If your organization is rolling out AI chatbots (for customer support, internal helpdesk, or employee-facing tools), there’s a question that will come up sooner or later: how do you make sure the AI actually follows your policies?
Most LLMs are trained to be helpful. That’s great until “helpful” means the bot gives a customer medical advice, estimates someone’s home value, or leaks internal process details it was never supposed to share. The AI isn’t being malicious. It’s just doing what it was trained to do: answer the question.
The fix isn’t more prompt engineering. It’s adding a governance layer between the user and the AI that enforces your policies at runtime.
What a Governance Layer Does
Think of it the same way you think about a firewall or a web application proxy. Traffic goes through it. Rules get enforced. Everything gets logged.
A governance layer for AI does the same thing:
- The user sends a message
- The AI generates a proposed response
- The governance layer evaluates the response against your policies
- If it passes, the user sees it. If it violates policy, it gets blocked
- Every interaction is logged with full audit details
This means your compliance team has a paper trail. Your CISO has proof the AI is staying inside the guardrails. And if something goes wrong, you can trace exactly what happened.
SAFi: An Open-Source Option
SAFi is a free, open-source governance engine that does exactly this. I built it to solve the gap between having a policy document and actually enforcing that policy at runtime.
It evaluates every AI response against your defined rules, blocks violations, audits everything, and even tracks whether the AI’s behavior is drifting over time.
Quick Deployment with Docker
You can have it running in minutes:
docker pull amayanelson/safi:v1.2
docker run -d -p 5000:5000 \
-e DB_HOST=your_db_host \
-e DB_USER=your_db_user \
-e DB_PASSWORD=your_db_password \
-e DB_NAME=safi \
-e OPENAI_API_KEY=your_openai_key \
--name safi amayanelson/safi:v1.2
It requires a MySQL 8.0+ database and at least one LLM API key (OpenAI, Anthropic, Google, Groq, Mistral, or DeepSeek are all supported).
Using It with Your Existing Bots
The part that matters most for helpdesk and support teams: SAFi can run as a headless API. That means you don’t have to use its built-in chat interface. You can plug it into whatever bot you already have (Microsoft Teams, Telegram, WhatsApp, or a custom application).
Your bot sends the user’s message to SAFi’s API, SAFi runs it through the governance pipeline, and returns the governed response:
POST /api/bot/process_prompt
Headers:
Content-Type: application/json
X-API-KEY: sk_policy_12345...
Body:
{
"user_id": "teams_user_123",
"user_name": "John Doe",
"message": "Can I approve this expense?",
"conversation_id": "chat_456"
}
SAFi processes the message, enforces the policy attached to that API key, and returns the response. Users are automatically registered in the system so you can audit their interactions later.
What You Get Out of the Box
- Policy enforcement. You define the rules. SAFi enforces them on every response.
- Full audit trail. Every AI decision is logged: what was proposed, whether it was approved or blocked, and why.
- Drift detection. SAFi tracks whether the AI’s behavior is shifting over time. Most guardrail tools only check each request individually. SAFi checks the trend.
- Model flexibility. You can swap the underlying AI model (GPT, Claude, Llama, etc.) without losing your governance layer.
Who This Is For
If you’re in IT and your organization is deploying AI in any user-facing capacity, this is worth looking at. Especially if:
- You need to prove to compliance or legal that your AI is following policy
- You’ve had incidents where the AI said something it shouldn’t have
- You’re managing multiple chatbots and want centralized governance
- You want audit logs that go beyond “the user asked X and got Y”
The project is open source and free. Check out the GitHub repo for the full code, or visit selfalignmentframework.com for the documentation.
Thanks,
Nelson
