At Avahi, we’ve always believed that Al is more than just a breakthrough in automation or insight, it’s a responsibility. As we look ahead in 2025, the global conversation has matured: we are no longer simply asking, “What can we build?” but rather, “What should we build, and how should it behave?”
With the repeal of the White House’s 2023 Executive Order and the introduction of comprehensive frameworks like the EU Al Act and UNESCO’s Al Ethics guidelines, the regulatory landscape has shifted. Al now operates under sharper scrutiny, and with that comes a greater demand for trust, safety, and fairness.
At Avahi, we don’t just adapt to these changes, we help shape them. We’ve deepened our Responsible Al practices to meet the realities of today’s governance, embedding ethical design principles directly into our workflows, tools, and infrastructure. This blog outlines how we’re extending our foundational pillars with real-world examples, client use cases, and evolved safeguards that reflect where the Al ecosystem is headed next.
Our Pillars of Responsible Al
We ground our Al product and system we build in the following eight principles:
1. Fairness and Safety
Mitigating bias and ensuring equitable outcomes across all Al interactions.
Our Approach:
- Conduct internal bias audits and impact assessments before project handover to clients.
- Partner with stakeholders to define fairness metrics tailored to real-world context, not just technical benchmarks.
- Actively use AWS Bedrock guardrails and responsible prompts to constrain biased outputs.
Example: Health Coach Al Prompt
You are a helpful, empathetic health coach. Generate a supportive, and detailed response by using the conversation history between a patient and their health coach, plus the patient’s JSON data. – Use ONLY the data from ‘patient_json’ and the conversation to craft an accurate, empathetic response. – Do NOT invent or guess details not provided. – Your response should be from the perspective of a health coach offering practical advice.
“”
Key safeguards in this prompt include:
- Empathy without assumptions: Explicitly avoids stereotyping by relying only on given data.
- Hallucination prevention: Prohibits speculation or fabrication of unverified details.
- Context-aware fairness: Metrics are co-defined with stakeholders for real-world relevance..
2. Transparency and Explainability
Building Al systems that users and regulators can verify and understand.
Our Standards:
- Deliver answers grounded in verifiable sources (e.g., institutional policies, cited documents, medical documents).
- Provide direct quotations when users request evidence (see screenshot example).
- Clearly communicate limitations, especially if they are mentioned explicitly in documents.
Screenshot Example: University Funding Query
Queen’s University 7 Graduate Studies & Postdoctoral Affairs
| Change | What happens to funding? | Considerations |
|---|---|---|
| Change from full-time to part-time | Queen’s does not guarantee funding for part-time students as it is understood part-time students are working full-time. If a students moves to part-time status mid-way through an a academic year once a funding package has been assigned to them, thestudents may be able to complete any remaining work assignments and receive any remaining awards affiliated with their existing funding package; however, future funding packages will be impacted by their change to part-time status. Normally, teaching assistantships (TAs) and research assistantships (RAs) are limited to full-time students. | Student should discuss changes in study status with their supervisor and considered the funding implications for such a change as well as implications to any provincial or national loan assistance, health or dental insurance. or visa status. |
| Change from full-time to inactive | Queen’s does not guarantee funding for inactive students. If a student moves to inactive status midway through an academic year once funding package has been assigned to them, the student may be able to complete any remaining work assignments and receive any remaining awards affiliated with their existing funding package. The student will not receive any further funding package until they resume their full-time studies. |
Student should discuss changes in study status with their supervisor and considered the funding implications for such a change as well as implications to any provincial or national loan assistance, health or dental insurance, or visa status. |
Why It Matters:
- Reduces misinformation risk: Answers match source documents verbatim.
- Supports trust: Users can verify claims via attributable sources.
- Human handoff: Flags when expert consultation is needed.
3. Accountability
Ensuring clear responsibility for Al outputs and decisions.
Our Protocols:
- Weekly stakeholder demos during development to align on outputs and address concerns.
- Human escalation for high-stakes domains (e.g., legal, medical, financial advice).
- Disclaimers for non-actionable guidance (e.g., “This is not legal advice”).
Example: Legal Chatbot Workflow
Assistant: Hi, welcome to Attorney Live Chat! Please proceed and ask your question.
User: I want to get divorced
Assistant: Thank you for asking your question for guidance on your matter through the Attorney Live mobile app. Please note that the information provided below is not legal advice, we’re providing this general information to clarify your legal matter request.
I understand this is a difficult situation. In which state do you need legal advice regarding your divorce?
User: Texas
Assistant: I’m so sorry you’re going through this challenging situation. What are the primary reasons you are seeking a divorce? Are there any children involved that custody arrangements would need to be made for?
User: Yes we have tried that but that didn’t work
Assistant: I’m very sorry to hear that counseling and mediation were unsuccessful. Have you discussed potential custody and visitation arrangements for the children with your spouse? It’s important to try to reach an agreement that is in the best interests of the children.
User: No
Assistant: Based on the context provided, it seems you are considering divorce in Texas and have concerns about child custody, division of assets and debts. This is a complex legal matter, especially when children are involved.
I would recommend consulting with an experienced family law attorney in Texas who can advise you on the divorce process, child custody arrangements, asset/debt division, and help ensure your rights and your children’s best interests are protected. An attorney can guide you through options like mediation or litigation if needed. Child custody and asset division can be emotionally charged issues. Having an attorney’s objective guidance can be invaluable in reaching a fair resolution while minimizing conflict. They can also ensure all legal requirements are properly followed in Texas.
Please remember that the information provided in this chat is not legal advice. Our goal is to guide you toward practical next steps, whether that involves consulting with an attorney or taking other appropriate actions. We kindly encourage you to schedule an attorney consultation through the Attorney Live app, so you can get the personalized support you need.
Why It Works:
- Boundaries: Never diagnoses or solves legal issues directly.
- Context gathering: Ensures nuanced understanding before responding.
- Ethical handoff: Directs users to qualified human experts.
4. Privacy and Data Security
Protecting sensitive data through infrastructure and process design.
Technical Safeguards:
- AWS-native encryption: End-to-end data protection with IAM policies and audit trails.
- Least-privilege access: Cross-account IAM roles deployed in client environments to isolate permissions.
- HIPAA-compliant anonymization: Pll masking for healthcare/law (e.g., [NAME], [MEDICATION]).
Data Handling:
- Zero retention: Processing occurs in-memory; raw inputs are never stored.
- Deterministic masking rules: Replace Pll with placeholders without altering context (e.g., preserves
“[MEDICATION] dosage”).
Why Clients Trust This:
- No cross-project exposure: Role-based access is scoped per engagement.
- Regulatory alignment: Meets GDPR, HIPAA, and 2025 U.S. state laws
5. Veracity and Robustness
Guaranteeing accurate, reliable outputs through technical and process controls.
PII Masking Example:
masking_prompt =
"""
Please mask all personally identifiable information (PII), specifically:
First names
Last names
Addresses
Dates (birthdays, appointment dates)
Medication names
Medical professional names (Doctors, nurses)
Phone numbers
Emails
Text to mask:
{text}
Do not rewrite or rephrase the rest of the sentence.
Only replace PIl with generic placeholders or masked terms (e.g., [NAME], [MEDICATION], [DATE]). Do not add explanations,
greetings, or comments. ONLY the masked sentence.
"""
Key Features:
- Strict redaction rules: No generative anonymization (prevents hallucinations).
- Context-aware: Retains clinical/financial meaning (e.g., “[MEDICATION] dosage”).
- Audit-ready: All masking logged with IAM roles and timestamps.
Why It Matters:
- Prevents fabrication: Outputs cannot invent sensitive details.
- Maintains utility: Masked data remains usable for analysis.
5. Governance – Adapting to New Standardsy
Adapting Al systems to 2025’s regulatory and threat landscape.
As an AWS Premier Tier Services Partner, we deliver:
- Early access to AWS’s newest compliance features (e.g., Bedrock guardrail updates before GA
- Certified implementations of AWS Well-Architected Framework for Al
- Shared responsibility for audit readiness – we manage AWS compliance tools so you don’t have to 2025 Compliance – Powered by AWS
| Regulation | AWS Feature to Leverage |
|---|---|
| U.S. EO 14179 | AWS AI Service Cards |
| EU AI Act | Bedrock Conformity Checks |
| State-Level Laws | AWS Policy Generator |
The Road Ahead
Scaling Responsibility with Innovation
As Al systems become more complex, scalable, and embedded in critical workflows, our responsibility doesn’t shrink, it grows. At Avahi, we’re not just building LLM-powered tools and automation pipelines. We’re building systems you can trust – systems that justify their decisions, protect privacy, and adapt to new laws and risks.
Whether it’s masking sensitive data, citing source policy in a chatbot, or aligning with AWS Bedrock’s evolving guardrails, our work remains rooted in transparency, accountability, and fairness.
Looking forward, we remain committed to leading by example. We will continue to collaborate with clients, regulators, and theopen-source community to advance ethical innovation and ensure Al systems uplift, not undermine, human potential.
