Responsible AI Policy

Dhari AI builds agentic AI systems for regulated enterprises. We hold ourselves — and the systems we deliver — to a standard that goes beyond what is legally required. This policy sets out our commitments, structured around the Monetary Authority of Singapore's FEAT principles: Fairness, Ethics, Accountability, and Transparency.

Our core commitment

We will not build, deploy, or recommend an AI system that we would not be comfortable having scrutinised by a regulator, an affected customer, and our own conscience.

Fairness

We design AI systems to produce equitable outcomes across the populations they affect. In practice:

• We test for disparate impact across customer and employee segments where lawful to do so
• We validate that confidence and escalation thresholds are calibrated consistently across groups
• We treat unexplained outcome disparities as defects requiring investigation, not as acceptable noise

Ethics

Every AI system we build operates within an explicit ethical envelope:

• Our agents never represent themselves as human in customer-facing contexts
• We do not build systems that exploit cognitive biases for institutional benefit
• Every agent has a defined refusal envelope — actions it will not take regardless of instruction, including assisting regulatory evasion, fabricating evidence, or acting outside delegated authority
• We decline engagements where the intended use of AI would cause foreseeable harm

Accountability

AI does not dilute human accountability — it must strengthen it. Our systems are built so that:

• Every material agent decision is logged with timestamp, model version, prompt version, input, retrieved context, tools called, output, and confidence
• Prompts and models are version-controlled and changes are managed like any production change
• Decisions above defined materiality thresholds require human sign-off, recorded immutably
• A named human owner is accountable for each deployed agent — there is no "the AI decided" defence

Transparency

We maintain transparency for two distinct audiences:

• For regulators and auditors — model cards, system cards, evaluation reports, and drift monitoring documentation are maintained for every production system
• For affected individuals — where an AI system materially influences a decision affecting a person, we ensure that fact is disclosed, the basis is explainable, and a meaningful path to challenge the outcome exists

Data and privacy

• We do not use client or personal data to train foundation models
• We design for data residency, applying PDPA standards as a baseline regardless of jurisdiction
• We conduct Data Protection Impact Assessments for use cases involving personal data
• We minimise data — agents access only the data necessary for their task

Human oversight

We design for human-in-the-loop and human-on-the-loop oversight appropriate to the risk of each use case. We never deploy fully autonomous agents for decisions with significant impact on individuals, markets, or institutional safety without a human accountability checkpoint.

Model and vendor neutrality

We are not tied to any single model provider. We select models — frontier or open-weight — based on the requirements of the task, including accuracy, cost, latency, and data residency. This neutrality lets us recommend what is right for the client rather than what is commercially convenient for us.

Alignment with frameworks

Our practice is designed to align with:

• MAS FEAT Principles (Fairness, Ethics, Accountability, Transparency)
• IMDA Model AI Governance Framework and AI Verify testing framework
• Singapore PDPA and PDPC advisory guidelines on AI
• The EU AI Act risk-based framework, for clients with European exposure
• NIST AI Risk Management Framework, as a technical reference

Continuous review

AI governance is not static. We review this policy and our practices regularly as regulation, technology, and our own understanding evolve. We welcome scrutiny and feedback.

Contact

Questions about our responsible AI practices? Contact us at security@dhari.ai.