AI assurances: A deep dive

“What assurance do we have that this AI product is safe and appropriately controlled?” Let’s do a deep dive.
My first question would be: What exactly are we assuring?
The model? The system around it? How it's deployed? Or the outputs and actions it produces?
These are four different assurance questions: MODEL → SYSTEM → DEPLOYMENT → OUTPUT

1.      Model assurance asks: What can the underlying model do, how was it evaluated, and where does it fail?

Think hallucination, reasoning, calibration, cyber capability, robustness, and safety evaluations. But knowing the model is not enough.

2.      System assurance asks: What controls surround the model?

Think system prompts, RAG, tools, permissions, model routing, content filters, citation verification, human approval, and logging. A capable model inside a poorly controlled system can still create unacceptable risk.

3.       Deployment assurance asks: Is this AI system appropriately controlled for this specific use, environment, users, and consequences?

The same model summarizing public filings presents a fundamentally different risk when connected to ERP, email, customer records, banking, or production infrastructure, with permission to act.

·         A safe model does not guarantee a safe system.

·         A safe system does not guarantee a safe deployment.

4.      Then comes Output assurance.

Can we independently verify the consequential claims, recommendations, calculations, or actions the AI produces?

·         A citation tells you where information came from. It does not tell you whether the AI interpreted the source correctly.

·         A calculation can be mathematically correct while relying on the wrong inputs or assumptions.

·         An AI agent can successfully execute an action without having authority to take it.

For consequential outputs, ask:
What did the AI claim? What evidence supports it? Does that evidence justify the claim? How was it independently checked? Can we reconstruct what happened afterward? 

 A vendor system card provides valuable evidence about an AI system. It cannot prove that an individual AI-generated claim or action is correct. That is the difference between AI that appears trustworthy and AI whose work can be examined, challenged, and audited. AI governance must move from assurance by assertion to assurance by evidence.