Which Tasks Should I Never Hand to AI? The Honest List for Small Businesses

Never hand AI the final say. That is the whole principle, and it covers seven specific jobs: tax and BAS lodgments, contracts, hiring and firing, payments, sensitive customer data, regulated advice, and anything you will not personally check. We say this as a government-funded service whose entire purpose is helping Australian businesses adopt AI. It is precisely because the technology is worth using that the honest boundaries matter.

The seven

1. Final tax and BAS lodgments. AI can prepare, reconcile and draft. But when a lodgment is wrong, the ATO's penalties land on you, not the software. The Tax Practitioners Board has told even registered agents they must review every AI output and cannot transfer professional responsibility to a tool. If the professionals cannot delegate that judgment, neither should you.

2. Contracts you have not read. AI is useful for flagging odd clauses and translating legalese into plain English. It is dangerous as a drafter you trust blindly, because it invents plausible-sounding terms and, notoriously, fictional case law. Courts in Australia and overseas have already dealt with lawyers filing AI-invented citations. Use it to understand documents, not to sign them.

3. Hiring and firing decisions. Anti-discrimination law does not accept "the algorithm decided" as a defence. AI screening tools can carry bias from the data they learned from, and unfair dismissal claims turn on the process. Let AI draft the job ad and sort applications into piles if you like. A human makes every decision about a person.

4. Payments and bank details. No AI tool should be able to move your money or change a payee's account details without a human sign-off. This is less about AI failing and more about criminals exploiting automated trust. Payment redirection scams cost Australians $166.8 million last year, and the defence is a human with a phone.

5. Customer data into free tools. Free tiers of some AI products may use whatever you type to train their models. Customer records, health details, financials and anything identifiable belong only in paid business-grade tools with clear data commitments, and your privacy policy should reflect what goes where.

6. Regulated advice. Financial advice, legal advice, health claims in your marketing. These fields are licensed for a reason, and AI-generated advice published under your business name is your liability. AI can research; a qualified human advises.

7. Anything you will not verify. Quotes built on AI-estimated measurements. Prices scraped and summarised. Technical specs in a tender response. AI states wrong things with total confidence, so the rule is simple: if you would not stake the job on it, check it before it leaves the building.

What AI should absolutely be doing

The flip side deserves equal honesty. Drafting emails and marketing copy, summarising documents and meetings, reconciling bank transactions, answering common customer questions, turning messy notes into proposals: AI does these well, cheaply, today, and businesses that refuse the help are giving competitors a head start. If you have not started, our guide on where to start with AI in your business is written for exactly that moment.

The rule of thumb

AI drafts, humans decide. If a mistake would cost real money, a licence, or a customer relationship, a person signs off. Apply that one test and you can adopt AI aggressively everywhere else with a clear conscience. If you want help drawing the line for your specific business, that conversation is what our free consultations are for. Book one here.

Frequently asked questions

Is it safe to give AI confidential information if I pay for it?

Safer, not automatically safe. Business-grade paid tools typically commit to not training on your data, but read the data terms, and keep genuinely sensitive records in tools with explicit Australian privacy compliance.

How do I check whether an AI answer is right?

Ask for sources and verify the ones that matter against the original. For numbers, recalculate independently. For legal, tax or compliance questions, treat AI output as a starting point for a conversation with a professional, never the final word.

Can AI legally make decisions about employees?

AI can support the process, but decisions about hiring, discipline and dismissal need human judgment. Anti-discrimination and unfair dismissal law hold the employer responsible regardless of what software was involved.

What is the safest first task to give AI?

Something internal, low-stakes and easy to check: summarising documents, drafting routine emails, or turning meeting notes into action lists. You learn the tool's strengths and failure patterns before anything customer-facing depends on it.