Responsible AI Policy

FrootAI helps teams build AI agents they can trust. That starts with us being transparent about how our own platform works — what we do, what we don't do, and what we believe.

Last updated: May 2026 · Version 1.0

What FrootAI Does (and Doesn't Do)

We are an evaluation and quality assurance platform. We help teams measure whether their AI agents work correctly, safely, and reliably.

We do:

Score agent outputs for groundedness, safety, and quality
Provide deterministic, reproducible evaluation results
Integrate with CI/CD pipelines for automated testing
Surface regressions, anomalies, and failures
Help customers comply with AI Act evaluation requirements

We don't:

Train AI models
Make deployment decisions for customers
Use customer data to improve our own models
Censor or modify agent outputs
Guarantee that evaluated agents are safe (we measure; you decide)

Our Six Principles

Model Transparency

✓We proxy models — we don't train them. You choose which models power your agents.
✓For every model used in evaluation, we link to the provider's model card (capabilities, limitations, training data).
✓All evaluation scores are deterministic and reproducible — same input, same score, every time.
✓We document which eval primitives (groundedness, faithfulness, safety) use model-based scoring vs. rule-based scoring.
✓No black-box scoring — every metric has a published methodology.

Bias Mitigation

✓Our evaluation framework tests for bias explicitly: demographic fairness, language equity, cultural sensitivity.
✓We provide bias detection eval suites out of the box — customers can run them on their agents before production.
✓We acknowledge that model-based evaluation can inherit biases from the scoring model itself.
✓We publish benchmark results showing how different scoring models perform on bias-sensitive evaluation tasks.
✓We encourage customers to include diverse test cases in their evaluation suites and provide templates to do so.

Human Oversight

✓FrootAI augments human judgment — it does not replace it. Evaluation scores inform decisions; humans make them.
✓Every automated eval suite can include human review checkpoints (e.g., 'flag for human review if safety score < 0.8').
✓We do not auto-deploy or auto-approve agents based on eval scores — that decision stays with the customer's team.
✓Enterprise customers can configure mandatory human sign-off before any agent promotion to production.
✓Our dashboards surface anomalies for human attention, not just aggregate scores.

Data Handling

✓No customer data is used for model training — ever. Not by us, not by model providers through our platform.
✓Evaluation data (inputs, outputs, scores) is stored in the customer's tenant and encrypted at rest (AES-256) and in transit (TLS 1.3).
✓Customers control data retention — default 90 days, configurable from 7 days to 3 years.
✓GDPR data subject requests (export, deletion) are supported and documented.
✓On contract termination, all customer data is deleted within 30 days — with a certificate of deletion available on request.

Content Safety

✓We provide content safety evaluation primitives out of the box: toxicity, PII detection, jailbreak resistance, prompt injection detection.
✓Content safety guardrails are configurable per tenant — customers define their own thresholds based on their use case and risk tolerance.
✓We do not censor evaluation inputs or outputs — we score them. The decision to block or flag is the customer's.
✓Safety evaluation results are immutable in the audit log — they cannot be retroactively modified.
✓We publish our safety scoring methodology and update it as new attack vectors emerge.

Environmental Impact

✓We publish an annual estimate of our compute and energy footprint.
✓Our infrastructure runs on Azure regions that offer renewable energy options — we select green regions where available.
✓We optimize evaluation pipelines for efficiency: batch processing, caching, and deduplication reduce unnecessary compute.
✓We provide customers with per-evaluation cost and compute estimates so they can make informed decisions.
✓We commit to carbon-neutral operations by 2028 through Azure sustainability programs and offsets.

Governance

This policy is owned by the FrootAI Founder / CEO and reviewed annually. Material changes are communicated to customers via email and changelog.

Annual review: Every January, starting January 2027.
Trigger review: Major product changes, new regulations, or incidents.
Customer input: Enterprise customers can request a review via their CS Lead.
Public changelog: All policy versions archived at /responsible-ai/changelog.

Questions?

If you have questions about this policy or FrootAI's approach to responsible AI, contact us at [email protected].