesk platform, and any earlier one.
2. Ticket count per month and years of history.
3. Channels and their rough share.
4. Languages.
5. Whether internal notes, macros and satisfaction ratings exist.
6. Whether card or health details could appear, and how you would find them.

That page answers most of a buyer’s first call. It is also what our form asks for if you want to know who would license your archive. The homepage explains how to sell support tickets on published terms.

Treat the thread as the product, and everything with a name on it as waste to remove.