AI DATA HANDLING

How Atlas uses AI with your portfolio

Atlas uses AI to help you analyse your IT portfolio faster. This page describes exactly which AI features exist, what data they send to AI providers, and what controls you have as a customer.

What we use AI for

The following AI features are available in Atlas. Each shows which model is used and the type of data sent.

Strategy parser	Claude Opus — parses uploaded strategy documents (PDF, DOCX, PPTX) into structured themes, objectives, and initiatives. Sends document text only.
Capability generation	Claude Opus — generates L1/L2/L3 capability maps from workspace industry + org metadata. No application data sent.
Application enrichment	Claude Sonnet + web search — enriches application records with lifecycle, vendor, and technology context from public sources. Sends application name and vendor.
Recommendation engine	Claude Sonnet — surfaces alignment gaps, redundancy candidates, and technology risks across your portfolio. Sends application attributes and capability assignments.
Policy generation	Claude Opus — generates AI governance policy drafts for EU AI Act compliance cockpit. Sends workspace metadata and risk classification inputs.
Walkthrough narratives	Claude Sonnet — generates executive summary narratives for the strategy-to-execution walkthrough. Sends portfolio summary statistics.
AI Help Assistant	Claude Haiku — answers in-product help questions. Sends the user's question text and current page context only. No portfolio data sent.

What data flows to AI providers

Workspace metadata: organisation name, industry, approximate team size.
Application attributes: application name, vendor, technology category, lifecycle stage, TIME quadrant assignment.
Capability structure: capability names and hierarchy (L1/L2/L3). Not individual employee assignments.
Strategy inputs: themes, objectives, and initiative names from strategy documents you upload.
Decision context: title and description of ADR drafts when AI drafting is triggered.
Uploaded document text: full text of PDF/DOCX/PPTX files when you use the strategy parser.

What is never sent to AI providers

User email addresses, names, or passwords.
Authentication tokens, session cookies, or API keys.
Audit logs or billing data.
Personally identifiable information (PII) from your users — Atlas is a portfolio management tool, not a CRM.
Data inside the applications you are tracking (i.e. data your customers' customers have in your apps).
Error context sent to Sentry (scrubbed via beforeSend hook before leaving the browser).

Data retention at AI providers

Anthropic	Zero data retention configured. Prompts and completions are not persisted by Anthropic beyond the live request. Opt-out of model training enforced via API agreement.
Azure OpenAI	Zero data retention default (no prompt logging). Opt-out of abuse monitoring available for EU-only customers under enterprise agreements.
Atlas prompt registry	Prompt templates and AI output metadata are stored in Atlas's own database (Neon, Frankfurt) for audit purposes. Retained for 12 months.
Atlas AI classifications	AI-generated enrichment results (classifications, recommendations) are stored per-workspace in Atlas's database. Retained until workspace deletion or 3-year limit.

Compliance and GDPR alignment

SCCs	Standard Contractual Clauses in place for cross-border transfers to Anthropic (US).
DPA with Anthropic	Signed. Available for customer review on request.
DPA with Azure	Signed for EU-only customers. Available on request.
PII at observability layer	Sentry beforeSend hook scrubs known PII patterns before error context leaves the client.
Right to erasure	Workspace deletion cascades to all AI outputs, classifications, and prompt history stored in Atlas. Implemented via FK CASCADE (ADR-0003).
EU AI Act alignment	Atlas maintains an AI registry of all AI features with risk classification (Annex III). Compliance cockpit available in-product.
What we do not claim	We do not claim AI Act certification (no such certification exists for AI providers in 2026). We claim GDPR-aligned operations and EU AI Act-ready documentation.

Customer controls

BYOL — Bring Your Own LLM Key

Supply your own Anthropic or Azure OpenAI API key per workspace. Billing flows directly to your own account. Configured in Settings → AI.

Banned model list

Per-workspace model blocklist (ADR-0007). Prevent specific models from being invoked — useful when procurement requires model approval.

Daily cost caps

Per-workspace daily spend limit on AI calls. Configurable in Settings → AI. Default cap applies to all workspaces to prevent runaway spend.

AI call audit log

Every AI call is logged in ai_classifications with timestamp, model, prompt version, and confidence score. Admins can export logs from Settings → Trust.

Data residency policy

Tri-state setting (eu_only / us_ok / on_prem) that controls which AI providers can be invoked. See the Data Residency page for details.

AI features toggle

Workspace admins can disable all AI features for their workspace. No AI calls will be made while the toggle is off.