AI data market
Support data investment segmentation: five markets, not one dataset
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Customer support is several data markets, not one dataset. Split it into five decisions, resolution, escalation, refunds and returns, entitlement and policy exceptions, and rate each on outcome evidence, preparation burden, privacy exposure and dated demand signals. The strongest alignment sits in policy-bound transactional decisions such as refunds, returns and entitlement checks, where outcomes can be checked against a system of record and public agent benchmarks already test them.
Checklist
Support data investment
Key takeaways
- Ticket volume says little about value: the five support decisions need different context, outcome evidence and buyers.
- Refunds, returns and entitlement checks have the clearest outcome evidence, because the right answer leaves a verifiable state in a system of record.
- Public agent benchmarks such as tau-bench already grade agents on these policy-bound actions by checking the final database state.
- Complete case episodes take far more preparation than single tickets because they need event histories, linked transactions and policy versions.
- Synthetic users, contested rights over customer-written text, privacy exposure and concentrated buyers are the risks that could break the thesis.
How should investors segment customer support data?
By the decision each record documents, not by the system it sits in. A help desk export mixes at least five different decisions: resolving a question, escalating a case, issuing a refund or return, checking an entitlement and approving a policy exception. Each needs different context to be usable, leaves different evidence of whether it was right and serves a different buyer task, so each behaves like its own data market.
Treating support data as one dataset leads to the wrong diligence question. Counting tickets measures supply abundance; it says nothing about whether the records can teach or grade an agent on the decision a buyer is paying to automate. The sharper question is which segment combines independent demand, outcomes that can be verified and preparation costs that do not swallow the value.
This article builds that comparison: a Support Data Opportunity Map with stated rating rules, a separate look at preparation effort for simple tickets and complete case episodes, a thesis about where demand and verifiable outcomes align, the evidence that would prove it wrong, and the diligence questions that follow.
Five decisions inside one support queue
The five segments differ in what the agent must do and in what proves it did it right. Public benchmark design reflects the same split. The tau-bench retail policy lets an agent cancel or modify pending orders, return or exchange delivered orders and update an address; it requires explicit user confirmation before any consequential database action, and allows a transfer to a human agent only when a request cannot be handled within the agent's scope.
The airline policy in the same benchmark shows entitlement as a decision of its own. Whether a flight can be cancelled depends on booking time, fare class and travel insurance, and the policy notes that the API does not check these rules for the agent, so the agent must. Compensation certificates are limited to certain membership tiers, insured travelers or business-class passengers who complain about cancelled or delayed flights. The action is easy; the judgment is reading the rule correctly. The five segments:
- Resolution: answering a question or fixing a problem, such as a how-to request or a configuration fault. The record is the conversation and the fix; the outcome is often only a closed status.
- Escalation: deciding that a case needs a person, a specialist team or a different queue, and handing it over with context. The outcome is whether the hand-off was needed and complete.
- Refunds and returns: deciding whether money or goods move back, how much and by which method. The outcome is a transaction in a billing or order system.
- Entitlement: checking what a customer is owed under a plan, warranty, membership or fare rule before acting. The outcome is an action that can be checked against a written rule.
- Policy exceptions: approving a departure from written policy, with an approver, a reason and a consequence. The outcome is the approval record and what happened afterwards.
The Support Data Opportunity Map
The map rates each segment on four criteria with stated rules, using public evidence dated in the methods note. Ratings are relative across the five segments, not absolute scores, and they describe the typical case in a mid-sized support operation rather than any particular company.
Rating rules: outcome evidence is strong when the right answer leaves a record outside the conversation, such as a refund transaction or an order status. Preparation burden is high when a usable record needs linked histories from several systems or policy versions. Privacy exposure is high when the decision turns on personal circumstances, payment details or hardship. The demand signal must be a dated primary source showing that developers or deployers are building or testing that exact decision; a market-size estimate does not count.
Read across the rows, refunds and returns and entitlement are the only segments that pair strong outcome evidence with a dated public signal that agents are being built and graded on that decision. Resolution has the most visible deployment and the weakest outcome evidence. Policy exceptions may be the most valuable records per case and are the hardest to supply.
| Segment | What a usable record must contain | Outcome evidence | Preparation burden and privacy exposure | Public demand signal (dated) |
|---|---|---|---|---|
| Resolution | Conversation, product context, the fix applied | Weak to medium: closed status, reopens, satisfaction rating | Low burden; medium exposure from free text | Klarna says its AI assistant handled two-thirds of its customer service chats in its first month (February 2024) |
| Escalation | Conversation, agent scope, hand-off note, downstream outcome | Medium: whether the receiving team needed the case | Medium burden; medium exposure | tau-bench retail policy (2024): transfer to a human only when a request is out of scope |
| Refunds and returns | Order, payment method, policy version, action taken | Strong: a refund or return transaction in the order system | Medium burden; high exposure from payment data | tau-bench retail tasks: cancel, return and exchange, graded on the final database state |
| Entitlement | Customer tier, plan or warranty terms, rule text, action | Strong: the action can be checked against a written rule | Medium burden; medium exposure | tau-bench airline policy: cancellation and compensation turn on fare class, insurance and tier |
| Policy exceptions | Request, blocking clause, approver, reason, later outcome | Medium to strong when approval and follow-up are linked | High burden; high exposure from hardship and staff details | Not separately evidenced: exceptions surface inside refund and compensation policies |
Abundant tickets, thin usable coverage
Support records are abundant; records that document a decision together with its outcome are not. The labor base is large: the U.S. Bureau of Labor Statistics counted about 2,666,000 customer service representative jobs in 2025 and projects a 5 percent decline by 2035, with about 289,500 openings a year on average. A separate BLS analysis of its 2024-34 projections links limited future demand for customer service representatives to AI and automation.
All that work produces records, which is why volume is the wrong test. A queue of a million tickets can hold few refund decisions with the payment record attached, few escalations linked to what the receiving team found, and almost no exceptions with an approver and a reason. Usable coverage is the count of complete decision records in each segment, and it can be a small fraction of the ticket count.
Buyer preferences narrow coverage further. OpenAI's data partnerships page says it wants large-scale datasets that are not already easily accessible online and prefers data that shows human intent, such as conversations, but that datasets carrying sensitive or personal information, or information belonging to a third party, are not what it is looking for. Support conversations fit the first description and routinely fail the second until they are prepared, which moves value toward whoever can de-identify and scope them credibly.
Simple tickets and complete case episodes need different preparation budgets
Estimate preparation separately for the two products, because a complete case episode can take many times the work of a single ticket. A simple ticket is one conversation with a status and a category. A complete episode links the conversation to every update, the systems the decision touched, the policy version in force and the outcome after a waiting window.
The event history is the expensive part. Zendesk's API documentation describes ticket audits as a read-only history of all updates to a ticket, where each audit is a single update made of one or more events, and it points bulk change capture to the incremental ticket event export rather than the audits endpoint. Rebuilding an episode means exporting that history, joining it to billing or order records and attaching the rule text that applied on the date.
The preparation gap is where segment economics diverge. Resolution data can often be supplied as simple tickets; refund, entitlement and exception records lose most of their grading value without the episode. An investor should ask which product a company's figures describe, and whether the effort per usable record is measured by segment.
| Preparation step | Simple ticket | Complete case episode |
|---|---|---|
| Export | Ticket fields and conversation | Ticket, full update history and linked billing, order or CRM records |
| Context | Category and product | Policy version in force, customer tier or plan, prior contacts |
| Outcome | Closed status and a satisfaction rating if present | Transaction, reversal, reopen or dispute after a set window |
| De-identification | Names, contact details and identifiers in free text | The same, plus payment data, account links and staff names across systems |
| Review | Sample check of text quality | Join accuracy, timeline order and outcome labeling checks |
Where buyer demand and verifiable outcomes align
The thesis: in customer support, independent buyer demand and verifiable outcomes align most clearly in policy-bound transactional decisions, refunds and returns and entitlement checks, because the correct answer leaves a state in a system of record that a grader can check. It is a thesis to test, not a finding.
The public evidence for it is dated and checkable. tau-bench grades an agent on whether the database ends in the expected state, and its reward code sets the score to zero when the database changes are not correct. Its README describes conversations between a language agent and a user simulated by language models, and now points to a successor benchmark that adds a banking domain and a voice evaluation modality, which is a sign that researchers keep extending policy-bound support tasks.
Resolution looks different. Klarna says its AI assistant had 2.3 million conversations in its first month, two-thirds of its customer service chats, and that resolution time fell from 11 minutes to under 2 (February 2024). Deployment at that scale suggests routine resolution is already being automated from knowledge content and conversation logs, which would leave the marginal value of new resolution data below that of the decisions automation still gets wrong. That is an inference from one company's statement, not a measured market fact.
What would confirm the thesis: buyer requests that name refund, return or entitlement tasks with acceptance tests tied to transaction outcomes; benchmarks that keep adding policy-bound domains; and suppliers able to show complete episodes by segment. What would refute it: buyers paying as much for resolution transcripts as for linked episodes, or synthetic environments matching real episodes on buyers' acceptance tests.
What could make the support segmentation thesis wrong
Five developments could shrink the opportunity or move it to a different segment. Each comes with a source an investor can track.
- Synthetic users and environments could substitute. tau-bench simulates the customer with a language model and builds its own policy-bound databases; if synthetic environments satisfy buyers' acceptance tests, the premium for real episodes falls.
- Rights over customer-written text are contested. FTC staff wrote in February 2024 that a company quietly changing its terms of service or privacy policy to use customer data for AI training could be acting unfairly or deceptively, so support histories collected under older notices may need new steps before licensing.
- Privacy exposure is structural. NIST's Generative AI Profile (AI 600-1, July 2024) lists data privacy, including leakage and de-anonymization of personal information, among its 12 generative AI risks, and refund and exception records carry payment and hardship details.
- Buyers may stay concentrated. In its quarterly report for the period ended March 31, 2026, Reddit traces substantially all of its licensing contract value to just two partners; a support data supplier facing few buyers carries renewal and pricing risk.
- The labor base is shrinking. BLS projects customer service representative employment to fall 5 percent between 2025 and 2035, so future records may be produced partly by automated agents, which changes what those records can teach.
Questions to ask a support data supplier, SourceX included
Ask for evidence by segment, in counts with denominators, and put the same questions to every company in the market.
SourceX is one application of this thesis. It turns selected business records into licensed AI datasets, and for each one it finds and vets the supplier, reviews rights, contracts, prepares the records, and runs delivery and payment. Nothing is delivered before the supplying company has approved what is in scope and on what terms and has signed, and the datasets AI buyers receive are sourced on request, not pulled from stock.
SourceX has not published segment, supply or commercial metrics; investors get its company and commercial figures directly. It is raising under Rule 506(c), which limits purchasers to accredited investors, and materials can be requested on the investor page, where any of the questions below can also be raised.
Checklist · 6 items
- How many complete decision records can you source in each segment, and what share of a typical supplier's tickets qualify?
- Which segments do buyer requests name, and how many distinct buyers issued them?
- For refunds, returns and entitlement, what share of records link to a transaction or a dated rule version?
- What does a usable record take to prepare as a simple ticket versus a complete episode, measured by segment?
- Under which notices and contracts were the support records collected, and who confirmed the right to license them?
- Which synthetic benchmark result would make real episodes unnecessary for your buyers?
0 of 6 ticked. Tick each item as you go.
Methods note: how the five support data segments were rated
The Support Data Opportunity Map, the preparation comparison and the thesis are editorial frameworks built for this article on 2026-10-26. Ratings are relative judgments against the rules stated above, informed by the tau-bench repository (retail and airline policies, reward code and README, read in October 2026), Zendesk's ticket audit documentation, OpenAI's data partnerships page, Klarna's February 2024 announcement, BLS occupational data and projections analysis, FTC staff guidance, NIST AI 600-1 and Reddit's quarterly report for the period ended March 31, 2026.
Not measured: record volumes or usable coverage by segment at any company, preparation effort in hours or money, prices, buyer budgets, and SourceX's own supply or demand. Klarna's figures are the company's own statements. Each missing measure requires company records that have not been published.
Frequently asked questions
Is customer support data valuable for AI training?
Some of it is, and the value depends on the decision a record documents. Refund, return and entitlement records linked to the transaction or rule they produced can grade an agent objectively; transcripts with only a closed status are weaker evidence. Volume alone is a poor guide.
Which customer support data segment looks most investable?
On public evidence as of October 2026, policy-bound transactional decisions such as refunds, returns and entitlement checks pair the clearest outcome evidence with dated signals that agents are being built and graded on them. Policy exceptions may be more valuable per record but are harder to supply. Treat this as a thesis to test with segment-level data.
Why does a complete support case episode cost more to prepare than a ticket?
An episode links the conversation to its full update history, the billing or order records it touched, the policy version in force and the outcome after a waiting window. Each link adds an export, a join and a de-identification step across systems.
Can synthetic data replace real customer support records?
For some uses, possibly. Public benchmarks such as tau-bench already simulate customers with language models in policy-bound environments. Whether synthetic episodes can replace real ones depends on whether they pass a buyer's acceptance tests on real cases, which is the main substitution risk to track.
Does SourceX publish its support data supply or deal metrics?
No. Supply, segment and commercial metrics are not public, and SourceX gives its company and commercial metrics to investors directly. A materials request through sourcex.si/investors is the place to ask how supply would be counted by segment, and only verified accredited investors may invest in any offering.
Sources
- BLS counts about 2,666,000 customer service representative jobs in 2025, projects a 5 percent decline by 2035 and about 289,500 openings a year on average. Source
- A BLS analysis of 2024-34 projections links limited demand for occupations such as customer service representatives to AI and automation. Source
- Klarna says its AI assistant had 2.3 million conversations in its first month, two-thirds of its customer service chats, with resolution time down from 11 minutes to under 2 (February 27, 2024). Source
- FTC staff (February 2024): quietly changing terms of service or privacy policies to use customer data for AI training could be unfair or deceptive. Source
- OpenAI seeks large-scale datasets not easily accessible online, prefers data showing human intent such as conversations, and is not seeking sensitive or personal information or third-party information. Source
- Zendesk ticket audits are a read-only history of all updates to a ticket, each audit a single update of one or more events; bulk change capture should use the incremental ticket event export. Source
- Reddit's quarterly report for the period ended March 31, 2026 says substantially all contract value associated with its licensing revenue is derived from two partners. Source
- NIST AI 600-1 (July 2024) lists data privacy, including leakage and de-anonymization of personal information, among 12 generative AI risks. Source
- tau-bench emulates conversations between a user simulated by language models and a language agent with domain tools and policies; its successor adds a banking domain and a voice modality. Source
- tau-bench's retail policy covers cancelling or modifying pending orders and returning or exchanging delivered orders, requires explicit confirmation before consequential actions, and allows transfer to a human only for out-of-scope requests. Source
- tau-bench's airline policy ties cancellation to booking time, fare class and insurance, says the API does not check these rules for the agent, and limits compensation certificates to certain tiers, insured or business-class travelers. Source
- tau-bench's reward code compares the final database state with the expected state and sets the reward to zero when the database changes are not correct. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.