Atlas · AI
What the AI in Atlas actually does.
The AI in Atlas keeps your architecture model current and serves it as governed context to your own AI tools. Agents draft entries, enrich them from real signals, and scan for duplicates. People approve every change in a review inbox. Nothing is written to the model without a human saying yes.
AI agents don't need more access. They need one governed truth to ask.
Notion made your documents agent-readable. Atlas makes your architecture agent-readable.
In plain terms, Atlas is the governed context layer: one place your agents can ask, and one place people stay in control of what changes.
01The work
What the agents do.
The agents take on the upkeep that used to eat an EA team's week. They never act alone.
Draft
From strategy documents, system lists and sources, agents draft entities and relationships, so the model starts from real material instead of a blank grid.
Enrich
Agents fill in vendor, category, lifecycle and ownership from signals and web search, and mark how sure they are.
Scan for duplicates
Agents flag likely duplicates and near-matches, so the same system does not live under three names.
02The principle
Write with review, always.
Reads come back with citations. Writes become proposals waiting for a person. A human approves or corrects in a review inbox, and every change is logged with who, when and why.
Agents write proposals. People approve. Everything is logged.
03The engine
Which models, and your own key.
Atlas uses Anthropic's Claude models for specific analyses, matched to the task: deeper models for strategy parsing, lighter ones for classification and help.
Models per task
Strategy parsing and capability generation use the deepest model. Enrichment and recommendations use a balanced model. The in-product help assistant uses the fastest, with only the question and page context.
Bring your own key
Supply your own Anthropic or Azure OpenAI key per workspace, and billing flows to your own account. A per-workspace model blocklist and daily cost caps are there when procurement needs them.
Controls
A data-residency setting controls which providers can be invoked, and workspace admins can switch all AI features off. While the toggle is off, no AI calls are made.
04Your data
What your data is used for, and what it is not.
Only content from your own workspace is included in AI calls. Data from other workspaces or users is never part of a request from your account. Anthropic does not retain the content of API calls and does not use it to train models. All AI-generated output is stored in your EU database.
What is sent
- Workspace metadata: organisation name, industry, approximate team size.
- Application attributes: name, vendor, technology category, lifecycle stage, TIME assignment.
- Capability structure: capability names and hierarchy. Not individual employee assignments.
- Strategy inputs: themes, objectives and initiative names from documents you upload.
- Decision context: title and description of ADR drafts when AI drafting is triggered.
- Uploaded document text: the full text of files when you use the strategy parser.
What is never sent
- User email addresses, names or passwords.
- Authentication tokens, session cookies or API keys.
- Audit logs or billing data.
- Personal data about your users. Atlas is a portfolio tool, not a CRM.
- Data inside the applications you are tracking.
- Error context sent to Sentry, which is scrubbed before it leaves the browser.
Full detail lives in the Trust center. This page reuses the same promises, so the story is the same wherever you read it.
05Regulation
Ready for the EU AI Act, on the real timeline.
The evidence you need is a by-product of the work, because the work happened in the system.
- Annex III obligations move to 2 December 2027, so there is runway to do this calmly rather than in a panic.
- The relief regime for organisations under 750 employees makes the requirements more manageable for the mid-market.
- The AI registry and cockpit generate documentation, audit trails and board reports from data you already hold.
We do not claim AI Act certification. No such certification exists for AI providers. We claim GDPR-aligned operations and EU AI Act-ready documentation.
FAQ
AI questions, answered.
- Can an agent change our model on its own?
- No. Every write is a proposal a person approves, and every approval is logged. Write with review is the backbone, not a preference.
- Do you train on our data?
- No, and neither does Anthropic. The content of API calls is processed transiently and is not used to train models.
- Can we use our own model key?
- Yes. Bring your own Anthropic or Azure OpenAI key per workspace, with billing to your own account, a model blocklist and daily cost caps.
- Where is our data stored?
- AI-generated output is stored in your workspace database in the Frankfurt EU region, scoped to your workspace, and can be exported or deleted like any other Atlas data.
- Can we turn the AI off?
- Yes. Workspace admins can disable all AI features. While the toggle is off, no AI calls are made.
Give your agents one truth to ask.
Get early access, or talk to us about how governed context fits your AI plans.