Responsible AI Policy

1. Purpose and Scope

This policy governs how Horus Technology designs, builds, deploys, and operates AI systems for clients and internal use. It applies to all engagements, all personnel (employees, contractors, and partner pod resources), and all AI models and services used in delivery, regardless of provider.

2. Accountability

Designated Responsible AI Owner: Michael Walker, CEO. The CEO is the single point of accountability for responsible AI use, with authority to approve or restrict use cases, commit resources, approve scope changes, and resolve escalations.

Escalation path: Delivery Engineer to Solution Architect (Security Lead) to CEO. Any team member or client may raise a responsible AI concern at any tier; concerns reaching the CEO receive a written disposition within 2 business days.

3. Use Case Authorization

Every AI use case is classified at engagement scoping and re-reviewed at each SOW milestone.

Authorized by default (standard review):

• Document extraction, classification, and summarization with human-in-the-loop review
• Retrieval-augmented question answering over client-owned corpora with tenant isolation
• Internal productivity and code-assistance tooling

Requires CEO approval before build (elevated review):

• Systems processing PHI, PII of minors, or financial account data
• Any straight-through processing where AI output triggers action without human review
• Client-facing generative output published under the client's name
• Use of client data for model training or fine-tuning

Restricted (we will not build):

• Systems making final adverse decisions about individuals (credit, employment, benefits denial) without meaningful human review
• Surveillance or biometric identification outside a clearly lawful, contracted scope
• Systems designed to deceive end users about whether they are interacting with AI
• Any use prohibited by the model provider's usage policies

4. Engineering Controls

• Human-in-the-loop by default. Consequential outputs route through confidence-gated review queues. Low-confidence extractions go to an exceptions queue for human review rather than proceeding.
• Ground-truth discipline. Training and evaluation data is drawn only from human-approved outputs. Ground-truth provenance is versioned and leakage between training and evaluation sets is checked.
• Evaluation gates before scale-up. Systems must meet field-level accuracy thresholds measured against human-approved ground truth, including on worst-case documents, before expanding scope or volume.
• Hallucination controls. Known failure modes are managed with blocklists, structured output validation, and deterministic post-validation where correctness is checkable.
• Least-privilege infrastructure. All deployments follow our cloud account governance standard: least-privilege IAM, guardrail policies, full audit logging, MFA, and SSO federation. Client data is processed in the client's own cloud account wherever feasible.

5. Data Handling

• Business Associate Agreements are executed before any access to protected health information
• Client data is never used to train models for other clients
• Multi-tenant systems enforce tenant isolation (per-tenant indexes, row-level security)
• Data retention and deletion follow the governing agreement; no client data is retained beyond contractual need

6. Responsible AI in Document Processing for Regulated Industries

Much of our work is intelligent document processing and AI/ML for banking, insurance, healthcare, and government — domains where an extraction error carries real consequences for a person or an audit. In these engagements responsible AI is not an abstract principle; it is a set of controls baked into the pipeline.

• Confidence-gated extraction. Fields extracted with low confidence are never written straight into a system of record. They route to an exceptions queue where a person reviews and corrects them before the record proceeds.
• Accuracy measured against ground truth. Before a document workflow expands in scope or volume, it must meet field-level accuracy thresholds measured against human-approved ground truth — including on the worst-case, hardest-to-read documents, not just the easy majority.
• Compliance preserved, not bypassed. Automation speeds up processing without removing the review and audit steps regulators expect. When we automated handwritten loan-application processing with Amazon Bedrock, faster turnaround was delivered alongside — not instead of — the human checks that protect accuracy and compliance.

7. Human-in-the-Loop Review

Human oversight is the default for consequential outputs, not an optional add-on. We design review workflows so that people stay in control of decisions that affect individuals, money, or safety.

• Consequential outputs pass through review queues before they trigger any downstream action.
• Straight-through processing — where AI output acts without a human in the loop — requires explicit CEO approval before we build it, and we will not build systems that make final adverse decisions about individuals without meaningful human review.
• Reviewer corrections feed back as approved ground truth, so the system's measured accuracy reflects real, human-validated outcomes rather than the model grading its own work.
• The volume of items reaching the exceptions queue is itself a monitored signal: a rising exception rate is treated as an early warning of drift or a new document type, not something to quietly suppress.

8. Data Privacy on AWS

Because we build primarily on AWS, our data-privacy controls are enforced through cloud account governance rather than policy documents alone. This applies to both our intelligent document processing and our generative AI services.

• Process data in the client's account. Wherever feasible, client data is processed in the client's own AWS account, so the data never leaves the boundary they already control and audit.
• Least privilege by design. Deployments follow our cloud account governance standard: least-privilege IAM, guardrail policies, encryption, VPC isolation, full audit logging, MFA, and SSO federation.
• Tenant isolation. Multi-tenant systems enforce per-tenant indexes and row-level security, and client data is never used to train models for other clients.
• Contractual data lifecycle. Business Associate Agreements are executed before any access to protected health information, and retention and deletion follow the governing agreement with no client data retained beyond contractual need.

9. Incident Response

AI incidents (material accuracy regressions, data exposure, harmful output reaching an end user) follow the escalation path in Section 2, with a written root-cause analysis produced for any client-impacting incident and corrective actions tracked to closure.

10. Review

This policy is reviewed by the CEO at least annually and upon any material change in model providers, regulatory requirements, or engagement risk profile.

11. Contact

Questions about this policy or a responsible AI concern: