Responsible AI at Roark

Last updated: July 2, 2026

Roark builds testing, simulation, and analytics tools that help teams understand and trust their voice AI. Because our job is to make AI legible — to show teams how their systems actually behave — responsible AI is core to the product, not a bolt-on. Roark is a measurement and quality-assurance layer: it evaluates and monitors the voice AI our customers operate, and it does not make or recommend consequential decisions about individuals.

We hold ourselves to the following principles.

Transparency

Every score, metric, and evaluation Roark produces is traceable to the evidence and the logic that generated it — the underlying conversation, the transcript segment, and the rule or evaluator that produced the result. No black boxes. A reviewer can always see why a result was reached.

Human oversight

Roark is a decision-support and quality-assurance tool. It surfaces evidence for people to act on; it does not take autonomous, consequential actions about individuals. A human always remains in the loop for any downstream decision.

Privacy and data protection

We never use customer data to train, refine, or enhance our models — a commitment we make contractually, not just as a policy. We act as a data processor on our customers' documented instructions, store and process data onshore within the United States, and delete customer content on request. See our Cloud Service Agreement, Data Processing Agreement, and Privacy Policy for details.

Fairness in measurement

Reliable measurement must be even-handed. We assess and monitor our evaluation and transcription components for disparities across accents, dialects, and demographics, because measurement that varies by speaker would undermine the value we provide. Where customers use Roark to test their own candidate- or user-facing systems, Roark can form part of their own fairness and quality-assurance process.

Accountability

We maintain documented governance, change management, and incident-response processes, and we allocate liability clearly in our contracts rather than disclaiming it onto our customers. We commit to notifying affected customers of any security incident within 24 hours.

Reliability and bounded behaviour

We document what our system does, where it works well, and its limitations, and we monitor real-world performance. Evaluation logic is versioned so that results are reproducible and auditable over time.

Certification

We are pursuing ISO/IEC 42001 (AI Management System) certification, with a target completion of September 2026. In the interim, we operate the security, governance, and data-protection controls described in our Data Processing Agreement and Cloud Service Agreement.

Questions about this document? Email support@roark.ai.