The next data market may be workflows and agent outcomes, not document volume

The next data market may place more emphasis on workflows and agent outcomes than on document volume, but this is a hypothesis, not an established market forecast. Research shows that agent benchmarks can use defined interactive environments, issue histories and task outcomes. It does not prove that companies will sell such data, that buyers prefer it, or that workflow records are inherently more valuable than documents.

The useful distinction is between three kinds of evidence: historical documents, captured real workflows, and agent-generated trajectories. Each can answer different questions and has different limits.

Documents preserve content, not always process

A collection of manuals, reports or messages can help answer questions about language and domain knowledge. But documents may omit which information was available at a particular decision, what action followed, whether an action worked, and how the result was verified. A document archive may still be useful for its intended task; it simply should not be treated as a complete interaction history by default.

Volume is a weak substitute for structure. One million pages with unclear origins and no task boundary may support fewer decisions than a smaller, well-described source. This is an editorial inference about evaluation design, not a measured market price or rule.

What can retained workflows show that documents omit?

Operational records may connect a request, context, decision, action and later outcome. That sequence can support questions about workflow behavior, provided timestamps, revisions and links are reliable. Many systems retain only partial records. An outcome may be absent, delayed, overwritten or produced by a different process than the one being studied.

Before claiming that a dataset contains trajectories, define what counts as one: state, observation, action, transition, outcome and source provenance. Distinguish observed events from reconstructed links and inferred intentions. A chronological log is not automatically a causal explanation.

How do agent trajectories differ from human work histories?

An agent trajectory can record the prompts or observations shown, tool actions taken, environment changes, evaluations and terminal state. In a simulator, the system can reset and reproduce conditions more easily, but the environment may not capture real constraints. In a historical dataset, the actions were performed by people or existing systems, not the evaluated agent. In an interactive replica, actions can be tested under constraints without exposing a live system.

Research examples illustrate these distinctions. WebArena describes a reproducible web environment for autonomous agent tasks. OSWorld studies agents in computer environments using executable tasks. SWE-bench builds software tasks around real GitHub issues and pull requests and tests generated patches. These works show ways to define tasks and outcomes; none establishes a commercial market shift or the availability of company data.

When might documents or simulation be the better evidence?

Hypothesis one: workflow context becomes more useful. If evaluation needs to test multi-step decisions, linked events and outcomes may be more informative than disconnected text. The test would be whether a defined task can be evaluated more reliably using those links, after accounting for collection and review cost.

Hypothesis two: documents remain sufficient for many tasks. Retrieval, classification and explanation tasks may not require full action histories. A document source can be easier to scope and may have clearer permissions. The test would be whether documents support the target task at an acceptable quality and cost.

Hypothesis three: synthetic or simulated environments substitute for scarce records. Controlled environments can produce repeatable trajectories and reduce reliance on sensitive company data. The test would compare simulation behavior with the real use context and identify where the simulator's assumptions fail.

These hypotheses can coexist. Which one matters depends on the task, rights, risk and evidence required.

Which evidence would change the sourcing choice?

These are proposed comparisons, not measured results. The buyer should identify the added decision evidence and preparation cost before asking an owner to reconstruct a larger archive. The owner should ask which retained links make that evaluation possible, rather than market every activity log as an agent trajectory.

What evidence would test the workflow-data thesis?

To test whether workflow outcomes are commercially or technically more useful, researchers would need comparable, permission-cleared examples; task-specific evaluation; documented provenance; cost of preparation; and transparent limitations. They would also need to distinguish public research artifacts, provider marketing statements, company sources and completed transactions. The available benchmark papers do not supply a market-wide comparison of those categories.

For an individual source, begin with an evidence map: task, inputs, actions, outcomes, timestamps, revisions, sample method, rights, privacy, and gaps. Ask which specific decision becomes possible because the workflow data are present. If no one can name that decision, “workflow data” may simply be a broader label for an archive.

What should an owner or buyer do with this thesis?

Document volume alone is not a useful proxy for evaluation value. But the alternative is not to assume that every real workflow is valuable. Workflow sequences can be incomplete, biased, sensitive, hard to link and difficult to authorize. Agent-generated trajectories can be reproducible yet unrealistic. Good sourcing identifies which evidence exists and what claim it can support.

Avelence's service model includes research and dataset positioning. It does not claim that workflow data are a superior asset class, predict market demand, or guarantee a buyer. The question for an owner is whether a specific source can support a clearly defined and permissioned use.

The W3C PROV data model is a useful vocabulary for describing which entities were used, what activities changed them, and which agents were responsible. An evaluation report can apply that distinction to source episode, transformation, and annotation so that the tested material can be reconstructed.