Transparency You Can Verify
Article 50 of the EU AI Act has applied since August 2026. This page states our role, shows how each obligation is implemented in the platform, and publishes the marking specification so anyone can verify our AI-generated content. No certifications are claimed, and nothing here is aspirational: every mechanism described is running in production.
Who Is Responsible for What
Edge Pharos AB places Pharos Platform on the market and is the provider under Article 3 of the AI Act, also when the platform runs inside your own Azure tenant. Your organization uses the system in its operations and is the deployer, with its own Article 50 duties. The platform is built so those duties are met by what the platform does automatically, and this page plus the compliance documentation are the evidence you can hand to your own reviewers.
Disclosure on Every Channel, in Platform Code
People are told they are interacting with AI at or before first contact, on every channel. The disclosures are implemented in platform code outside the AI prompts, so an agent cannot skip or override them, and each disclosure is logged as evidence.
Voice
Every call opens with a spoken AI disclosure in the caller’s language, on phone, Teams, and browser calls alike. The disclosure is played by the platform itself as a pre-recorded notice before the AI agent speaks, so it is delivered identically on every call and cannot be talked over. If playback is ever unavailable, the platform falls back to instructing the model verbatim and verifies delivery, repeating the disclosure if the caller interrupted it.
A per-call log records that the disclosure was delivered, in which language, and by which mechanism.
Every AI-authored email carries a human-readable disclosure in the recipient’s language and a machine-readable header for automated processing. Forwarded mail is handled honestly: the AI-written note discloses, while the forwarded original is never relabeled.
The disclosure and header are appended by the sending code path itself, on every message without exception.
Chat and Teams
New conversations open with a clear notice that the counterpart is an AI assistant, and every reply carries the channel’s native AI-generated label.
The label is attached by the platform to each outgoing message, not generated by the agent.
SMS
The first reply in a text conversation carries an AI disclosure line, repeated periodically for long-running threads.
Each disclosure is logged with language and recipient conversation.
Operator dashboard
Every AI message is badged, and each conversation carries a persistent AI indicator alongside the model that produced each reply.
Indicators are rendered from message provenance, with screen-reader accessible labels.
The Marking Specification
Article 50(2) requires machine-readable marking of AI-generated content. This is our published specification: it is deliberately simple enough that anyone can verify a file with tools they already have. AI-generated content produced by the platform carries the literal text AI-generated in the following locations.
Word, PowerPoint, Excel documents
The category and keywords document properties, embedded in the file itself, travel with download and forwarding.
PDF documents
The Keywords entry in the PDF metadata, carried over automatically when a marked document is converted to PDF.
An X-AI-Generated message header set to true.
Markdown and text files
A machine-detectable marker comment on the first line.
Stored files
A provenance flag in the document store for programmatic queries.
Verify It Yourself
- check_circle PDF: open the file in any PDF reader and view document properties. The Keywords field reads AI-generated.
- check_circle Windows: right-click the file, choose Properties, then Details. The Tags field shows the mark.
- check_circle Office files: the mark appears under document properties as the category.
A ready-made verification tool covering every mark above is available free of charge on request, as is the full compliance documentation.
Honest Limitations
Plain text itself can only be watermarked inside the AI model, at generation time. We deliberately add no hidden characters to text, because such tricks are fragile and break accessibility; instead, every channel our text is delivered through carries its own marking or disclosure, and we adopt each model vendor’s in-text watermarking and detection as they become available for the models we run, tracked per model with dates. Metadata marking is lost when a file is rebuilt: re-saving through another editor, printing to PDF, screenshots, or copy-pasting text. The absence of a mark is therefore never proof that content is human-made, and content from other AI systems is marked by their vendors, not by us. Only genuinely AI-generated content is marked: documents an agent receives, redlines on a human’s document, or forwarded third-party content are never labeled AI-generated, so a mark you do find can be trusted.
Human Oversight with a Named Reviewer
Consequential actions require explicit human approval before execution, and the approval record captures the verified identity of the person who approved: who, when, and through which channel. For content published to inform the public, a per-pipeline review workflow routes publishing through that same named-reviewer approval.
Out of High-Risk Scope, by Design
The AI Act reserves its heaviest obligations for high-risk uses listed in Annex III. The platform and its agent templates are deliberately kept outside those categories: no agent screens or evaluates job candidates, scores creditworthiness, decides access to education, or assesses eligibility for essential services, and configurations attempting those purposes are blocked at validation time. Each agent template carries a documented intended purpose and classification, maintained as part of the platform, and the exclusions are part of the terms of service.
Governance You Can Ask For
Agent inventory
A machine-readable inventory of every deployed agent: intended purpose, classification, model, channels, and disclosure state, generated from the platform rather than maintained by hand.
Audit evidence
Disclosure events, marking, approvals with reviewer identity, and full interaction logs, queryable per environment.
Learning without training
Agents improve through a reviewed memory layer. AI model interactions are inference-only: customer data never trains or fine-tunes any model.
Framework and Status
Our marking and detection approach follows the EU Code of Practice on Transparency of AI-generated Content as its reference framework. We do not claim any certification, because none exists for this category of system: what we offer instead is implementation you can verify, documentation you can request, and honest statements about limitations.
Compliance questions before procurement?
Ask for the compliance documentation and we will walk your team through the implementation, channel by channel.
Contact Us