Which tasks should you hand to AI? A task-by-task guide for Australian SMBs

Most Australian SMBs I talk to have the same list in their heads: the invoice reminders nobody enjoys, the same customer questions every week, the CRM that's always behind. They know AI could take some of it. What they don't know is which tasks to hand over first, and how much to let go of each one.

That's a task-by-task decision, not a whole-of-business one. An invoice reminder and a legal sign-off are not the same job, so they shouldn't get the same amount of AI. I help Australian SMBs work through that list in plain English. Not legal advice.

Which tasks should a small business hand to AI first?

Start with tasks that happen often, follow a pattern, and can be undone if something goes wrong. Leave rare, high-judgement or irreversible work with a person.

Every task gets a level, from 0 (a person does it all) to 4 (the AI runs it unattended). In between are Assist (the AI drafts, a person sends), Approve (the AI does the work, a person approves it) and Audit (the AI acts, a person reviews a sample). That's the Autonomy Spectrum behind how I approach Human-Led AI, and that page explains each level properly. This post is the practical part: which tasks go where.

Two rules sit under everything below:

• A task earns its level. It moves up when your own work shows it's ready, not when a product update adds an "autonomous" toggle.
• A named person owns the outcome, unless the task has genuinely earned full autonomy.

What tasks can AI do for a small business, and at what level?

Here's where common SMB tasks usually start. Anything customer-facing starts at Level 2 or lower. The Level 3 rows are where tasks end up after they've proven themselves, not where they start.

Which tasks should stay with a person?

Legal sign-offs and first contact with a major prospect, for a start. The same goes for anything where you can't write down the judgement involved, or where one mistake costs you a client. Leaving a task at Level 0 isn't being behind. It's knowing where the real risk sits. A surprising amount of my job is telling people a task doesn't need AI at all.

Which tasks can AI draft while a person sends?

This is Level 1, Assist: Microsoft 365 Copilot drafting emails and documents, summarising long threads, or turning meeting notes into a list of actions. The person still does the job. They just start from a draft instead of a blank page.

There are two ways this goes wrong. People send drafts without reading them. Or someone pastes client emails into a free consumer chatbot because it's quicker. The first is a habit to fix. The second is a data problem, and it's covered in the privacy section below.

Which tasks can AI complete if someone approves them first?

This is Level 2, Approve, and it's where most SMB task automation should live for a while. The AI does the whole job: invoice follow-ups, replies to common questions, CRM updates from emails and calls. Nothing external or irreversible goes out until a person approves it. The job changes from doing to checking.

In Microsoft land, this is where a lot of Copilot Studio work actually lands. Microsoft's guidance on "autonomous agents", and agent flows that pause to ask a person for input or approval, are mostly Level 2 patterns in practice. "Autonomous" in a product name doesn't make the task Level 4 in your business. Microsoft also talks about maturity levels for organisations adopting agents. That's a different scale: it measures how mature your organisation is, not how much a single task should be trusted to do.

There are two things to watch here. Approvals can become a rubber stamp. And the connectors behind the agent can end up with more access than the task needs. Before anything customer-facing goes live, run the pre-launch checks for Copilot Studio agents.

Which tasks can AI handle on its own with spot checks?

This is Level 3, Audit: booking confirmations, routine document requests and CRM logging. The AI acts within rules, flags anything unusual, and a person reviews a sample plus the exceptions.

A task only moves here once it has proven itself at Level 2, and once three things exist:

• a clear rule for what "unusual" means, with a hard stop on anything the AI shouldn't attempt
• a log good enough to answer "what did it send last Tuesday?" in about 30 seconds
• a named owner who can halt it

Without those, "Audit" is really Level 4 with better intentions.

What about customer emails? Can AI send those without approval? For most SMBs, not to start with. Customer-facing tasks stay at Level 2 or lower until the person approving them has stopped changing anything. Skip that and the first time you find out what "unusual" looks like is when a client tells you.

Are there any tasks AI can run completely unattended?

Yes, a few. Internal-only work, like syncing data between two systems you own, can earn Level 4 once it has a track record at the levels below. Where I draw the line is client-facing or irreversible work. Put that on autopilot without evidence and a small mistake becomes the same mistake at scale, in your name.

How do I check whether a task is ready to hand over to AI?

Run each task through this checklist before you change its level.

1. Is it customer-facing? Keep it at Level 2 or lower until the edits dry up.
2. Does it touch sensitive personal information? Use tenant-bound tools and more oversight.
3. Is the outcome reversible or irreversible?
4. What's the blast radius if it goes wrong overnight?
5. Can you answer "what did it send last Tuesday?" in about 30 seconds?
6. Has the person stopped changing the outputs? That's the signal it's ready to move up.
7. Is "unusual" defined and hard-stopped before you consider Level 3?
8. Is there a named owner who can halt it?
9. Does it have a significant effect on individuals, or raise Australian Privacy Principles (APP) obligations? Take extra care.
10. Is the change recorded: task, level, who decided, review date?

If you can't answer 5 or 8, the task isn't ready, whatever the vendor says.

Can we use Copilot or ChatGPT for tasks involving client data?

It depends on whether you're an APP entity, and on which tool you're using.

The OAIC's guidance on commercially available AI products comes down to three things: do your due diligence on the product, keep meaningful human oversight, and don't put personal information into publicly available generative AI tools. That last point is why "Level 1 in a free consumer chatbot" is the riskiest version of a low-risk task.

Many small businesses under the roughly $3 million turnover threshold aren't automatically APP entities. There are exceptions. If you are covered, APP 1 transparency requirements for automated decisions start on 10 December 2026, so it's worth knowing which tasks the AI is deciding on and which a person is. This is not legal advice. If you're unsure, ask a lawyer, not a blog post.

For broader risk language, the NIST AI Risk Management Framework recognises that AI systems run at varying levels of autonomy, and it treats defining human–AI roles as part of managing AI risk. It's useful grounding if someone asks where the idea comes from.

Where should an Australian SMB start?

Pick one task from your list that happens often and can be undone. Start it at Assist or Approve, name the person who owns it, and watch what they change. When the changes stop, you have the evidence to move it up. Then do the next task.

Before an AI assistant can take a task off your hands, it has to be allowed to see things. That's usually when a business finds out what's actually exposed: the shared folder everyone can open, or the contractor who still has a login (I wrote about leftover guest access). AI doesn't create those problems. It finds them.

When I do a Security Health Check for an Australian SMB, "what can the AI see, and who owns what it does?" is now one of the first questions I ask. If you want help sorting your task list into what to hand over and what to keep, without the big-firm theatre, get in touch.