Skip to content

AI Governance and Acceptable Use Policy

Organization: Cloudnaut Technologies LLP
Classification: Internal
Document status: Effective
Version: 1.0
Owner: Management / Security Owner
Effective date: 2026-10-15
Review cycle: Annual, or sooner given the pace of change in AI tools and regulation

Document control

Role Name
Prepared by Management / Security Owner (Cloudnaut)
Reviewed by Sandeep AC
Approved by Sandeep AC

Revision history

Version Date Summary Reviewed by Approved by
0.1 2026-10-15 Initial draft: AI tool use and AI/agentic system engineering. Sandeep AC —
1.0 2026-10-15 Initial effective release. Sandeep AC Sandeep AC

Purpose

This policy governs (a) how Cloudnaut personnel use AI tools in their work, and (b) how Cloudnaut designs and builds generative AI and agentic AI systems for customers, so that customer information stays protected, outputs are reviewed by people, and customer requirements are respected.

Scope

  • AI tool use: AI coding assistants, chat assistants, and other AI services used by personnel for any Cloudnaut or customer work.
  • AI engineering: GenAI and agentic solutions Cloudnaut builds for customers (for example on Amazon Bedrock).

Roles and responsibilities

Role Responsibilities
Management / Security Owner Approves AI tools; maintains the approved tool list; approves exceptions.
Engagement Lead Confirms customer permission for AI tool use on the engagement and records any restrictions.
Personnel Use only approved tools within permitted data limits; review all AI output before use.

Part A — Use of AI tools

  1. Approved tools only. Only AI tools on the approved list may be used for Cloudnaut or customer work. Approval considers data retention, whether inputs are used for model training, data location, enterprise controls, and the supplier review under the Third-Party and Supplier Security Policy. The approved list, including editions and required settings (for example business or enterprise tiers with training on inputs disabled), is maintained internally by the Security Owner.
  2. Customer permission first. AI tools are used on a customer engagement only where the customer contract and the customer’s instructions permit it. Where the customer prohibits or restricts AI use, IDE and tool settings are configured accordingly (see step 7 of the procedure in the Approved Development Tools and Plugin Policy).
  3. Data limits. Information classified Restricted (credentials, secrets, personal data, regulated data) is never entered into AI tools. Confidential information, including customer source code, is entered only into approved tools the customer has permitted. See the Data Handling Policy.
  4. Human review. All AI-generated code, configuration, and documents are reviewed by a qualified person before use and pass the same peer review, scanning, and testing as human-written work under the Secure SDLC Policy. The author remains accountable.
  5. Intellectual property and licensing. Personnel check that AI output does not reproduce third-party code in breach of licence terms, and follow customer IP requirements.
  6. No autonomous action on customer systems. AI agents or assistants are not given credentials to customer environments or allowed to execute changes there unless the customer has approved that specific use and its permissions.
  7. Transparency. Where a customer asks, Cloudnaut discloses which AI tools were used in a deliverable.

Part B — Engineering GenAI and agentic AI systems for customers

  1. Customer ownership. The customer owns the AI system’s use case, data, model selection approval, and acceptance of risk; Cloudnaut designs and builds to agreed requirements. See the Engagement Responsibility Model.
  2. Data protection. Designs keep customer data within the customer’s account and chosen region, use managed services whose terms exclude training on customer inputs (for example Amazon Bedrock), and encrypt data at rest and in transit.
  3. Least privilege for agents. Agents and their tools/actions receive only the permissions they need; destructive or high-impact actions require human approval; tool inputs are validated.
  4. Prompt injection and output handling. Designs treat user input and retrieved content as untrusted, apply guardrails (for example Amazon Bedrock Guardrails or equivalent) for content filtering and sensitive-data redaction where appropriate, and validate model output before it triggers actions.
  5. Testing and evaluation. Systems are evaluated before release for accuracy, harmful output, sensitive-data leakage, and prompt-injection resistance, with results shared with the customer.
  6. Logging and monitoring. Designs include logging of prompts, responses, and agent actions as the customer requires, with logs treated as potentially containing personal data.
  7. Test data. Development uses synthetic or customer-approved data.

Exceptions

Exceptions require written management approval and, for customer work, customer approval, recorded in the engagement record.

Records

Approved AI tool list and approvals, customer permissions per engagement, exception approvals, and AI system evaluation results.