Portfolio inventory

EU AI Act Article 50 Inventory for Document AI Teams

Before writing disclosure copy or buying detection tooling, document AI teams should start with one current inventory showing which systems interact with users, generate or transform content, publish output, and trigger Article 50 review.

Last verified: 2026-07-15. This page is operational guidance, not legal advice.

Start with the inventory, not the conclusion

Article 50 requires different evidence depending on the system and role. A chatbot disclosure is not the same artifact as a machine-readable marking control, and a provider duty is not the same as a customer publication label. The inventory is the control surface that keeps those decisions reusable.

Answer first: an Article 50 inventory should identify every document AI workflow, map it to a possible Article 50 paragraph, name the provider and deployer, record any exception analysis, and link to marking, disclosure, contract, and review evidence.

Fields to track

FieldWhy it matters
System name and ownerPrevents orphaned AI workflows when product, support, and documentation teams each use different tools.
Workflow typeChatbot, summarizer, translator, drafting assistant, classifier, extraction pipeline, or publishing workflow.
Content typeText, image, audio, video, mixed media, or structured document output.
EU use signalEU customers, EU users, EU publication, EU deployer, or EU market placement.
Provider and deployerArticle 50 separates provider duties and deployer duties; the inventory should not blur them.
Article 50 trigger50(1) interaction notice, 50(2) marking, 50(3) exposed-person notice, or 50(4) publication disclosure.
Exception decisionFor example, whether the system is only assistive editing or does not substantially alter input data or semantics.
Evidence linkUX screenshot, marking test, counsel review, contract clause, release ticket, or Confluence page.
Review dateShows the inventory is current when procurement, counsel, or a customer asks.

Document AI examples

Customer support chatbot

Usually starts with 50(1): direct interaction disclosure. If it generates text output, review 50(2) as well.

Contract summarizer

Likely needs an applicability assessment because summarization can substantially alter semantics.

Drafting assistant

Track generated paragraphs, downstream publication, and whether output marking is technically implemented.

CMS publication workflow

Separate the tool provider's marking duty from the publisher's disclosure or editorial-control evidence.

Make it a living record

A spreadsheet can start the inventory, but it becomes fragile once product teams ship new workflows and customer contracts assign duties differently. In Confluence, the inventory can sit next to AI policy, vendor review, data protection, and product documentation, while the controlled terms are governed in Compliance Glossary.

Useful Confluence structure: one inventory page, one term owner, one review cadence, and links to Article 50 terms such as provider, deployer, AI system, deep fake, AI-generated text, and public interest publication.

Sources

Compliance for Confluence

See how Compliance for Confluence turns approved terminology into audit evidence inside Confluence.