SOP: Responsible Use of AI Tools

1

Purpose & Scope

Why this exists and who it applies to

This SOP helps the Primaris IT team use AI tools with confidence - knowing what is safe, what is expected, and what to avoid. It turns policy into practice.

Content is tagged by activity mode throughout this document:

📋 General AI Usage

Using AI to get things done faster - drafting, summarising, researching, communicating. Primarily Microsoft 365 Copilot.

⚙️ AI Development

Building with AI - writing code, creating automations, using Claude Code CLI. Additional guardrails apply.

Any team member may move between these modes. The tags are about the activity, not the person.

Policy aligned with: Canadian privacy regulations, EU AI Act (2026), and Primaris data governance standards.

AI Committee: A cross-functional committee guides AI priorities, manages risk, and ensures our approach aligns with Primaris values.

AI Hub in PrimeNet (coming soon): A space to learn, share practical use cases, and build confidence with AI tools day-to-day. Watch for its launch.

2

AI Tools at Primaris

Approved, restricted, and blocked tools

Rule: Using unapproved or public AI tools with company data - even out of curiosity - is not permitted. If unsure whether a tool is approved, contact IT first.

Tool	Status	Key Consideration
Microsoft 365 Copilot Teams, Outlook, Word, Excel, PowerPoint	✓ APPROVED - Primary	Operates within Primaris M365 tenant. Does not train external models.
Claude Code CLI Azure-deployed - IT team only	✓ APPROVED - IT Team	Primaris Azure deployment. Same data security as OneDrive. Not the public Claude.ai service.
ChatGPT (personal / free)	✗ NOT APPROVED	Free tier may use inputs to train models. No enterprise data agreement.
Google Gemini (personal)	✗ NOT APPROVED	No enterprise data protection for personal accounts.
DeepSeek	✗ NOT APPROVED	Unclear data handling and jurisdiction.
Groq	✗ NOT APPROVED	Public inference API. No enterprise data protection agreement with Primaris.
Perplexity AI (personal)	✗ NOT APPROVED	Public model. Do not input business data.
DALL-E / Midjourney / Sora / Veo	⚠ RESTRICTED	IP and copyright considerations. Generic tasks only, with IT approval.
Any other unapproved tool	✗ CHECK WITH IT	Never test with real business data. Contact IT before using.

Key distinction: A free personal ChatGPT account has entirely different data handling than our Azure-deployed Claude or Microsoft Copilot. Always verify which tier you are using.

3

What AI is For

Approved use cases by activity mode

📋 General AI Usage

Drafting emails, reports, and communications
Meeting preparation, agendas, and note summaries
Summarising vendor documents and technical specs
Creating presentations and status updates
Research on IT concepts and best practices
Brainstorming solutions and approaches
Drafting knowledge base articles and guides

⚙️ AI Development

Code generation, review, debugging, and refactoring
Workflow and automation prototyping
Internal tooling and integration development
SOP and technical documentation creation
Interactive HTML guides and workshop artefacts
Dashboard and portal development (internal)
SQL query writing and optimisation

📌 Validated by the Primaris IT Team: The team has successfully used Claude Code CLI to build interactive HTML workshop guides, dashboards, and SOPs - including this document. Artefacts are stored in Teams > Primaris IT Team > Projects > Productivity.

AI is not a replacement for: legal advice • financial or investment decisions • system architecture sign-off • executive or strategic decisions • HR decisions affecting employees

4

Data Safety & What Not to Do

Red / Amber / Green - what can and cannot go into AI

These rules apply to all activity modes. When in doubt, treat data as Red.

🔴 RED - Never

Do not input into any AI tool

Employee names, emails, phone numbers, HR records
Tenant / customer / visitor PII (lease details, contact info)
Non-public financial results or forecasts
Board materials, strategic plans, M&A targets
Legal matters, active contracts, litigation
Credentials, API keys, passwords

🟡 AMBER - Approved tools + anonymise first

Copilot or Azure Claude only

Internal process descriptions (anonymised)
Aggregate or general operational data
General financial categories (not specific figures)
Project plans with sensitive details removed
IT system descriptions (non-sensitive)

🟢 GREEN - Safe for all approved tools

All approved tools

Publicly available information
Generic communication templates
Code (without secrets or proprietary logic)
Brainstorming and ideation
General IT best practice research

🚫 Specifically Prohibited

Inputting Red data into any AI tool without anonymisation
Using free-tier public AI (ChatGPT, Gemini, DeepSeek, Groq) for any business task
Testing unapproved tools with real work data “just to see how they work”
Sharing AI-generated outputs externally without human review
Storing credentials, API keys, or passwords in AI sessions or prompts
Allowing AI agents to access production systems without prior IT approval
Using AI to create misleading, discriminatory, or deceptive content

Quick test before pasting into AI: “Would I be comfortable if this information were visible to an external third party?” If No or Maybe - stop and anonymise or check with IT first.

5

AI Development Guardrails

⚙️ AI Development - Claude Code CLI

Claude Code CLI at Primaris runs against an Azure-deployed internal model - not the public Claude.ai service. Prompts stay within Primaris infrastructure. However, Claude Code can read, write, and delete files on your local machine - which requires deliberate guardrails.

🤖 Approved Model Selection Guide

Primaris has two approved models deployed via Azure AI Foundry. No other models may be used for any development or AI activity.

SONNET

Everyday workhorse

pmz-eai-productivity-sonnet-4-6

Research and planning
Documentation and SOPs
Building documents and presentations
Quick analysis and dashboards
General productivity tasks

OPUS

Complex work

pmz-eai-productivity-opus-4-6

Complex development requirements
Multi-step reasoning and architecture
Deep code analysis and refactoring
Use when Sonnet is insufficient for the task

No other models are approved. Using any model not listed above - including public Claude.ai, OpenAI models, or any other provider - violates Primaris IT policy. When in doubt, check with IT before configuring any model endpoint.

🤖 Recommended: The organization recommends using the Primaris Model Selection skill file to manage model usage workloads. This skill file configures Claude Code to automatically route tasks to the correct approved model based on complexity and task type.

⬇ Download primaris-model-selection.md

⚠ Mandatory Guardrails

📁

Scope Your Working Directory

Always launch from a specific project folder. Never point at root drives, shared drive roots, or production paths.

🗑

File Deletion Risk

Claude Code can delete files. Never ask it to “clean up” on shared or important directories without a backup.

✅

Review Before Accepting

Always read proposed changes before writing them. Never bulk-accept without reviewing what is being modified or deleted.

🔌

Anthropic-Published Skills Only

Only install skills and MCP extensions published by Anthropic. Never install community or third-party public skills - they can execute code or exfiltrate data.

🔎

No Unreviewed Code in Production

Never deploy AI-generated scripts directly to production. Test first. Human review of all generated code is required.

🔗

Review Agentic Actions

When Claude Code proposes to call APIs, access external services, or run system commands - review carefully before confirming.

📄

Save & Share Outputs

Store artefacts in Teams > Primaris IT Team > Projects > Productivity for shared access and audit trail.

🔒

Protect Secret Keys & Credentials

Secret keys, API keys, and credentials must never be shared between users. All secrets must be stored in Keeper Vault. Never commit .env files or any keys to Azure DevOp, use Azure App Service Application Settings or a Key Vault reference instead.

6

Responsible Use

Ethics, risks, and cost awareness

▶ Do’s and Don’ts

✓ DO

Verify AI outputs before publishing or distributing
Maintain human oversight - AI augments, never replaces decisions
Disclose AI assistance in formal or external content
Apply critical thinking to communications that may be AI-generated
Respect intellectual property in AI-generated content or code
Report concerns about AI misuse or data handling to IT

✗ DON’T

Create discriminatory, misleading, or deceptive content
Distribute AI outputs without human review
Assume AI is accurate - hallucinations are real
Use AI for decisions affecting employees, tenants, or investors without human accountability
Create personal API keys using company billing information
Leave automated agents running unattended without monitoring

🔌 Key Risks to Know

Risk	Level	How to Avoid It
Data leakage via public AI tool	Critical	Only use approved tools (Copilot, Azure Claude)
AI hallucination published externally	High	Human review before any external distribution
Agentic AI file corruption or deletion	High	Scope working directory, review all proposed changes
Unreviewed AI code deployed to production	High	Code review required before any production deployment
AI-powered phishing / social engineering	Medium	Verify suspicious communications through a second channel
API cost overrun from uncontrolled workflows	Medium	IT lead approval required for bulk / automated workflows

Human oversight is always required. AI should augment decision-making, not replace it. A person should always be accountable for decisions that affect tenants, employees, investors, or the public.
- Joseph Martino, VP Information Technology

7

Quick Reference

Checklists and decision guide - progress saves in your browser

📋 General AI Usage - Before You Start

I am using an approved tool (Microsoft 365 Copilot or Azure-deployed Claude)
My data does not contain Red Zone content (PII, financials, board materials, legal, HR)
If using Amber data, I have anonymised it first
I understand AI can produce inaccurate content and will review the output
If the output is for external use, a human will review it before distribution

⚙️ AI Development - Claude Code CLI Session

Claude Code is configured to the internal Azure-deployed model (not public Claude.ai)
Working directory is scoped to a specific project folder - not a root drive or shared drive root
No Red Zone data (PII, credentials, financials) is in my working directory
I will review all proposed file changes before accepting - no bulk-accept without reading
Only Anthropic-published skills/extensions are installed - no community or third-party skills
Generated code will be reviewed before deployment to any environment
Team artefacts will be saved to: Teams > Primaris IT Team > Projects > Productivity
Any automated or bulk workflow has been approved by the IT lead before running at scale

📊 Can I Use AI for This? - Quick Decision Guide

I want to use AI for a task

↓

Is the tool on the Primaris approved list?
(Copilot / Azure Claude)

↓

NO

❌ Stop. Check with IT before using any unapproved tool.

YES

Does the task involve Red Zone data?
(PII, financials, board, legal, HR)

↓

YES

❌ Anonymise first, or use a non-AI approach.

NO

Will the output be shared externally or used in a financial/legal context?

↓

YES

Proceed - human must review before distribution.

NO

✓ Proceed with AI

8

Incident Reporting

What to report and how

🚨 What Counts as an AI Incident?

Sensitive or restricted data submitted to a public or unapproved AI tool
An AI agent modified or deleted files unintentionally
AI-generated content containing inaccurate information was distributed externally without review
An unapproved AI tool was used for business tasks
Suspicious communications believed to be AI-generated targeting Primaris staff
Unexpected or excessive Azure API usage suggesting a runaway process

📞 How to Report

1

Stop the activity. If a data incident is in progress, stop immediately. Do not continue the task.

2

Contact the IT Department. Include: what happened, which tool was used, what data was involved, and approximate time.

3

Escalation: VP Information Technology. For significant breaches affecting tenants or employees, the VP IT determines next steps including any disclosure obligations.

No-blame culture for good-faith reporting. Reporting promptly is the right thing to do and will be treated as such. Concealing an incident is far more serious than the incident itself.