User responsibilities and best practices
This topic covers responsibilities and best practices for using agentic and non-agentic AI coding tools safely, responsibly, and in compliance with policy.
Responsibilities are defined at two levels:
- Project level β for project and tech leads
- End user level β for all users, including agentic users
Project level best practices
Project and tech leads are responsible for ensuring AI coding tools are used safely, responsibly, and in compliance with Requirements for AI tool clearance and Security and Sensitive Classification policies.
CII/SII systems: If your system is classified as Critical Information Infrastructure (CII) or Sensitive Information Infrastructure (SII), you must obtain system-level approval before using agentic AI tools. SHIP-HATS tool clearance alone is not sufficient. See Additional approvals.
| Responsibility | Actions |
|---|---|
| π Align usage with policy | Ensure your team uses AI tools in compliance with policies related to system class, classification, and sensitivity rules. See Security and Sensitive Classification. |
| π Manage access | Enforce least privilege and role-based access for agentic AI tool usage. Prohibit the use of admin tokens, elevated credentials, or unrestricted access with AI assistants or agents. |
| ποΈ Oversee development process | Follow a rigorous code review workflow with human involvement. Review all AI-generated or AI-assisted code before it is merged into production. |
| π‘ Evaluate use cases | Assess the suitability of AI tools for each project use case. Approve AI usage only where it adds value. |
| π© Report incidents | Report unsafe or unexpected AI behaviour promptly to the platform or security team. |
End user level best practices
All users β including developers writing, modifying, or reviewing code, and users generating content (for example, blogs or related materials) β must comply with these responsibilities when using AI tools as part of their work.
Agentic users are a subset of end users who use AI agents with capabilities such as running commands, iterating on code, or creating merge requests. They must comply with all end-user responsibilities, plus the agentic-specific responsibilities below.
End user responsibilities
| Responsibility | Action |
|---|---|
| π§ Responsible usage | Use AI tools responsibly by following the Instruction manuals β Use of Generative Artificial Intelligence in the Public Sector (accessible to public officers only via intranet). |
| β Output validation | Assess AI-generated output before committing and make necessary adjustments. You are ultimately responsible for the output. |
| π Data compliance | Do not share or input data ineligible for AI features. Ensure code, prompts, and data comply with your projectβs classification level. See Code Classification Overview. |
| π Credential security | Do not share API keys, tokens, or access secrets used with AI coding tools. Follow organisational policies for credential rotation and secure storage. |
| π» Secure your environment | Keep your developer environment and workflows aligned to policy, security, and project requirements. Do not circumvent safeguards. |
| π© Report incidents | Report unsafe or unexpected AI behaviour to your project lead, platform team, or security team. |
Agentic user responsibilities
In addition to the end user responsibilities above, users enabling AI agents must also follow these practices:
| Responsibility | Action |
|---|---|
| π Human oversight | Keep a human in the loop for all AI-driven decisions and code changes. Developers remain ultimately responsible for the code. |
| β Merge request validation | Review all merge requests thoroughly, especially those generated by AI. Ensure one or more developers validate the changes. |
| π Least privilege | Apply the principle of least privilege. Neither agents nor developers should push code directly to production. |
| βοΈ Traceability | Sign and label all agent-generated commits to ensure traceability. |
| π Trusted connections | Connect to MCP servers only from trusted sources. MCP is available for Claude Code (GovTech users only) via whitelisted MCPs. MCP remains disabled for other tools. |
Risks and mitigation
AI coding tools can improve development speed and quality, but introduce risks that users and project teams must manage:
| Risk | How to mitigate |
|---|---|
| Copyright infringement | Scan code for copyright and open-source licence violations before merging |
| Security vulnerabilities | Scan and fix security issues before merging into any environment |
| Defective code | Inspect, test, and peer-review all AI-generated code before merging |
| Logical flaws or malicious content | Review code for prompt injection instructions and logical errors at the development stage |
Any AI-generated content used in development is the responsibility of the user and project team. While SHIP-HATS implements policy, security, and contractual safeguards, responsible AI usage remains with the team.
