Responsible AI Policy

The short version. Pazau's AI agents speak to millions of people on behalf of brands in regulated industries, so we hold them to a high bar. Every model and Procedure change is evaluated before release. We red-team continuously. Humans stay in control of consequential actions. Every AI decision leaves a trace your team can inspect and export. We test voice and language features for performance gaps across accents, dialects, and languages. We never train foundation models on Customer Data. Our AI management system is certified to ISO/IEC 42001.

Why this policy exists

Pazau Agent resolves 67% of customer conversations end-to-end on average: rebooking flights, filing insurance claims, disputing card charges, and changing delivery addresses. When an AI agent takes those actions, mistakes affect real people. This policy describes how Pazau designs, tests, deploys, and monitors its AI systems so those people are treated fairly, honestly, and safely.

It applies to all AI features in Pazau Agent, Pazau Voice, Pazau Inbox, and Pazau Insights, and to the third-party models they rely on. Contractual commitments to customers are in the AI Addendum. Customer obligations are in the Acceptable Use Policy.

Accountability. Dr. Mei Tanaka, Head of AI, owns this policy. Omar Haddad, CISO, owns AI security. Our AI Governance Council, chaired by the Head of AI with members from engineering, security, legal (Daniel Reyes, General Counsel), privacy (Aoife Brennan, Data Protection Officer), and customer success, meets monthly and approves every new model provider, high-impact feature, and policy exception.

Our principles

1. Resolve, don't deflect. An AI agent should solve the customer's problem or hand them to a person who can. Containment rate is not our goal. Resolution is.
2. Honest by default. Pazau agents say they are AI, don't pretend to be human, and don't invent policies, prices, or promises. When an agent doesn't know, it says so or escalates.
3. Humans in control. Customers decide what an agent may do. Consequential actions can require human approval, and any end customer can ask for a person.
4. Explainable and auditable. Every AI decision produces a trace: the Procedure followed, the data retrieved, the Actions called, the guardrails checked, and the reason for each step. Traces are exportable.
5. Private by design. Customer Data is used only to serve the customer it belongs to. We never train foundation models on it, and model providers operate under zero data retention.
6. Fair across people and languages. An agent should work as well for a caller with a Glaswegian accent or a Tagalog-speaking customer as it does for anyone else. We measure this and publish our thresholds.
7. Proportionate controls. The higher the stakes of a use case, the stronger the safeguards: stricter guardrails, more evaluation, and more human review.

How we build: the AI lifecycle

Our AI management system follows ISO/IEC 42001 and maps to the NIST AI Risk Management Framework (Govern, Map, Measure, Manage). Every AI feature passes through these gates:

Model providers

Pazau uses foundation models from Anthropic, OpenAI (through Azure OpenAI Service), and Google (through Vertex AI), plus specialized speech providers. We evaluate each model version on our own suites before it can serve customer traffic. We do not switch a customer's production model without running those evaluations, and Enterprise customers can pin a model version or restrict providers. The full list is on the Sub-processors page.

Evaluation

We maintain evaluation suites built from synthetic conversations and from de-identified conversations contributed by customers who opted in to our model improvement program. Customers who haven't opted in never have their data used in evaluation sets.

Blocking metrics for every release include:

• Resolution accuracy: whether the agent reached the correct outcome under the Procedure.
• Policy adherence: whether it followed the Procedure's steps, including required disclosures and verification.
• Groundedness: whether every factual claim is supported by retrieved knowledge or Action results. Our production target is fewer than 0.5% unsupported claims.
• Action safety: whether it called the right Action with the right parameters and requested approval where configured.
• Escalation quality: whether it handed off when it should, with a useful summary.
• Safety: harmful, toxic, or inappropriate output rates on adversarial sets.

Customers can run the same evaluations on their own Procedures in Pazau Agent → Test suites before publishing a change, and Pazau Insights scores 100% of live conversations for quality and compliance.

Red-teaming and AI security

Our AI security team red-teams every major release and every new model version. We also commission an external red-team exercise at least twice a year. We test for:

• direct and indirect prompt injection, including instructions hidden in emails, attachments, web pages, and Action responses;
• data exfiltration across conversations, customers, or workspaces;
• guardrail bypass and jailbreaks;
• social engineering of Actions, such as tricking an agent into issuing an unauthorized refund;
• voice-specific attacks, such as synthetic voice impersonation and audio injection.

Findings are tracked to closure with severity-based deadlines. We align our testing with the OWASP Top 10 for LLM Applications and MITRE ATLAS. External researchers can report AI vulnerabilities under our Vulnerability Disclosure Policy, and prompt injection that crosses a tenant boundary is in scope for our bug bounty.

Human oversight

Pazau gives customers layered controls:

• Procedures define exactly what an agent may do, in natural language your CX team can read and change.
• Actions have per-Action permissions, spending limits, and optional human approval. For example, refunds over $250 can require an approval in Pazau Inbox.
• Handoff on request: any end customer can ask for a human, and the agent must offer one. Customers configure when and how, but they can't turn it off entirely.
• Confidence-based escalation: agents escalate when they are uncertain, when a guardrail fires, or when sentiment signals frustration or vulnerability.
• Traces: supervisors can see why an agent did what it did, in real time or after the fact, and export traces for audit.
• Kill switch: workspace admins can pause any agent, channel, or Action instantly.

Bias and fairness testing for voice and language

Speech and language systems can perform worse for some people than others. We test for this before release and monitor it in production.

Speech recognition. We measure word error rate (WER) for Pazau Voice across accent and dialect groups, including regional US, UK, Irish, Indian, Nigerian, Filipino, and Latin American English, and across age bands and call conditions (mobile, landline, speakerphone, background noise). A release is blocked if any tested group's WER exceeds the best-performing group's by more than 1.5×, or if any group regresses by more than 10% relative.

Language coverage. Pazau Agent supports 45+ languages (30+ on voice). Eleven are Tier 1: English, Spanish, French, German, Portuguese, Italian, Dutch, Swedish, Japanese, Korean, and Arabic. Tier 1 languages must reach resolution accuracy within 5 percentage points of English on our benchmark. Other languages are labeled in the product with their measured accuracy, so customers can make an informed choice.

Outcome parity. Using synthetic personas that vary names, dialects, and writing styles while holding the request constant, we test whether agents reach the same outcome, such as a refund, rebooking, or escalation, at the same rate. Differences greater than 2 percentage points on consequential outcomes are investigated before release.

Voice synthesis. Pazau's stock AI voices are licensed from consenting voice actors. We don't create custom voices of real people without their documented consent.

We publish a summary of our fairness results each year in the Trust Center.

Transparency to end customers

• Pazau agents disclose that they are AI in their greeting by default, and always when asked.
• AI-generated voice calls include a spoken disclosure at the start of the call.
• Customers can show end customers a plain-language explanation of how the agent works and how to reach a person.
• We give customers the information they need to meet their own transparency obligations, including model documentation and this policy.

AI incident handling

An AI incident is an event in which an AI feature causes, or could cause, harm: for example, an agent giving materially wrong information, taking an unauthorized Action, exposing data, or behaving in a discriminatory way.

1. Detect. Signals include automated monitors, Pazau Insights compliance flags, customer reports, and researcher reports.
2. Contain. On-call engineers can roll back a model or feature, disable an Action type, or tighten guardrails platform-wide within minutes.
3. Notify. We notify affected customers without undue delay. If Customer Data is involved, we follow the breach notification timelines in the Data Processing Addendum: no later than 72 hours after becoming aware.
4. Investigate. Every Sev 1 and Sev 2 AI incident gets a blameless post-incident review, with root cause and corrective actions shared with affected customers.
5. Report. Where the EU AI Act or other law requires reporting a serious incident to authorities, we support customers and report as a provider where that obligation falls on us.

To report an AI incident, customers can contact support or email security@pazau.com.

EU AI Act readiness

We track the EU AI Act's phased application and have mapped our obligations:

• Prohibited practices (applicable since February 2, 2025): our Acceptable Use Policy prohibits every practice listed in Article 5, and we screen new use cases against it.
• AI literacy (since February 2, 2025): all Pazau employees complete annual AI literacy training, with role-based modules for engineering, sales, and support. We provide admin training for customers through the Pazau Academy.
• General-purpose AI models (since August 2, 2025): Pazau does not place general-purpose AI models on the market. We rely on providers that have signed the GPAI Code of Practice, and we review their technical documentation.
• Transparency (Article 50, from August 2, 2026): Pazau agents disclose AI interaction by default, and AI-generated audio is marked. Customers acting as deployers can meet their disclosure duties with built-in settings.
• High-risk systems: Pazau's standard support use cases are not high-risk under Annex III. We maintain a register of customer use cases that approach Annex III categories, such as essential services in banking and insurance, and review them with customers. We are preparing technical documentation, logging, and quality management artifacts in line with Articles 9 to 17 so we can support customers with high-risk deployments if and when they arise.

Our AI Act conformity work is overseen by the AI Governance Council and reviewed annually by external counsel.

Certifications and standards

• ISO/IEC 42001: certified AI management system covering the design, development, and operation of Pazau Agent, Voice, and Insights.
• ISO/IEC 27001 and 27701: information security and privacy management.
• SOC 2 Type II: including AI-relevant controls for change management and monitoring.
• NIST AI RMF: our control framework maps to its functions and the Generative AI Profile (NIST AI 600-1).

Reports and certificates are available under NDA in the Trust Center.

Questions and feedback

We welcome questions, challenges, and suggestions from customers, end customers, researchers, and civil society. Email responsible-ai@pazau.com. Privacy questions go to our Data Protection Officer at dpo@pazau.com.