AI transparency

This page explains, in plain terms, how AI works inside Ovidens: what the AI is, which models power it, what data reaches those models, and the controls that keep a person in charge. It is informational and describes the product as it works today; it is not legal advice.

What the AI is

Ovidens has one AI, named Vera. You meet it in three places, but it is the same assistant and the same controls behind all of them:

• The in-app assistant you chat with while building or querying an app.
• Skills: reusable AI tasks an app author saves so anyone can run them.
• The hosted connector, a per-workspace door that lets your own AI (for example Claude) work in your workspace on your behalf. See Connect your own AI.

Models and provider

Which model powers Vera is a per-workspace choice. A workspace admin selects it; every member can see the active model. These are the models in the catalog:

These three models are the available choices. Vera uses one configured model per workspace for all AI work: chat, AI widgets executing skills, background helpers and conversation summarisation. There is no separate helper model. The default is Claude Sonnet 5.5.

The API's default inference routing is global. Vera's AI processing is not limited to the EU, even though Ovidens' application is hosted in Frankfurt. See Anthropic's data-residency documentation. Anthropic does not train on commercial API inputs or outputs by default; see its commercial data policy.

Ovidens is the provider of this AI system and a deployer of general-purpose AI models. Ovidens does not train its own models; it uses the model providers' models as delivered through their APIs.

Choosing the workspace AI model

A workspace admin picks the model on the General tab of the System page (the workspace configuration page in the sidebar), under AI configuration: the "Workspace AI model" card lists every selectable model with a short description and its processing location, and shows which one is active. Members see the same card read-only. The choice applies to the whole workspace - every Vera chat, setup job, skill run and internal AI call uses the selected model. Each new request or resumed request resolves the workspace's current setting, including when you reopen an older conversation. Conversation history does not pin a model. The audit log records the model used for every AI action, so a change of model is visible in the workspace's history. See Administration for the activity log.

Intended use and limitations

Vera assists with building and querying analytics: setting up an app's data steps, indicators, and Visualisation, and answering questions about your data.

Its outputs are AI-generated and may be inaccurate. Treat them as a fast first draft, not a final answer. As the in-app notice puts it: Vera is an AI assistant, its responses are AI-generated and may be inaccurate, so please verify before acting on them. You remain responsible for what you publish or decide.

Human oversight and controls

Vera is built so a person stays in charge:

• Changes follow your request and permissions. Requested edits use the app's draft. Vera publishes only when you ask and its publication tool is enabled for you. AI widgets are read-only and never edit data or create a draft.
• Per-user AI permissions and opt-outs. An admin sets the maximum any member's AI may ever do, and each member can switch off anything they do not want their AI to do.
• External AI is deny-by-default. The hosted connector is off for every member until an admin grants it per person, per connection.
• Per-surface rate limits keep AI use bounded.
• Bounded skill runs. A skill run is capped at roughly 30 steps and about 150 seconds, so it always finishes in a predictable envelope.
• A tenant audit log records every AI action with per-resource AI provenance: which surface ran it, which model or client was used (the workspace's selected model, per action), and which person instructed the AI. See Administration for the activity log.

Data handling

Vera sends your request, relevant conversation context, attachments and tool results to Anthropic's API. Depending on the task, those results can include app configuration, knowledge-base content and queried data. The complete stored conversation is not sent on every request; the model receives a bounded recent portion automatically. Older turns remain available to you in conversation history; restate important details when they fall outside the model's recent context.

Every call uses the workspace's configured model through this direct API, whose processing is not limited to the EU. Ovidens checks your current permissions before every resource operation and keeps workspace conversations private to their owner. Anthropic's API data handling follows its commercial data policy and data-residency controls.

An external AI connected through MCP uses its own provider and terms. Review those before connecting it; it is separate from Vera's model selection.

Transparency markers

Two markers make AI use visible:

• The "you are interacting with AI" disclosure shown wherever Vera answers.
• The audit-log AI-provenance marker, which tags each AI-made change with the surface, the model or client, and the person who instructed it.

How this maps to the EU AI Act

For the AI features described here, Ovidens treats its obligations as transparency-risk obligations under Article 50 of the EU AI Act: people are told when they are interacting with an AI system, and AI-generated actions are logged and attributable.