AI governance you can show an auditor

The EU AI Act is an obligation with dates, not a white paper. Valystrum treats it the way it treats every regulation it tracks: articles mapped to controls, controls proven by evidence, and a readiness figure an auditor can check. This site covers that part of Valystrum, and how we run AI ourselves. The platform, its plans and everything else it does are on valystrum.com.

The EU AI Act, as an obligation and a framework

Applicability, by your role. The platform does not guess whether the Act binds you. The obligation asks whether your organisation provides or deploys AI systems, and it is suggested, never adopted for you. The framework's first domain records your role (provider, deployer, importer or distributor, product manufacturer, or several) and whether the high-risk, general-purpose AI and transparency duties apply; providers' and deployers' obligations each have a domain of their own.
The high-risk requirements as controls. Risk management, data governance, technical documentation, record-keeping, transparency to deployers, human oversight, and accuracy, robustness and cybersecurity (Articles 9 to 15), plus the quality management system (Article 17), each a control with its questions. Prohibited-practice screening (Article 5), the deployers' duties (Article 26), the transparency duties of Article 50, general-purpose AI models (Articles 53 to 55) and post-market monitoring (Articles 72 and 73) are mapped the same way. Article 4's AI-literacy duty is on the obligation's timeline but has no control of its own yet; until it does, ISO/IEC 42001's competence control is where you evidence it.
Evidence, not declarations. Each control carries its evidence. The obligation view shows your readiness per article in two numbers: what fresh, passing evidence proves right now, which leads, and what you have declared implemented.
The dates. The Act's phase-in dates are on the obligation's page, from the prohibited practices and AI literacy (February 2025) and the general-purpose AI model duties (August 2025) onward. The high-risk dates shown there predate the Digital Omnibus on AI, which is reported to move the Annex III obligations to December 2027 and Annex I's to August 2028. We are checking that against the Official Journal; until we have, treat the high-risk dates on that page as under review.

ISO/IEC 42001 and the NIST AI RMF

ISO/IEC 42001, the AI management system standard, and the NIST AI Risk Management Framework are frameworks on the platform beside the EU AI Act, so the standard you certify against and the regulation you answer to are worked in one place. No ready-made crosswalk links their controls to the EU AI Act's yet, or ISO/IEC 42001's to the NIST AI RMF's. On Silver and above you link them yourself, or import the pair and let your AI provider match them. AI matches are saved as soon as the import runs, labelled "AI match", and count toward what you have declared, never toward what evidence proves.

Bronze has one framework and two regulatory overlays. Adopting the EU AI Act obligation there takes one overlay for the Act and brings in ISO/IEC 27001 as well, because its controls evidence three of the Act's articles (10, 12 and 15); that is Bronze's one framework, so ISO/IEC 42001 or the NIST AI RMF beside the obligation needs Silver, which has five regulatory overlays and unlimited frameworks. Gold and Enterprise have unlimited of both.

AI output that says where it came from

When Valystrum's own AI features write something, it says so:

AI drafts and insights are labelled as AI-generated on screen.
An assessment or board-report PDF whose narrative a model wrote carries a machine-readable marker in its metadata (ai-generated, with the model or the provider that wrote it) and a sentence the reader sees.
In a data export, AI-written rows are flagged aiGenerated: true, with the gaps named below.

That is how we meet Article 50's marking duty for our own AI, with gaps still to close: a policy version's AI change summary and the questions the AI writes for a supplier questionnaire are neither labelled on screen nor flagged in the export, and posture recommendations, the control mappings the AI matches, AI remediation suggestions and the actions the AI proposes are not yet flagged in the export. Our register lists them. Two more are shown on screen without an AI-generated label: a threat brief's narrative and the AI Security OS's brief. The other way round, two kinds of row are flagged although no model wrote them: a what-if scenario you save from the scenario panel and an insight posted through the API.

An AI-drafted policy text or a supplier score waits for a person: it becomes part of your record only when someone in your organisation accepts it. Two outputs do not wait yet: control mappings the AI matches, on a framework-pair import or through the API, land at once, labelled "AI match", and you unmap the ones you disagree with; matching through the API can also turn a mapping you made by hand into an AI match; and a policy version's AI change summary, requested through the API, is saved as the model wrote it.

Your model provider, your key

Valystrum's AI features run on the provider and the key your organisation configures: Anthropic, Mistral (France, the EU choice among them), OpenAI or Perplexity under your own contract, or an OpenAI-compatible endpoint you run yourself. Without a key the AI features stay off and everything else keeps working; reports are then composed by the platform from the same figures and marked as computed. Before a prompt leaves the platform it is stripped of credentials and of the personal identifiers a pattern can recognise (email addresses, phone numbers written with a country code, IBANs). Names cannot be recognised that way, and we do not claim they are removed. The vectors behind search are computed on the platform itself, with no network call, unless your organisation sets its own embeddings endpoint; text sent there is redacted the same way first.

Planned, and not available yet: Valystrum EU AI, Mistral's models (France) and OpenAI's open-weight gpt-oss through Scaleway's Generative APIs in Paris, for organisations that want no model-provider contract of their own. It will be metered in credits: a monthly allowance on every plan, a trial included, and prepaid packs beyond it. A call that the credits left cannot cover will be refused before anything is sent, until the allowance resets or a pack is added. Its prompts will be redacted and its outputs labelled exactly as with your own provider. The vectors behind search could be computed there too, on an embedding model of Alibaba's Qwen family (China) whose open weights Scaleway runs in Paris; it turns text into vectors and writes no text. Bring your own key stays beside it.

Your AI assistant, connected read-only

Your people can connect an AI assistant that speaks the Model Context Protocol (MCP) to Valystrum and ask it about your organisation's compliance, findings and deadlines.

Off until you switch it on. An owner or admin turns AI assistants on for the organisation, for every member or for owners and admins only, and can allow only recognised assistants (claude.ai, claude.com, chatgpt.com, an app on the person's own computer, or your organisation's own confidential clients). Every owner and admin is emailed each time someone connects one.
Read-only. An assistant reads what a viewer can see, only in the modules your admins allow, and changes nothing.
What leaves, and to whom. What an assistant reads goes to that assistant's provider under each person's own account with them. The redaction described above applies to Valystrum's own AI features, not to an assistant's answers. An incident and the change timeline are read without the people in them, and a workflow's approvals without who was asked or who decided; the remediation items they lead to are not, so an incident's action items still show who reviewed the incident and who owns each action, and an item an approval creates shows who decided. Other records are read as you would see them, including who owns a risk or a control and who is assigned a task, with their names and, in some records, their email addresses (a risk's owner, a remediation item's assignee, a supplier questionnaire's recipient). A setting that leaves people's names out of answers by default, and never sends an email address, is decided and not built yet.
What it never reads: members and invitations, evidence files and every download or export, API keys, secrets and integrations, settings, audit logs, comments, and anything that calls an AI provider.
It acts as a person. Each connection is approved by someone signed in at app.valystrum.io, for one organisation, and ends after 90 days at most. Every call is recorded under that person's name before it is answered.

How we run AI ourselves

We keep the records we would ask you for. Valystrum B.V. is the provider of Valystrum AI; when your organisation uses its AI features, you are their deployer, and we are a deployer too when we use it on our own organisation. Our records:

An AI system register: the system, our role, the model chain, the classification (not high-risk under Annex III, not a prohibited practice, Article 50 transparency applies), human oversight, logging and known limitations, reviewed yearly, whenever a feature changes what the system infers about a person and whenever the model chain changes.
An AI-literacy record under Article 4: the curriculum, and a row for each person who builds, supports or demonstrates the AI features.
A data protection impact assessment for the AI features.

We train no language model. We train one model ourselves: the breach-likelihood model that scores assets (never people) from derived numbers per asset, with no names, hosts, addresses or finding text. It learns across customers only from organisations whose owner turns on "Help train the breach model" in Settings, which is off by default. Every organisation is scored either way, and each score shows the factors behind it. Our register records this model beside Valystrum AI.

The classification is summarised on our own obligations page, and the register itself is published there with our Trust Center when Valystrum goes live on Scaleway. Your DPO can have the classification record on request.

Coming: a register for your own AI systems

A register of your own AI systems, counted in the EU AI Act view, is coming. Today the asset inventory does not record which assets are AI systems, so the Act's scope is the answer you give.