AI Act Transparency Report
How FrootAI classifies under the EU AI Act, our obligations as a GPAI provider, and how we help customers meet their own compliance requirements.
Last updated: May 2026 · Version 1.0 · Regulation (EU) 2024/1689
Our Classification
FrootAI is classified as a General-Purpose AI (GPAI) provider under Article 53.
We provide software components (evaluation protocol + platform) that can be integrated into a wide range of AI systems across industries and use cases. We proxy general-purpose models for scoring but do not train them.
What we are
- • GPAI provider (evaluation tooling)
- • Downstream provider of model-based scoring
- • Data processor (under GDPR)
What we are not
- • Not a foundation model trainer
- • Not a GPAI with systemic risk
- • Not a high-risk AI system (we evaluate them)
Customer Responsibilities
Customers are responsible for the AI systems they build and deploy. FrootAI helps you evaluate — but the deployment decision is yours.
- Model selection: You choose which models power your agents. You inherit that provider's obligations.
- High-risk classification: If your agent operates in a high-risk domain (Annex III), you must comply with Articles 6–15.
- Risk management: FrootAI provides evaluation data — you must implement the risk management process (Article 9).
- Human oversight: You must ensure appropriate human oversight of your AI systems (Article 14).
- Documentation: You must maintain technical documentation. FrootAI eval reports can serve as evidence.
- Incident reporting: If your AI system causes harm, you must report it to authorities. FrootAI audit logs can support the investigation.
Obligation Mapping
Article 53 — Transparency obligations for GPAI providers
Compliant- Technical documentation published (FAI Protocol specification, architecture docs)
- Model cards linked for all scoring models (see table below)
- Copyright policy: we do not train models — we link to provider copyright disclosures
- EU AI Office notification: will file when enforcement guidance is finalized
Article 55 — Obligations for GPAI with systemic risk
Not applicable- FrootAI is not a GPAI model with systemic risk — we are an evaluation platform that proxies models
- We do not train foundation models
- Systemic risk classification applies to model providers (OpenAI, Anthropic, etc.), not to evaluation tooling
Article 6 & Annex III — High-risk AI classification
Not applicable (platform), advisory for customers- FrootAI itself is not a high-risk AI system — it is a testing/evaluation tool
- However, customer agents evaluated by FrootAI MAY be high-risk (e.g., healthcare, insurance, employment)
- We provide guidance to customers on high-risk classification and how FrootAI helps meet Article 9 (risk management) requirements
Article 9 — Risk management for high-risk AI
Enabling tool- FrootAI's evaluation framework directly supports Article 9 compliance
- Systematic evaluation = documented risk identification and mitigation
- Eval results serve as evidence of risk management processes
- Customers can export evaluation reports for regulatory filing
Article 13 — Transparency for high-risk AI
Enabling tool- FrootAI dashboards provide the transparency reports that Article 13 requires
- Evaluation scores, failure analysis, and trend data = documented AI system behavior
- Customers can generate compliance-ready reports from Studio
Article 14 — Human oversight for high-risk AI
Enabling tool- FrootAI supports human-in-the-loop evaluation workflows
- Configurable thresholds for human review (e.g., safety score < 0.8 → flag for human)
- Mandatory sign-off gates before agent promotion to production
Model Cards
FrootAI proxies the following models for evaluation scoring. We link to each provider's model card for transparency.
| Provider | Model | Usage in FrootAI | Model Card |
|---|---|---|---|
| Azure OpenAI | GPT-4o | Default scoring model for groundedness, faithfulness, relevance evals | View |
| Azure OpenAI | GPT-4o-mini | Cost-optimized scoring for high-volume eval suites | View |
| Anthropic | Claude 3.5 Sonnet | Alternative scoring model (customer-selectable) | View |
| Gemini 1.5 Pro | Alternative scoring model (customer-selectable) | View | |
| Meta | Llama 3.1 70B | Self-hosted scoring option for data-sovereign deployments | View |
How FrootAI Helps You Comply
Risk Management (Art. 9)
Systematic evaluation identifies and quantifies risks in your AI agents before they reach production.
Data Governance (Art. 10)
Evaluation suites test for data quality, bias, and completeness — ensuring your training/retrieval data meets standards.
Technical Documentation (Art. 11)
Export evaluation reports, score histories, and configuration as compliance documentation.
Transparency (Art. 13)
Dashboards and reports provide the transparency that regulators and customers require.
Human Oversight (Art. 14)
Human-in-the-loop checkpoints, review gates, and mandatory sign-off workflows.
Accuracy & Robustness (Art. 15)
Evaluation primitives test for accuracy (groundedness), robustness (adversarial inputs), and reliability.
Copyright & Training Data
FrootAI does not train AI models. We use third-party models for evaluation scoring. Copyright obligations related to training data rest with the model providers:
- Azure OpenAI / OpenAI: openai.com/policies
- Anthropic: anthropic.com/policies
- Google DeepMind: policies.google.com/terms
- Meta (Llama): Llama 3 Community License
The FAI Protocol specification and FrootAI platform source code are licensed under MIT. No third-party copyrighted material is included in our codebase without proper attribution.
Questions?
For questions about FrootAI's AI Act compliance posture, contact [email protected].
This report will be updated as EU AI Office enforcement guidance evolves. Subscribe to our compliance changelog for updates.