A file is a fragment of a workflow
A purchase order records a request. The supplier response explains a constraint. A changed delivery date records a decision. Joined together, those records tell a more useful story than any one file on its own.
That context is what an AI team may need to study a capability: can an agent identify the constraint, find an authorized alternative and update the right system?
Look for a chain you can reconstruct
Start with one recurring workflow and map its inputs, decisions, tools and outcomes. Stable identifiers and timestamps help establish which records belong together and what information was available at each stage.
- Inputs: the request, policies and relevant starting state.
- Decisions: actions, alternatives and necessary approvals.
- Outcomes: what changed, what succeeded and what required correction.
- Expert context: a practitioner who can explain ambiguous cases.
Useful messiness is different from unusable data
Missing values and conflicting messages can be realistic parts of a task. Broken joins and unexplained timestamps can make the same task impossible to evaluate. Preparation should make that distinction explicit.
Do not silently fill every gap or remove every failed example. A correction or escalation may contain the evidence needed to understand a difficult decision.
Buyer fit and permissions determine value
A rare workflow is only commercially useful if a buyer needs it and the required permissions can be obtained. Large volumes of generic documents may be less useful for a specific brief than a small, coherent set of operational episodes.
Start with a metadata inventory and a capability brief. Review access, confidentiality and permitted uses before moving source records. A readiness assessment is a way to find the next question, rather than a promise of a payout.