Skip to main content
Draft pending counsel review. This document addresses regulatory disclosure requirements (including EU AI Act Article 50, which becomes applicable on 2 August 2026) and is being finalised by counsel.
Effective from: (pending) Last updated: 27 May 2026 This Policy describes the AI features in Brand Atlas, the underlying AI providers, what content is sent to those providers, and the disclosure obligations we satisfy.

​
1. AI features in Brand Atlas

Brand Atlas offers two named AI features:

​
Oswald (Guardian tier)

An editorial AI assistant. Used by the brand owner to draft Horizons, propose Update Request responses, and review assets against the brand record. Oswald is vendor-hosted; usage is included in the Guardian subscription.

​
Henry (Keeper and Guardian tiers)

A read-only AI assistant. Used by team members to answer questions about the brand record. Henry uses a customer-supplied AI provider API key (BYOK). Together they are referred to as “Brand Atlas AI.”

​
2. Underlying AI providers

The user-facing branding (Oswald, Henry, Brand Atlas AI) is product naming. The underlying AI models are operated by third parties:
  • Oswald. Powered by Anthropic Claude, on a vendor-hosted enterprise API tier.
  • Henry. Powered by the customer’s chosen provider: OpenAI or Google Gemini, on the provider’s enterprise API tier.
The underlying provider may change over time as we improve the product. Material changes will be announced and this Policy will be updated.

​
3. What content is sent to providers

​
Oswald

  • The team member’s prompt (the request being responded to) or the brand owner’s prompt (when used directly).
  • Relevant brand-record context: the standard sections, Horizons, and history that bear on the prompt.
  • The current Oswald session’s conversation history.
  • Brand Atlas’s own system prompt (which is not customer content but is part of the request).

​
Henry

  • The team member’s prompt.
  • Relevant brand-record context, scoped to what the user can read.
  • The current Henry session’s conversation history.
  • Brand Atlas’s system prompt for Henry.

​
4. What is not sent to providers

  • Material from other atlases. Each atlas is isolated.
  • Personal data beyond what is in the brand record. No team-member personal data or billing data is sent.
  • The plaintext form of any secrets (the BYOK key itself is not sent in prompts).
  • External system data (no email, calendar, file system access).

​
5. Training disclosure

Brand Atlas does not use Customer Content to train any AI model. The AI providers we use do not train their models on API content when accessed through the enterprise tiers we use:
  • Anthropic (Oswald): Anthropic’s enterprise API does not use API content to train Anthropic models. Anthropic’s commitment is documented in their API terms.
  • OpenAI (Henry): OpenAI’s API content is not used to train OpenAI models by default. The customer’s account controls this setting; if the customer’s account has data-sharing turned on for ChatGPT or model improvement, that setting may apply to API content. We recommend customers verify this in their OpenAI account.
  • Gemini (Henry): Google’s API content is not used to train Gemini models on enterprise tiers. The customer’s Google Cloud project’s settings govern this; we recommend verification in the customer’s account.

​
6. Retention by providers

  • Anthropic. API content is retained for up to 30 days for operational debugging, then deleted. Anthropic personnel do not access content except under exceptional circumstances (security investigation, legal requirement).
  • OpenAI. API content is retained for up to 30 days, then deleted. Customers on zero-retention agreements have shorter retention.
  • Gemini. Standard latency-purposes retention; customer-account settings can adjust.

​
7. Transparency obligations

​
Where AI is used

Brand Atlas surfaces AI use to users:
  • Oswald is clearly identified as an AI in the portal interface.
  • Henry is clearly identified as an AI in the chat interface.
  • Every Oswald-drafted change in the brand record carries an attribution note indicating Oswald drafted it.
This satisfies the principle that users should know when they are interacting with AI.

​
EU AI Act compliance

The EU AI Act’s transparency obligations under Article 50 become applicable on 2 August 2026 (per the European Commission’s confirmation that “the rules on transparency […] become applicable” on that date). Brand Atlas’s AI features are designed to satisfy these obligations:
  • Generative AI outputs are labelled as machine-generated where Article 50 requires.
  • Users are informed they are interacting with AI when chatting with Henry or Oswald.
  • Synthetic content produced by the AI is identifiable.
The full compliance position will be confirmed by counsel before the August 2026 applicability date.

​
8. Bias, fairness, and limitations

AI models exhibit biases inherited from their training data. We test and observe Oswald’s outputs for the brand-management use cases Brand Atlas is designed for; we cannot eliminate all biases. We recommend brand owners and team members treat AI outputs as drafts to review, not as final decisions. The deliberate design of Oswald (drafts require brand-owner approval) and Henry (read-only) reflects this principle.

​
9. Customer rights regarding AI

A customer can:
  • Disable Oswald or Henry from the AI settings.
  • Remove the BYOK key, ending Henry usage.
  • Object to specific AI features through legal@brandatlas.pro if a feature is incompatible with their compliance position.
  • Receive a report of AI usage in their atlas on request.

​
10. Changes

We may update this Policy. Material changes are announced 30 days in advance.

​
What changed

  • 27 May 2026: Initial draft published for counsel review.