How do I set up a new business with AI from day one?

The advantage you have for about a year

Most of what makes AI expensive in an established business is integration: connecting new capability to systems chosen years ago for different reasons, holding data in incompatible shapes. A new business has none of that, and the window in which it has none of that is roughly the first year.

Spending that window well mostly means making boring decisions correctly. Choose tools that have documented APIs, even if you have no immediate plan to use them. Keep one authoritative record of customers and one of financial transactions, rather than letting each tool become its own partial truth. Those two choices determine whether automation in year two is a project or an afternoon.

What a very small team can genuinely run with AI

The honest list is narrower than the marketing suggests and wider than most sceptics expect. These work because output is reviewable before it matters.

• First-draft written material — proposals, service descriptions, standard responses, website copy. A person still edits, but the blank page problem disappears.
• Inbound triage. Classifying and routing enquiries, drafting an acknowledgement, extracting what was asked. The customer still gets a human answer; they get it faster.
• Document handling. Pulling structured fields out of invoices, forms and statements instead of retyping them, with a person checking the extraction.
• Research and summarising before a decision — competitor positions, supplier comparisons, condensing long documents to the parts that bear on a choice.
• Internal reporting. Assembling the weekly or monthly numbers from systems that already hold them.

What still needs a human, and what needs a licensed one

Some functions are not slow because nobody has automated them. They are slow because they carry consequence. Choosing a company structure, registering for the right taxes, understanding which licences your activity requires, and signing contracts all sit in this category. An AI tool will produce a fluent and plausible answer about Australian business structures, and being wrong about that costs considerably more than the advice would have.

Use AI to prepare for those conversations — to arrive understanding the options and with the right questions — and use a registered tax agent, an accountant and a lawyer to make the decisions. NTDG is not a registered tax agent and does not hold an AFSL or ACL; nothing here is tax, legal or financial advice. For registration and obligations, business.gov.au and the ATO are the authoritative starting points.

Three decisions that are expensive to reverse

• Where the customer record lives. Every tool will offer to be the system of record. Pick one deliberately and make the others read from it. Businesses that skip this spend year three reconciling four partial customer lists, and no amount of AI fixes contradictory source data.
• How you handle personal information, decided before you collect any. Retrofitting a privacy position onto data already gathered under no stated purpose is genuinely difficult. Decide what you collect, why, where it is stored and who it is disclosed to — including AI vendors — while the answer is still "nothing yet".
• Whether a human is in the loop on customer-facing output, and where. This is easy to set and hard to tighten later, because tightening it means telling customers the service got slower.

The failure mode specific to starting this way

A business that automates before it has customers automates its assumptions. You build a quoting flow for the work you imagined selling, a triage system for enquiries you predicted, and both encode guesses. Then real customers arrive wanting something adjacent, and the automation is now an argument for serving the market you invented.

The protection is sequencing, not restraint. Do the work manually with AI assistance until you have seen enough real volume to know what the process actually is. Automate the parts that have stopped changing. In a new business that is usually month six, not week two — and the thing being automated by then is rarely what you would have built on day one.