Your AI's Blast Radius Just Got Bigger

"The moment AI moves from drafting to sending, the tool has not changed. The risk has."

In an AI governance conversation recently, the answer sounded complete.

The tool had been approved. So had the people using it.

Then I asked what the tool could actually do.

Could it read email? Access company files? Draft a response? Send it? Create or delete something?

That was where the certainty disappeared.

Most AI policies are built around two binary questions. Is the tool approved? Is the person approved?

That made sense when a piece of business software had a defined purpose, predictable permissions and a relatively stable set of features. But an AI tool is no longer one thing. The same product can be a chatbot one day, connected to your email the next, and acting across your company's files soon after that.

Approving the tool is not the same as approving what it can do.

Reading is not sending. Drafting is not deciding. Every new permission changes the risk.

That is where AI governance is now being won or lost.

Think about an agent that helps manage your inbox.

At the first level, it scans your emails, identifies what matters and gives you a summary. The risk is relatively contained. The work stays inside your environment and you remain in control.

Then it starts drafting responses in your voice. That creates another significant productivity gain. You still review each message, change anything that does not sound right and decide whether to send it.

Then you give it permission to send.

It feels like a small step. One permission. One less click. But everything has changed.

An error no longer stays on your screen. It leaves the organisation wearing your name.

Private context could make its way into the wrong conversation. An internal comment about an employment matter could appear in an email it was never intended for. A poorly interpreted instruction could trigger messages to the wrong people. Something that previously took one person making one mistake could now happen repeatedly before anyone notices.

The tool has not changed. The user has not changed. The blast radius has.

And that is before the AI is connected to other systems.

Give it access to your calendar, company files, customer records or project platforms and each connection creates another pathway through the business. Connected well, these systems can turn hours of searching, sorting and administration into minutes. They can also allow one mistake to travel much further than anyone intended.

This is not an argument against agents or connected AI. I use these tools every day, and I am constantly surprised by what they now allow us to do. The reality is that the opportunity and the risk are expanding together.

As people become more capable with AI, they start seeing possibilities everywhere. They connect another system, add another instruction, automate another step and give the AI permission to take one more action.

That is exactly how the value gets unlocked. It also means your most proficient AI users are not automatically your lowest-risk users. They may be the people capable of creating the most powerful systems before the rest of the organisation understands what has changed.

That is not carelessness. Most leaders, boards and internal governance teams simply do not have enough experience with these tools yet to recognise every material change in risk. The technology is moving quickly, users are learning quickly, and new capabilities can appear inside a product that was approved months ago.

So organisations fall back to the controls they understand.

• Is the tool approved? Yes or no.

  
  

• Is the person approved? Yes or no.

  
  

• Is there a policy? Yes or no.

Those questions create comfort. They do not necessarily create control.

A policy that records an approved product name is a snapshot of one moment in a system that keeps changing. It can give the organisation confidence without giving it protection.

The answer is to govern the use, not just the tool.

That does not mean producing a new thirty-page assessment every time somebody changes a setting. It means recognising the points where the AI's access, authority or reach has materially changed.

There are three questions worth asking whenever an AI workflow develops.

1. What can it see?

What information can the AI access? Has somebody connected a mailbox, shared drive, customer system or other source that was not part of the original approval?

2. What can it do?

Can it only search, summarise and draft? Can it also send, change, delete or approve something? Can it act only when instructed, or can it decide when to act?

3. Who or what can it affect before a person can stop it?

Is the output contained on one person's screen? Could it reach a colleague, client or the public? Would a human see it first? If it went wrong, could the action be detected and reversed?

That is the blast-radius check. Each time one of those answers changes materially, the use deserves another look.

Start with the smallest permission that creates useful value. Use read-only access where possible. Let the AI recommend or draft before it acts. Keep a person between the AI and external communication, financial commitments, sensitive decisions or anything difficult to reverse.

Then test it in the real workflow. Watch where it struggles. Record what happens. Expand its authority only when the evidence justifies the next step.

That is not bureaucracy. It is how autonomy is earned.

No IT team, governance owner or leadership group can monitor every new way AI is being connected and used. The critical decision often happens at the user's desk, when they connect another data source, enable a new permission or allow the AI to take an action. That person needs enough understanding to recognise that the risk has changed and pause before they switch it on.

This is why governance and capability cannot be separated. A policy can establish the boundaries, but the people using the technology have to keep those boundaries alive as the tools change.

At your next leadership meeting, ask one question: what can our AI do today that it could not do when we approved it?

The name on the tool tells you very little about the risk. The permissions tell you almost everything.

Find out where your AI policy is leaking control

Most businesses don't need a bigger policy document. They need to know where their AI tools' permissions have expanded past what was originally approved.

The 5 Minute AI Readiness Snapshot helps you see where the gap between approval and actual access might be opening up, before a permission change turns into an incident.

Take the 5-Minute AI Readiness Snapshot

--

About the Author 

Mark Lucas is Co-founder and Managing Partner at Generation AI, where we help New Zealand business leaders build real AI capability.

Information provided is general in nature and does not constitute legal advice