Guide

AI Security and Governance for Business: What to Fix First

By Ryan Lanyon, Founder and AI Systems Architect, Nexus Digital
Published 9 November 2026

AI security and governance for a business is the set of controls that stops an AI system from leaking your client data, writing to systems it should not, or giving your team an answer a regulator would question. Done well, it reads boring. Done badly, it is the thing that ends a contract.

What is AI security and governance?

It is two jobs in one phrase, a technical half and a human half. Security covers who the system is, where your data goes, and what the system is allowed to touch. Governance covers who decides those things, who signs off when they change, and who answers the question when a client, an insurer or a regulator asks.

The two live together because neither works alone. A well-secured system with no named owner is one resignation away from being nobody's problem. A governed system with a weak account is one leaked login away from a bad weekend. We treat them as one review.

The five controls we check first

Every AI review we run for an Australian business checks the same five things. We check the same five controls in every review, and the two that fail most often are account access and data leaving the network.

1. 1
   Access, who can see what. Which humans and which systems hold a login, what each one can read, and what each one can change. Most AI incidents inside a business start as access incidents. A former contractor still holds a Google account. A shared admin login is pinned in a password manager. A system account granted during a vendor trial is still running on full rights two years later. In our experience, the access half of the review usually surfaces at least one stale login nobody is accountable for.
2. 2
   Data flow, where it goes when you ask. When somebody on your team types a question into an AI tool, their words leave the room. Some vendors retain prompts for training. Some pass requests to a third country. Some promise neither in writing and do both in practice. The question is not whether this is bad. The question is whether you have it in writing from the vendor, and whether your clients would read that document the same way you do.
3. 3
   Oversight, who signs off when the system does something new. Any AI tool that writes an email, places an order, pays an invoice or updates a record needs a named human in the loop for the class of output that could be wrong in a way that matters. Payments. Client-facing copy. Anything a regulator would review after the fact. A one-line approval step is cheap. Explaining to a client why you had no approval step is not.
4. 4
   Logs, a trail by default. Every prompt, every output and every tool call kept for long enough to answer a question somebody will ask six months later. Most AI tools do a half job of this out of the box. Turn the rest on. If you cannot say what the system did on a Tuesday afternoon four weeks ago, the system is not auditable.
5. 5
   Owners, named for each system. One person inside the business owns each AI system. Their job is not to run the thing every day, it is to notice when it is going sideways and to be the one who stops it. No name means no accountability, no training, and no review cycle. Systems without owners get quietly switched off inside twelve months.

Who writes the policy

Somebody on the leadership team, not the IT provider. The policy belongs to the business because the business is the one that signs the client contracts and fills in the insurance renewals. The IT provider helps enforce it, and the vendor helps meet it, but neither of them is the author.

A workable AI policy for a small or mid business runs one page. It names which AI tools may be used, what data may be put into them, what signoff is needed before a new tool is added, and who holds the owner role for each system in use. One page is enough to pass most client security questionnaires and most insurer AI riders, which is where the policy earns its keep.

The two patterns Australian businesses miss

Two patterns show up in almost every review.

Pattern one, the browser tab problem. Somebody on the team has been using ChatGPT, Claude or Gemini in a browser tab for a year. The tool is on a personal account, not the business one. It is paid on a personal card and expensed quietly. The business has no record of what has been pasted into it. This is not a reason for a crisis meeting, it is a reason to move the account under the business, turn on the available retention controls, and put the user on the policy.

Pattern two, the shared admin login. A single admin account gets used by three or four people because setting up proper ones took effort and the trial was running out. The password is in a shared vault. When somebody leaves, the password changes. When somebody else leaves, it does not. By the time an AI tool is wired into that account, every history, every prompt and every output is attributable to nobody specific.

The Australian paper trail

Australian businesses answer to a set of standards stricter than most teams assume. The Australian Privacy Principles set what has to happen when personal information moves outside the country, and when a breach has to be notified. The Essential Eight is the baseline cyber control set the Australian Signals Directorate expects of businesses that handle client or government data. Neither one is AI specific. Both of them cover the AI case, because an AI tool is a system that reads and writes, and the rules apply to the data, not to the tool.

If your clients run a supplier security review, this is the ground their questionnaire covers. The time to meet it is before the questionnaire arrives.

How to start this quarter

Pick a two hour window, pull in your operations lead and the owner of each tool, and answer five questions on paper.

• Which AI tools is the team using today, on which accounts.
• Who has a login to each, and who no longer should.
• Where the data goes when a prompt is sent, from the vendor's own page.
• One page that names which tools are allowed, which data they may see, and who signs off before a new tool is added.
• A named owner for each system, listed next to it in the runbook.

That is the first pass. It will not catch the technical faults a proper review does, but it will clear the loud ones before the real review runs.

A proper review sits inside the paid Strategic Assessment, which checks data, systems, security, repeated work and ownership against the same five lenses across the business. If you are past that stage and want the controls built into the system itself, the AI implementation guide covers what the Implementation includes, the security line among them.

Book the Strategic Assessment when you want the whole picture written down.

Sources

• Australian Privacy Principles, OAIC
• Essential Eight, Australian Signals Directorate

FAQ

Want the next one in your inbox? Join the weekly AI newsletter and get the free AI security guide.