EU AI Act Article 50: Disclosure Rules for Agreement AI

EU AI Act Article 50 requires that people be told when they are interacting with an AI system, unless that is already obvious, and it became enforceable across all 27 EU member states on August 2, 2026. If your agreement workflow puts an AI system in front of a signer, counterparty, or employee anywhere in the pipeline, from a Docusign Iris-drafted clause to a Slack bot negotiating on your behalf, that duty already attaches to a specific point in your Docusign IAM setup, not to a general AI policy sitting in a drawer.

Most compliance write-ups explain Article 50 in the abstract: four scenarios, two roles (provider and deployer), one set of fines. That is useful background, but it does not tell you which button in Docusign Workflow Builder needs a disclosure step attached to it. This article does that part.

What Does EU AI Act Article 50 Require Starting August 2, 2026?

Article 50 of the EU AI Act (Regulation (EU) 2024/1689) sets out four distinct transparency duties, and it splits them between providers and deployers:

1. AI interaction disclosure. Providers of AI systems that interact directly with people (chatbots, voice assistants, agents) must design the system so it discloses that a person is talking to AI, unless this is already obvious from the context.
2. Synthetic content marking. Providers of generative AI systems must mark outputs (audio, image, video, text) in a machine-readable format so the content is detectable as AI-generated.
3. Emotion recognition and biometric categorization disclosure. Deployers of these systems must inform anyone exposed to them.
4. Deepfake and public-interest text disclosure. Deployers must disclose when image, audio, video, or text content has been artificially generated or manipulated, with a narrower carve-out for AI-generated text published specifically to inform the public on matters of public interest.

The obligations apply immediately to all in-scope systems from August 2, 2026, regardless of when the system was placed on the market. Content generated before that date does not need retroactive labeling. There is one carve-out on timing: the machine-readable marking obligation in Article 50(2), for generative systems already on the market, got pushed to December 2, 2026 under the Digital Omnibus negotiations. The disclosure duties for direct AI interaction and deepfakes did not get that grace period. They apply now.

Non-compliance is not a slap on the wrist. Fines run up to 15 million euros or 3% of worldwide annual turnover, whichever is higher, enforced by national market surveillance authorities in each of the 27 member states. Article 50 also applies extraterritorially: any organization whose AI outputs are used within the EU is in scope, whether or not the company itself is EU-based.

Which Points in a Docusign IAM Workflow Count as AI Interaction Under Article 50?

This is the question the generic legal write-ups skip, because they are not looking at a specific product. A modern Docusign IAM deployment has AI touching an agreement at several distinct points, and each one needs to be evaluated separately against Article 50, not folded into one blanket disclosure.

Walk a typical agreement through Docusign Workflow Builder (formerly Maestro) and here is where AI shows up:

• Intake and drafting. Docusign Iris suggests clause language and AI-generated markups during contract review and negotiation.
• Review and negotiation. A counterparty gets an AI-generated summary of redlines, or an AI assistant answers their questions about the contract's terms.
• Approval routing. An internal agent evaluates risk and decides whether a clause needs legal escalation.
• Signer-facing communication. A signer interacts with a chatbot, voice assistant, or Slack bot to ask about the agreement before signing.
• Post-signature management. Iris flags obligations, renewals, or risk in a portfolio-wide review.

The Article 50 disclosure duty attaches specifically where a person interacts directly with the AI system, or where AI-generated content reaches a person in a way that is not already obvious. Internal AI-assisted drafting that a human lawyer reviews and takes editorial responsibility for is a materially different case from a chatbot a counterparty types into directly. Treat each touchpoint on its own merits rather than assuming one disclosure covers the whole pipeline.

Do AI-Drafted Clauses From Iris AI Need a Content Label?

Usually not as a standalone label, but the workflow still needs to account for two separate things. First, Iris AI-generated markups and suggested clause language are typically reviewed and edited by a human before the agreement goes anywhere, which is the kind of human review and editorial responsibility that keeps AI-assisted drafting out of the Article 50(4) public-interest text disclosure duty in the first place. Second, Iris is built with human-in-the-loop review for high-impact actions, which supports treating the final agreement as a human work product rather than unreviewed AI output.

Where this changes: if your workflow auto-inserts Iris-suggested clause language into an agreement without human review, and that agreement (or a summary of it) is later published or shared in a way meant to inform third parties about matters of public interest, you have re-entered Article 50(4) territory. That is an edge case for most commercial agreements, but it is a real one for public-sector and regulated-industry contracts. Build the review checkpoint into Workflow Builder as an explicit approval step, not an assumption someone will remember to do it.

What About Slack, ChatGPT, and MCP Agents Negotiating on Your Behalf?

This is where most Docusign IAM deployments now have the clearest Article 50 exposure, and it is growing fast. Docusign has shipped two agent-facing surfaces in 2026 that put a conversational AI directly in front of a person:

• A Slack app powered by Iris AI, connected through Model Context Protocol (MCP), which lets teams query agreement data and trigger workflows using natural language inside Slack conversations.
• A Docusign app inside ChatGPT and Codex, also powered by Iris, that lets users create, analyze, and act on agreements with natural language prompts.

Both are textbook direct-interaction scenarios under Article 50(1): a person is typing into an interface and getting answers, approvals, or drafted content back from an AI system. Unless it is already obvious to that person that they are talking to AI (a Slack bot with an obvious bot label and a clear system prompt disclosure usually clears this bar; a chat window with no branding usually does not), the provider or deployer needs to design in an explicit disclosure.

The same logic extends to any custom MCP agent you have wired into Docusign, such as a Claude-based agreement bot negotiating terms with a counterparty, or an HubSpot- or Salesforce-triggered agent that a customer interacts with during the deal cycle. Secondary legal guidance on Article 50 suggests that where a provider cannot reliably predict whether an autonomous agent will end up interacting with a human (unverified as binding statutory text, but a reasonable reading of the Commission's stated intent), the safer default is to have the agent disclose its nature whenever that is plausible, rather than only when you are certain a human is on the other end.

Practically: if your MCP agent can end up in a live conversation with an external counterparty, build the disclosure into the system prompt and the first message, not into a terms-of-service page nobody reads.

Building the Disclosure Step Into Docusign Workflow Builder

The compliance failure mode we see coming is not