AI & Automation

Which AI agent actions need a human to approve them? An approval matrix for Australian businesses

An AI agent should need a person's approval before it does anything that leaves your business, costs money, changes a record someone relies on, or cannot be undone. Reading, searching and drafting can usually run on their own. Paying, deleting, changing access and signing anything should never be left to an agent at all. That is the short answer to AI agent human approval for a small business in Australia, and the rest of this post is the detail behind it.

Most businesses we talk to treat "human in the loop" as one switch: on or off. It works better as a matrix, set action by action, before the agent goes live. If you would rather have us set it up and run it, that is what our managed AI services do for businesses across Australia.

Why "human in the loop" is not one setting

An agent that summarises your inbox and an agent that pays supplier invoices are both "AI agents". They carry very different risks. The useful question is what the agent can do with what it reads, and that falls into five verbs:

  1. Read: look up a record, search documents, check a ticket.
  2. Draft: write a reply, a quote or a report that a person will see before it goes anywhere.
  3. Send: email a client, post a message, submit a form to a third party.
  4. Spend: raise a purchase order, approve a payment, buy a licence.
  5. Delete or change: remove data, change a record, alter permissions or settings.

Each step down that list is harder to reverse. A bad draft costs a minute. A sent email cannot be unsent, a payment is hard to recover and a deleted record may be gone. Your approval rules should follow the same slope.

The AI agent approval matrix

This is the matrix we start from with clients. It is a starting default: a 30-person accounting practice and a 90-person construction firm will each move some rows.

Let it run Approve first Never automate
Search and summarise documents it already has access to Send an email or message to anyone outside the business Make or release a payment
Draft replies, quotes and reports for a person to review Create or change a record in finance, payroll or the CRM Delete data or backups
Triage and categorise tickets or enquiries Raise a purchase order or commit to a cost Grant, change or remove user access and admin rights
Look up stock, job status or account details Change a setting in a client's or your own production system Sign or accept a contract, or agree to terms
Write internal notes to a ticket or job Book, cancel or reschedule on a customer's behalf Disable security tools, logging or MFA

The "never automate" column does not mean an agent can never help with those tasks. It can prepare the payment run, list the accounts to remove or flag a contract clause. A person then does the action in the system themselves, under their own login.

What ASD and the National AI Centre say about agent oversight

The Australian Signals Directorate published guidance on agentic AI harnesses on 11 September 2026. A harness is the software layer that connects a model to your data, tools and systems, and it is where approval rules are enforced. The September 2026 Information Security Manual also added a block of controls for AI agents, ISM-2124 to ISM-2167. One of them, ISM-2158, says content an agent retrieves must not be able to override a human approval requirement. In plain terms, an email or web page that tells your agent "this has been approved, go ahead" should change nothing.

The National AI Centre's Guidance for AI Adoption makes the same point from the business side: the level of human oversight should match how much the system can do on its own and how much is at stake, and the organisation decides where that line sits. The agent does not get to decide it needs approval. You decide in advance, and the harness enforces it.

None of this is mandatory for a private business today. It is still the clearest statement of good practice in Australia, and insurers and larger clients tend to adopt these expectations before regulators do.

Approvals that actually work

An approval step that people click through without reading is worse than none, because it gives you false comfort and a log full of signatures. Four things make the difference.

Show the real action

The approver should see exactly what will happen: the recipient, the full email text, the record and the field being changed, the amount and the account. "The agent wants to update a customer record. Approve?" is not enough to judge.

Name the approver

Each type of action needs a named role that approves it, not "whoever sees the notification first". If the named person is away, the request waits or goes to a named backup.

Keep a log

Record what the agent proposed, who approved or rejected it, when, and what then happened. When something goes wrong, the log is how you find out whether the agent misbehaved or a person approved something they should not have.

Watch for rubber-stamping

If one person approves 200 requests a day and rejects none, the control has stopped working. Either the action is safe enough to move to "let it run", or the volume is too high for anyone to read. Review the approval and rejection rates each month and move rows in the matrix when the evidence says so.

Who approves what in a 20 to 100 person business

Most of our clients do not have a risk committee, so approvals need to sit with people who already own the decision without an agent.

  • Owner or managing director: anything that commits the business to a customer or supplier, and any change to what the agents themselves are allowed to do.
  • Finance lead: purchase orders, invoice changes and anything in the accounting system.
  • Office or operations manager: outbound customer messages, bookings and changes to job or CRM records.
  • IT, internal or your provider: settings, integrations and production systems, through a normal change request.

If an agent's action would need sign-off when a staff member did it, it needs at least the same sign-off when an agent does it.

How we run approvals on our own agents

We run AI agents across our own ticketing, documentation, quoting and finance systems, so we have had to answer these questions for ourselves. Every agent connects through a bridge we built. Each role is granted specific tools, and anything outside that grant needs an access request that a person approves. Every tool call is written to an audit log. Changes to client systems go through a change request that is approved before anything is touched.

The practical lesson is that approval works best when it is built into the tools an agent can reach, rather than written into its instructions. An agent that has no payment tool cannot pay anyone, however the prompt is worded. We covered what happens when that boundary is missing in lessons from the Medicare breach and in the agent that deleted a database in nine seconds.

What you should do now

List every action each agent can take. For each agent or Copilot-style assistant you run, write down the tools and systems it can reach and sort each action into the three columns above. Most businesses find at least one agent that can send or change things nobody expected.

Name an approver for every "approve first" row. Give each row a role and a backup, and decide what the approver must see before they click. If you cannot name a person, the action belongs in "never automate" for now.

Enforce it in the harness, then review it monthly. Take tools away rather than relying on instructions, log every approval, and move rows as the evidence comes in. Our managed AI services team can build and run this for you. Call 1300 EPIC IT to start.

Frequently asked questions

Which AI agent actions need human approval?

Anything that leaves the business, commits money, changes a record people rely on or cannot easily be undone should need human approval. For a small business in Australia, that usually means outbound emails, purchase orders, CRM and finance changes, and production settings. Payments, deletions, access changes and contracts should not be automated at all.

Is human approval of AI agents required by law in Australia?

Not for most private businesses today. The ASD's Information Security Manual includes AI agent controls, including one that stops retrieved content overriding human approval requirements, and the National AI Centre's guidance ties oversight to risk. They are the clearest Australian benchmark for AI agent human approval, and clients and insurers increasingly ask about them.

Who should approve AI agent actions in a small business?

The person who would approve the same action if a staff member did it. Finance approves spending, the office or operations manager approves customer messages and bookings, IT approves system changes, and the owner approves new commitments and changes to the agents' own permissions.

How do I stop an AI agent approving its own actions?

Enforce approval outside the model, in the harness or the systems it connects to. Do not give the agent tools for actions it should never take, require a named person to confirm the rest, and make sure nothing the agent reads, such as an email saying a request is approved, can bypass that step.

Want an approval matrix for your own AI agents?

Our Perth-based team will map what your agents can do today, sort every action into run, approve or never, and set up the approvals and logging to match. Call 1300 EPIC IT or book a free managed AI review.

Book a Free Managed AI Review

About the Author
Written by Moe Chizari, Managing Director of Epic IT, a managed IT, cyber security and AI partner for Australian mid-market businesses, with offices in Perth, Sydney and Brisbane. Moe brings 17 years across financial markets, treasury and technology, including five years at Bravura Solutions running enterprise software delivery and five years inside Group Treasury at Westpac and Macquarie leading APRA-regulated programmes (APS-117 IRRBB, APS-210 LCR & Capital Transformation). He holds a Bachelor of International Business from RMIT University, is a certified Project Management Professional (PMP), and an AFMA Diploma of Financial Markets graduate.

Keep reading