How small business owners can get started with AI: a measured pilot

Choose one workflow, measure its current cost, and test whether assistance improves the whole task—not just the first draft. You do not need to code or buy an agent first. You need a process someone owns, examples of correct work, a safe test environment and a reason the improvement would matter.

This guide is for owners and managers of established small businesses. It helps you decide whether to configure an existing tool or commission a bounded pilot. It is not a coding tutorial, product ranking or promise that AI will improve your margins.

Before a pilot: set data and action boundaries before connecting systems

The ACSC's small-business AI guidance identifies data exposure, unreliable or manipulated outputs and supplier dependencies as risks. Check data handling, access settings and incident procedures; assign responsibility for output checks and ongoing review. A business-plan label alone does not settle these questions.

The OAIC's commercial-AI guidance recommends avoiding personal information, particularly sensitive information, in publicly available generative-AI tools. It addresses product due diligence, transparency and personal information in inputs and outputs. Begin with synthetic data and obtain appropriate advice before connecting real records; this guide is not a compliance assessment.

• List the systems and fields the pilot may access; begin with the least access required.
• Keep publication, payments and customer commitments behind named human approval.
• Define who changes prompts/configuration, how logs are protected and when data is deleted.
• Test manual fallback and account/access removal before depending on a supplier.

If you still need to choose the problem, use the business-workflow selection guide. Once the scope and boundaries are clear, the one-task assistant guide shows a specific implementation path. Begin every example here with fictional or expressly permitted data.

1. Choose a task with a clear way to check the answer

Describe an input, an output and an approver: “From approved product facts, draft a description for a staff member to check.” Avoid “automate marketing”; it hides too many decisions. First try a better template, clearer handover or fixed rule. If that solves the problem, you may not need AI.

Reject a first experiment if you cannot inspect its output, lack permission to use the inputs or cannot safely undo an action. A familiar task is not automatically low-risk: an incorrect price, recipient or health claim can still have consequences.

2. Record the baseline before choosing a solution

Observe a representative operating period. Record volume, case type, human handling time, corrections, waiting time and final outcome. Keep waiting time separate from paid effort: cutting a queue by a day is useful, but not a day of wage savings.

• Use one consistent unit: an approved description or correctly resolved enquiry, not one model response.
• Include difficult cases: missing inputs, unusual requests and existing escalations.
• Name an owner: who judges correctness, manages access and can stop the test?
• State the business reason: avoid overtime, reduce contractor spending, prevent errors or release capacity. These are different outcomes.

3. Decide whether to improve, configure or commission

There is no universal under-$500 budget or cheapest subscription that fits every pilot. Obtain prices for the actual seats, usage, integration and support. A custom build is not automatically better; an existing tool is not automatically safe. For commissioning decisions, see building AI for a real business problem.

4. Separate useful capacity from cash savings

Worked illustration, not an MLAI client result: suppose a business produces 60 descriptions a month. Each currently takes 20 human minutes. A proposed assisted process takes eight human minutes, including review, corrections and the average effort of manual fallbacks. These timings are assumptions to replace with measurements.

1. Effort released: 60 × (20 − 8) ÷ 60 = 12 hours/month.
2. Allow another two internal maintenance hours: 10 net hours/month.
3. At an illustrative loaded cost of A$45/hour, capacity is worth A$450. Subtract A$100 of additional running cash costs: A$350/month net capacity value.
4. If payroll and other spending stay unchanged, avoidable labour spend is zero. The cash change is minus A$100/month, not a A$350 profit increase.
5. If you can substantiate that half the net labour value replaces avoidable spending, the estimated cash change becomes A$225 − A$100 = A$125/month. An illustrative A$600 setup invoice takes 4.8 months to repay from that cash change, if sustained.

Six internal setup hours add A$270 of effort, taking total economic setup cost to A$870. That internal effort is not included in the cash-invoice payback. Do not count the same labour in both the invoice and internal hours. The example excludes additional revenue, tax effects, financing and seasonality.

If released capacity lets you sell more work, measure the additional contribution after delivery costs separately. Do not assume every released hour can be sold or value it at an unadjusted billing rate.

Complete the worksheet below, then choose Use my worksheet in a Studio brief to carry your estimates and non-confidential notes into the enquiry for review. That does not submit them or treat the example's figures as agreed requirements. You can also start a blank brief with the separate CTA.