You Can’t Block AI — You Have to Govern It

‍

Over the last few months, we’ve heard a consistent concern from risk, security and compliance leaders:

‍

“Our staff are uploading sensitive documents into their personal ChatGPT accounts, and we can’t fully stop it.”

‍

In most cases, public AI tools are already blocked on corporate networks. Microsoft Copilot is enabled internally. Data Loss Prevention policies are in place.

‍

And yet — the behaviour persists. This is not a failure of technology. It’s a signal that the problem has been framed incorrectly.

‍

The uncomfortable truth: AI usage is already happening

The idea that organisations can “decide” whether employees use AI is no longer realistic.

People are using AI because:

• It makes them faster
• It removes friction
• It feels easier than formal processes

Trying to eliminate that behaviour entirely leads to shadow usage:

• Personal devices
• Home networks
• Personal accounts
• No logging, no audit trail, no governance

From a regulator or board perspective, that posture is hard to defend.

Blocking tools doesn’t address intent

Most organisations respond to AI risk by focusing on tool restriction:

• Block external LLMs
• Approve one sanctioned tool
• Rely on policy statements

That approach assumes compliance through prevention.

But regulators, including those shaping APRA‑aligned and privacy‑driven controls, increasingly recognise a harder truth:

you cannot eliminate human intent — you can only reduce risk and prove control.

This is precisely the scenario Microsoft and regulators are now designing for.

‍

The shift: from “AI blocking” to AI Access Governance

Microsoft refers to this emerging control model as AI Access Governance, aligned to its Responsible AI and Secure AI Adoption frameworks.

We see it working in practice as a layered governance model, not a single control.

‍

At a high level, it has four parts:

1. Prevent: reduce the chance sensitive data leaves the organisation

This is where most organisations already have foundations:

• Microsoft Purview DLP policies
• Endpoint, device and browser controls
• Conditional access and session restrictions

The goal here isn’t perfection — it’s risk reduction at scale.

If sensitive data can’t easily leave managed environments, opportunistic misuse drops sharply.

2. Detect: gain visibility into real‑world AI behaviour

This is where many strategies fall down.

Tools like:

• Purview AI Hub
• Data Security Posture Management (DSPM) for AI

allow organisations to:

• Detect when sensitive content is used in AI prompts
• Identify trends and repeat behaviour
• Separate accidental misuse from sustained risk

This matters because regulators care less about isolated incidents — and far more about patterns and response.

‍

3. Educate: make safe AI the default, not the exception

Blocking without enablement pushes behaviour underground.

Clear guidance matters:

• What’s acceptable to use in Copilot
• What must never go into external tools
• Why Copilot is the preferred path

A “Copilot‑first” policy, backed by real training, is often one of the highest‑return controls available.

People generally want to do the right thing — if the path is obvious and supported.

‍

4. Enforce: consequences still matter

Governance without enforcement is just advice.

Effective programs:

• Escalate repeat or high‑risk behaviour
• Apply role‑appropriate consequences
• Treat AI misuse like any other data policy breach

This closes the loop — and critically, creates a defensible position if scrutiny arrives.

‍

Why this approach holds up under regulatory pressure

What this layered model provides is not just alerts after the fact, but:

• Real‑time guardrails
• Auditability
• Clear accountability

It acknowledges reality:

• AI will be used
• Mistakes will happen
• Controls must be proportionate, documented, and enforceable

From a board, regulator or insurer perspective, this is a far stronger position than “we blocked the tools and hoped”.

Where Argenti sees the real gap

The technology largely exists.

The challenge we see repeatedly is connection:

• Between Purview, identity, endpoint and Copilot controls
• Between security teams and business users
• Between policy intent and lived behaviour

AI governance fails when it’s treated as:

• A single security project
• A single Microsoft feature
• A once‑off policy update

It succeeds when people, process and technology are designed together — with AI treated as a permanent operating shift, not a temporary threat.

‍