Skip to content

AI data market

Agent data procurement investment: six purchase terms agents rewrite

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Agent rollouts change data procurement first in the specification: buyers of interactive work need starting states, permitted actions and outcomes a grader can check, not just labeled records. Whether agent data procurement becomes an investment-grade spending stream is unproven. Labs already grade end states, but no dated cohort of paid episode purchases is public. Count written requests and paid commitments by specification and commitment class, not launches.

Agent run

Agent data procurement

This page covers agent data procurement investment: six purchase terms agents rewrite.Illustrative

Key takeaways

  • Agent work is judged by the state it leaves behind, so the purchasable unit shifts from a labeled record to an episode with a starting state, actions and a checkable outcome.
  • Six purchase terms change with it: unit, contents, acceptance test, rights scope, refresh cadence and privacy exposure.
  • Evidence of the shift is strong on the lab side, where benchmarks and fine-tuning tools grade outcomes, and thin on the buyer side, where no dated cohort of paid episode purchases is public.
  • Capability launches and forecasts are context, not demand; Gartner itself expects more than 40 percent of agentic AI projects to be canceled by the end of 2027.
  • Measure the shift with dated written requests from distinct budget owners, coded by specification and commitment class, with denominators and open requests kept in the count.

Will agent rollouts change what AI teams pay for?

They change the specification before they change the budget, and only the budget makes an investment case. An agent is judged by what it leaves behind: a booking that exists, a refund that posted, a ticket routed to the right queue. Buying data for that kind of work means buying starting states, the actions an agent may take and outcomes a grader can check, which is a different purchase from a corpus of labeled records.

For an investor weighing agent data procurement as an investment theme, the decision is whether that specification shift is turning into paid, repeatable buying that an outside data supplier can capture. Supporting evidence would be dated written requests from distinct budget owners that name states, actions or outcomes; samples evaluated under an outcome-based acceptance test; and paid commitments, each counted against a stated denominator. Launch posts and benchmark scores show direction. They cannot show spending.

Why interactive tasks need states, actions and checkable outcomes

Because the grade comes from the environment, not from the reply. Anthropic's engineering guidance on agent evaluations defines the outcome as the state of the environment when a trial ends, and illustrates it with a flight-booking agent: the agent saying the flight is booked does not count, while a reservation existing in the database does. The guidance favors judging the final state over checking whether the agent followed a particular path.

Public benchmarks are built the same way. In τ-bench's published code, the environment replays the reference actions on a fresh copy of the database and compares a hash of the result with the database the agent produced; if the two differ, the reward is zero. The repository's leaderboard also shows why repeat trials matter to a buyer: in the retail domain, the top listed configuration succeeded on 69.2 percent of tasks in a single try but on 46.2 percent when all four of four tries had to succeed.

Training tools have moved toward graders as well. OpenAI's reinforcement fine-tuning guide describes a programmable grader that scores each candidate response, in place of a fixed correct answer, applied to an uploaded prompt dataset with a validation split. A buyer training this way needs tasks and graders, not only examples.

Evaluation of professional deliverables shows the gap that remains. OpenAI's GDPval spans 1,320 tasks across 44 occupations drawn from the nine industries contributing most to U.S. GDP, written by professionals averaging more than 14 years of experience, with 220 tasks released as a gold set. OpenAI notes that the evaluation is one-shot and does not capture work where a model must build context or improve through multiple drafts. Records of interactive work are the input that would fill that gap.

Six purchase terms that agent data rewrites

The terms are an editorial framework drawn from the sources on this page. The last column states what diligence should look for; none of it is yet observed in public data.

Compared line by line, an agent data purchase differs from a static dataset purchase in six terms, each listed with the public evidence that the difference is real and the evidence that would show buyers paying for it.

Six purchase terms that agent data rewrites
TermStatic dataset purchaseAgent episode purchasePublic evidence of the shiftEvidence it is paid
UnitRecords or documents, priced by count or volumeEpisodes: one task from starting state to outcomeτ-bench and GDPval define work as tasks, not rowsOrders priced per task or episode
ContentsFields, text and labelsStarting state, tools and policy, action trace, final stateAnthropic's guidance separates the transcript from the outcomeRequests that name states, actions or outcomes
AcceptanceLabel agreement or field completeness on a sampleOutcome checked by a grader or end-state comparisonτ-bench scores zero unless the end state matchesContracts with grader-based acceptance tests
RightsUse copies for trainingAlso replay environments, policies and tool schemasNo public contract terms found for episode dataLicenses that name replay or evaluation use
RefreshOne delivery, occasional updatesVersioned as tools, policies and benchmarks changeτ-bench's maintainers point users to a successor with fixed tasksRenewals or change orders after a tool update
ExposureRecords reviewed for personal dataAction logs and screens reviewed as wellNo public measure foundPrivacy terms covering action traces
A two-column comparison of a dataset purchase and an agent purchase across six terms, unit, contents, acceptance, rights, refresh and exposure, with each changed agent term highlighted in turn.
Editorial framework, not measured demand: six purchase terms that agent data changes.

Which signals count as agent data demand, and which only resemble it

Only a dated, written request from a budget owner counts as the first rung of demand; launches and forecasts are context. The grid crosses five commitment classes with what each would look like for episode data, and places the public evidence found for this article in its class.

The public record is lopsided. Capability evidence is plentiful: Anthropic released computer use as a public beta on October 22, 2024, called it experimental and at times error-prone, and reported 14.9 percent on the OSWorld benchmark using screenshots alone. Forecasts are plentiful too: Gartner predicted in August 2025 that 40 percent of enterprise applications would feature task-specific AI agents by 2026, up from less than 5 percent in 2025. Neither is a purchase.

Paid data licensing is publicly documented, but for content rather than episodes, and it is concentrated. Reddit's FY2024 Form 10-K states that "substantially all of the contract value associated with our licensing revenue is derived from two of our partners." OpenAI's data partnerships page asks for data that is not already easy to reach online, which is a request for access, not a specification of states, actions and outcomes.

Which signals count as agent data demand, and which only resemble it
Commitment classDefinition used hereWhat to count for episode dataPublic evidence found, dated
AnnouncedA launch, forecast or product claimNothing; record as context onlyAnthropic computer use beta, Oct 22, 2024; Gartner agent forecast, Aug 2025
RequestedA dated written request from a named budget ownerRequests naming states, actions or outcomes, per distinct buyer per quarterOpenAI's partnerships page seeks non-public data but specifies no episodes
EvaluatedA sample tested under a stated acceptance testTests that use an outcome grader or end-state checkGDPval and τ-bench grade tasks, but as benchmarks, not purchases
PaidA signed order or minimum commitmentPaid commitments with an episode specificationNone public for episode data; Reddit's content licensing is paid but concentrated
RenewedA repeat order after the first deliveryRenewals or change orders after a tool or policy changeNone public; benchmark upkeep shows maintenance is needed

Measuring the shift without counting headlines

Code buyer requests, not news. The unit of analysis is one written request from one distinct budget owner, dated on receipt. Each request carries two codes: a specification class (record, episode or mixed, where episode means it names at least one of a starting state, permitted actions or a checkable outcome) and a commitment class from the grid above, updated as the request moves.

SourceX has not published request or commitment data under this design, and the results are not yet published. The company sources datasets on request through its buyer form, and it shares company and commercial metrics with investors directly rather than on public pages.

The rest of the design follows from that unit of analysis:

  • Sampling frame: every request received in a fixed window, with inbound and outbound requests reported separately, because outreach inflates apparent interest.
  • Cohorts: requests grouped by quarter received, so progress to evaluation and payment is read within a cohort rather than across a pooled pipeline.
  • Denominators: every share stated against all requests in its cohort, including requests that went silent.
  • Censoring: open requests kept in the denominator as open, never dropped and never counted as lost.
  • Coding: two people classify a sample of requests independently and report how often they agree.
  • Interviews: budget owners at buyers that did not pay are interviewed too, so reasons for not buying are as visible as reasons for buying.
  • Exclusions: no demand inferred from funding rounds, launch posts or hiring announcements.

Could agents improve without anyone buying episode data?

Yes, and this is the explanation an investor should test first. Capability progress can come from data a lab builds itself: synthetic environments, simulated users and tasks written by contractors. τ-bench's own conversations pair the agent with a user simulated by a language model, which shows how much of an environment can be generated rather than licensed.

Enterprise rollouts may also stall before they create data budgets. Gartner predicted in June 2025 that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, and it warned about agent washing, the relabeling of existing chatbots and automation tools as agents. An announced agent is weaker evidence of data spending than it looks.

What would still favor licensed episodes is the part generation cannot supply: real policies, real exceptions and outcomes that actually happened. The test is empirical. If requests that name states, actions and outcomes do not climb the commitment classes in successive cohorts, the alternative explanation wins and the thesis should be dropped.

What could make the agent data procurement thesis wrong

Five risks could break the thesis, and each comes with a signpost an investor can watch.

At a glance

What could make the agent data procurement thesis wrong

  • Buyer concentration: if a handful of labs buy most episode data, suppliers inherit their bargaining power. Paid content licensing already shows the pattern in the two-partner concentration Reddit disclosed in its FY2024 Form 10-K.
  • In-house supply: labs may build environments and tasks internally and leave outside suppliers the commodity parts. Signpost: written requests that ask only for raw records, not episodes.
  • Rollout attrition: if Gartner's cancellation forecast proves right, fewer deployed agents means fewer budgets for the data behind them. Signpost: pilots that end without a production deployment.
  • Rights to replay: episodes can include screens, tool schemas and policies owned by software vendors or other third parties, which a license from the data owner may not cover. Signpost: deals that stall in rights review rather than technical review.
  • Fast decay: benchmarks and tool interfaces change, and τ-bench's maintainers now direct users to a successor repository with fixes to the airline and retail tasks, so episode data may need paid upkeep that buyers decline to fund. Signpost: renewal rates after a tool update.

Questions for any agent-data supplier, SourceX included

Ask for evidence at the level of the grid, with dates and denominators. A supplier that answers with launch news or total pipeline has answered a different question.

For SourceX specifically, the public facts are these: on behalf of companies licensing operational records to AI developers, it handles everything from qualifying and sourcing the supply to reviewing rights, contracting, preparing and delivering the data and handling payment, and a supplier's approval of scope and terms, plus its signature, must come before any data is delivered. Its raise is from accredited investors under Rule 506(c), and materials are requested through its investor page. None of that is evidence of agent data demand; the questions below are how to test it.

  • How many dated written requests in the last four quarters named starting states, actions or outcomes, from how many distinct buyers, out of how many requests in total?
  • What share of each quarterly cohort reached an evaluated sample, and what share reached a paid commitment?
  • Which acceptance tests did buyers use, and did any grade outcomes rather than label agreement?
  • Which rights cover replaying an environment, including screens, tool schemas and policies?
  • What happens to a delivered episode set when the source software or policy changes, and who pays for the update?
  • How concentrated are requests and commitments by buyer?

Methods note: how the purchase-term matrix and evidence grid were built

The six-term matrix and the five-class evidence grid are editorial frameworks. They draw on public sources listed on this page, each checked on its publisher's own domain or repository on October 26, 2026: Anthropic's evaluation guidance and computer use announcement, τ-bench's code and repository, OpenAI's GDPval, data partnerships and reinforcement fine-tuning pages, two Gartner press releases and Reddit's FY2024 Form 10-K.

Not measured: no buyer interviews were conducted for this article, no request data from SourceX or anyone else is reported, and no market size is estimated. Benchmarks sit in the grid as evidence of how work is graded, not of purchases, and placing each public item in a commitment class is a judgment a reader can recheck against the cited pages.

Frequently asked questions

What is agent data, compared with ordinary training data?

Ordinary training data is usually records with fields and labels. Agent data packages a task: the starting state of a system, the tools and policies an agent may use, the actions taken and a final state a grader can check. Agent benchmarks such as τ-bench score the final database state, which is why the starting state and the outcome matter as much as the text.

Do agent benchmarks show that companies are buying agent data?

No. Benchmarks show how agent work is graded and where models fall short, but researchers and labs build them; nobody buys them under procurement contracts. Evidence of buying is a dated written request, a sample evaluated under an acceptance test, or a paid commitment, each counted against a denominator.

What evidence would show agent rollouts are changing data budgets?

Successive quarterly cohorts of written requests in which a growing share name starting states, actions or outcomes and reach paid commitments, from a growing number of distinct buyers. Funding rounds, launch posts and forecasts do not count, because they can rise while purchases stay flat.

Why does buyer concentration matter to an agent data investment?

If a few buyers account for most purchases, a supplier's revenue depends on their renewals and terms. Content licensing offers a public example: in its FY2024 Form 10-K, Reddit discloses that two partners are the source of substantially all of its licensing contract value. Ask any supplier how concentrated its requests and commitments are.

Does SourceX publish its buyer request data?

No. None of SourceX's request, evaluation or commitment figures are public; the company provides company and commercial metrics to investors directly. This article sets out a measurement design an investor can ask any supplier to follow; materials can be requested through sourcex.si/investors.

Sources

  • Anthropic's agent evaluation guidance defines the outcome as the environment's state at the end of a trial, using a flight booking that must exist in the database, and favors end-state grading. Source
  • τ-bench's environment code replays reference actions on a fresh database and gives zero reward unless the agent's resulting database hash matches. Source
  • τ-bench's repository leaderboard lists retail pass^1 of 0.692 and pass^4 of 0.462 for the top configuration, describes LM-simulated users, and directs users to a successor with fixed airline and retail tasks. Source
  • OpenAI's reinforcement fine-tuning guide describes a programmable grader that scores each candidate response instead of a fixed answer, with a prompt dataset and validation split. Source
  • OpenAI's GDPval covers 1,320 tasks across 44 occupations in the top nine U.S. GDP industries, written by professionals averaging 14+ years of experience, with a 220-task gold set; it is one-shot. Source
  • Anthropic released computer use as a public beta on October 22, 2024, described it as experimental and error-prone at times, and reported 14.9 percent on OSWorld with screenshots only. Source
  • Gartner predicted in August 2025 that 40 percent of enterprise apps will feature task-specific AI agents by 2026, up from less than 5 percent in 2025. Source
  • Gartner predicted in June 2025 that over 40 percent of agentic AI projects will be canceled by end of 2027 due to escalating costs, unclear business value or inadequate risk controls, and warned of agent washing. Source
  • Reddit's FY2024 Form 10-K states that substantially all of the contract value associated with its licensing revenue is derived from two of its partners. Source
  • OpenAI's data partnerships page seeks large datasets not already easily accessible online and excludes sensitive, personal or third-party information. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies