BETAPro is not billed during beta. Lock in the price and we will honor it at launch.Freeze this price
Draft — this policy has not been approved for production use.

AI Policy

Starter — review with counsel before public launch. This is the operational shape of BrandBanta's AI transparency disclosure required under EU AI Act Art. 50 (limited-risk transparency obligation, applicable from August 2025). The technical facts are accurate to how the system works today; legal opinion is required on jurisdiction-specific wording before publication.

Effective date: to be filled at publication Last updated: to be filled at publication

What BrandBanta is

BrandBanta is a deployer of third-party AI assistants. We use external large language models (LLMs) from OpenAI, Anthropic, Google, and Perplexity to observe how those AI assistants describe customer brands — we are a measurement tool sitting in front of the AI ecosystem, not an AI provider ourselves.

The product surfaces the AI assistants' outputs, attribution analysis, and aggregated visibility metrics. We do not generate creative content, take autonomous actions on a user's behalf, or make decisions that affect third parties.

Risk classification under the EU AI Act

We classify BrandBanta as a limited-risk AI system under Regulation (EU) 2024/1689. We are not:

  • Engaged in any prohibited practice listed in Art. 5 (social scoring, manipulative subliminal techniques, real-time biometric identification in public spaces, etc.)
  • A high-risk system under Annex III (BrandBanta does not make decisions about employment, credit, justice, education, migration, or critical infrastructure)
  • A general-purpose AI model provider (we use foundation models supplied by Anthropic, OpenAI, Google, and the OpenRouter aggregator)

Our obligations are therefore concentrated in Art. 50 (transparency to natural persons) and the general AI literacy duty under Art. 4.

AI assistants we use

| Provider | Model family | Used for | | ---------- | -------------------------------- | ---------------------------------------------------------------------------------------- | | OpenAI | GPT-4o, GPT-4o-mini | Scan execution (routed via OpenRouter); embeddings (direct, for mention-context vectors) | | Anthropic | Claude Sonnet, Claude Haiku | Scan execution (via OpenRouter); insight narratives (direct, via Batches API) | | Google | Gemini 2.5 Flash, Gemini 2.5 Pro | Scan execution (via OpenRouter) | | Perplexity | Sonar | Scan execution with web search (via OpenRouter) |

Customers may bring their own provider credentials (BYOK) via Settings → API keys; when set, the corresponding workload bills to the customer's account at the upstream provider and that provider becomes the customer's direct data processor for that workload.

What BrandBanta does with AI outputs

  1. Stores raw text. Every AI response to a tracked query is persisted verbatim so customers can audit what the model said. Retention is governed by the Privacy Policy.
  2. Extracts structured mentions. A separate LLM call validates regex-detected brand mentions and assigns sentiment, position, and context-type metadata. The original text is the source of truth; extracted metadata is best-effort enrichment.
  3. Aggregates into visibility metrics. Mention rate, share-of-voice, citation share, and content-gap signals are computed deterministically from the stored mentions. Methodology is documented at /methodology.
  4. Generates summary narratives. A weekly digest and per-topic insight narratives are generated by Anthropic Claude via the Batches API. These narratives are summaries of the customer's own scan data — we do not train models on customer data, and Anthropic's API does not use Batches inputs for training under their published policy.

How AI is involved in decisions

BrandBanta does not make automated decisions about people. Every insight, alert, and metric is a summary of historical observations meant to inform a human's brand strategy. No automated content publishing, no autonomous outreach, no automated communication with third parties.

When the product surfaces a recommendation (e.g., "competitor X is gaining citation share — consider commissioning content on Y"), the recommendation is advisory. The user remains the decision-maker.

Limitations of AI outputs

AI assistants can:

  • Hallucinate facts. A response may state a brand attribute that is not actually true. We surface the raw AI response so customers can verify.
  • Be biased. Foundation models reflect the biases of their training data. BrandBanta does not correct for these biases; the measurement layer is honest about what the AI said, not about what the AI should have said.
  • Drift over time. Provider model updates change behavior. We pin model versions where possible and document drift in the methodology page.
  • Differ by provider. ChatGPT, Claude, Gemini, and Perplexity often give meaningfully different answers to the same query. BrandBanta measures this differential rather than averaging it away.

Customers should not treat AI-generated outputs as fact without independent verification, especially for adversarial use cases (competitive intelligence, regulatory disclosures, financial decisions).

Training and improvement

We do not train models on customer data. Our scan pipeline calls upstream provider APIs read-only. Provider-side data use is governed by each provider's terms:

  • OpenAI: API data is not used to train OpenAI models by default (per OpenAI Enterprise/API terms).
  • Anthropic: API and Batches API inputs are not used to train Anthropic models (per Anthropic API terms).
  • Google: API data handling is governed by Google's Generative AI terms.
  • Perplexity: API queries are subject to Perplexity's commercial API terms.
  • OpenRouter: Routes requests to the above; OpenRouter itself does not train on routed traffic per their published policy.

If a provider's training-data policy changes, we will update this page and notify customers materially affected.

Human oversight

The BrandBanta operator (today: a single-founder team) reviews:

  • New AI integrations before deployment
  • Material changes to extractor logic or summary-narrative prompts
  • Customer-reported issues that suggest systemic AI bias or hallucination
  • Quarterly review of provider terms-of-service changes that might affect customer data use

We commit to documenting AI literacy training for operating personnel as required by EU AI Act Art. 4 (applicable from February 2025).

How to flag an issue

If you believe BrandBanta has surfaced an inaccurate or harmful AI-generated output:

  • Email: ai-feedback@brandbanta.com
  • In-product: the report-a-problem link on any scan results page

We commit to acknowledge within 5 business days and to investigate genuine concerns within 30 days.

Changes to this policy

Material changes will be announced in-product and via email at least 30 days before they take effect. Non-material edits (typos, link updates, model-version refreshes) may happen without notice.

Contact

  • AI feedback: ai-feedback@brandbanta.com
  • Privacy questions: privacy@brandbanta.com
  • Security disclosures: security@brandbanta.com