What manufacturing data can be licensed for AI?

Start with a collection, not the entire factory

Choose records produced by one identifiable process, team or site over a known period. A collection might be work orders for a class of machines, inspection records for a defined product family, or company-written procedures with revision history. These are illustrative starting points, not evidence of current demand. Keep healthcare, finance, legal and consumer datasets outside this initial industrial review.

Do not combine a maintenance system, supplier document library and production database simply because they are accessible. Each can have different rights holders, sensitivities and interpretation requirements. A bounded scope makes exclusions visible and helps you explain what remains useful. Company-generated is a starting hypothesis to verify, not a blanket declaration of ownership.

• Name the process and the responsible team.
• Record the time period, format, units and known changes in collection method.
• Separate original company records from imported, purchased or customer-provided content.

Inventory four practical record families

Maintenance records can connect reported symptoms, machine context, actions and outcomes. A closed ticket is not necessarily a successful repair: check what closure means and whether later recurrence is recorded. The companion maintenance guide walks through this distinction in detail.

Quality inspections can connect measurements to the method, tolerance or acceptance rule used at the time. Record calibration context and changed specifications where available. A pass/fail field without the applicable criterion may be difficult to interpret. Customer drawings and specifications need their own permissions review.

Process documentation can explain sequences, decision points and exceptions. Identify company-written text separately from supplier manuals and customer instructions. Preserve revision dates so an older instruction is not mistaken for current practice. Production records can add batch, line, product-family or changeover context, but confidential recipes and commercially sensitive volumes may need exclusion.

• Maintenance: fault, machine class, intervention and recorded outcome.
• Quality: measurement, method, applicable criteria and exception handling.
• Documentation: authoring origin, process scope, version and review status.
• Production: permitted process context, timestamps, units and known missing periods.

Use questions to qualify context, not to fabricate a score

Ask whether someone outside the team could understand the records using a permitted data dictionary and a short process description. Abbreviations, changed machine identifiers and local terminology can make seemingly structured exports ambiguous. Record those limitations before investing in a cleanup.

NIST describes manufacturing AI work involving measurement science, standards and interoperability evaluation. That establishes industrial context, not commercial interest in your collection. Possible uses to discuss include retrieving permitted troubleshooting guidance or evaluating a model against documented inspection decisions. A model use must be defined before deciding whether the material is a fit.

• Can records be linked without exposing restricted identifiers?
• Are units, codes and field meanings documented?
• Are outcomes observed, inferred or simply absent?
• Would exclusions remove the context needed for the proposed use?
• Can known gaps and changes be explained without inventing missing history?

Map permissions and provenance collection by collection

Identify who created each part, under which contract, and whether your company can grant the proposed external use. Review employee and contractor contributions, supplier reports, customer specifications and confidentiality clauses. Access within the factory is different from permission for external AI training, evaluation or redistribution.

The EU Data Act sets rules concerning access to connected-product and related-service data and sharing with recipients. It is not a statement that a factory owns every machine record or can sell every document. Scope, user and data-holder roles, trade secrets, personal data and contractual context need review. Do not use the Act as a shortcut around a rights inventory.

Maintain a provenance note for each collection: origin, author or system, relevant agreements, permitted uses, uncertain rights and exclusions. Missing permissions are review items, not an automatic rejection of the whole business. Narrowing the scope may be more practical than trying to clear everything.

Non-consumer data can still identify people

Work orders and inspection notes can contain technician names, shift patterns, signatures or incident narratives. Removing names alone may leave a person identifiable through timestamps, job role, location or a rare event. Customer names can also reveal confidential commercial relationships even when they do not identify an individual.

The ICO distinguishes anonymisation from pseudonymisation: pseudonymised personal data remains within data-protection law. Assess the risk of identification in the intended release context, including other information a recipient could reasonably use. The ICO flags its anonymisation guidance as under review, so obtain current advice rather than treating this article as legal clearance.

• Identify direct identifiers and identifying combinations.
• Keep the mapping for replacement IDs separate and restricted.
• Review free text, attachments and document history, not just table columns.
• Check confidentiality and IP restrictions even if personal data is removed.

Create a description-only assessment brief

Summarise one collection in a few sentences: what process produced it, its coverage, what context and outcomes are available, how it originated and which restrictions remain unresolved. Describe approximate scope without copying real records, customer details or sensitive process settings. Note whether a responsible colleague can help verify the rights.

Do not upload exports, manuals, control-system details, credentials or identifiable examples to get an initial assessment. AssetGate provides a free qualitative read, not a monetary valuation, buyer confirmation or guarantee. If follow-up is useful, agree an authorised package, defined use, access safeguards and retention terms before any sample is shared.

• Ready to describe: origin and scope are known, with limitations stated.
• Needs review: rights, privacy or interpretability questions are documented.
• Exclude for now: credentials, restricted customer material and uncontrolled personal data.