Before you buy an enterprise workflow dataset, test whether a small sample supports the task you plan to train or evaluate. Ask for records from real cases, a description of the links between them, a count of what is missing, and license terms for your intended use. A long format list is useful, but it cannot answer those questions by itself.
Ask for a task-shaped sample
Give the provider a specific request, such as “find the documents behind a project scope change” or “identify the steps from a support ticket to its resolution.” Ask for ten cases from different projects or time periods, including ordinary and difficult examples. For each case, request a manifest that names the source files, their dates, identifiers, and any excluded material.
Try to answer the task using only the sample. If an expert needs to explain hidden context, write down what is missing. Keep the provider's sample selection method, since a collection of clean examples may not reflect the full archive.
Run five acceptance checks
| Check | Question to answer on the sample |
|---|---|
| Coverage | What share of requested cases has the input and final records? Which project types or years are absent? |
| Links | Can you join documents, messages, models, and actions through stable IDs rather than guessed filenames? |
| Time | Can you tell which version existed before the decision? Are revisions and duplicate exports identified? |
| Outcome | Is there a record that checks the answer or action, or will your team have to label it? |
| Rights and handling | Which source agreements, personal information, and third-party files affect the licensed scope? What does the contract permit? |
Add task-specific checks for language, units, formats, and software versions. Split training and evaluation by project or source company when related cases could otherwise appear in both sets. Record failures as well as passes. A missing outcome is a different problem from an unreadable file.
The distinction matters for agent work. WONDERBREAD provides workflow demonstrations and separate evaluation material. WorkArena defines tasks with validation. A company archive may hold useful raw records without either of those prepared layers. Price the work needed to create them separately from the files themselves.
Compare the offer with the delivered scope
Ask the provider to state what exists now, what will be prepared after agreement, and what is estimated. Confirm the unit of delivery: files, projects, linked cases, action traces, or validated tasks. Then request a count of records excluded for missing files, duplicate content, and rights review.
The SolarEPC-1TB and SolarThermal-4TB cards describe available Orla archives, with estimated sizes and project counts. SolarEPC spans design, purchasing, work orders, commissioning, and handover; it is primarily in Portuguese. SolarThermal contains engineering, project, and business files across multiple formats. Their public samples are selections of technical files, not representative workflow samples or prepared agent tasks. Licensing and pricing are on request.
If either archive is relevant, send Orla the task, the project coverage, the required file links, and the rights scope you need. Ask for a sample built around those acceptance checks. For help defining the task first, read what makes enterprise workflow datasets useful for AI.