Planning an AI Trial in Australia: Questions to Ask Before Signing a Contract

Planning an AI trial in Australia should begin with the decision you want to test and the terms under which the supplier will let you test it. A short demonstration can look convincing while leaving data handling, evaluation and exit arrangements unresolved. The contract is part of the trial design, not an administrative detail to read afterwards.

Use these steps to prepare questions before signing. The aim is a limited, reviewable experiment with clear responsibilities. Obtain appropriate legal, privacy or technical advice where the proposed use raises questions beyond your team’s expertise.

Step one: define one task and its boundaries

Describe the input, expected output and person using the result. Drafting an internal summary is different from automatically sending advice to customers. A trial should specify which activity is included and which decisions remain outside it.

Ask the supplier whether the product has been tested for that intended use and what evidence is available. The National AI Centre’s adoption guidance emphasises appropriate testing and human oversight.

Record the business problem before seeing another demonstration. This helps prevent a trial from expanding to match every feature the supplier happens to show.

Step two: identify exactly what information enters the system

List the proposed data, its source and who is authorised to provide it. Include documents, prompts, attachments, feedback and any connections to business systems. A product can receive information through integrations as well as through a visible chat box.

Ask whether the supplier or other parties can access inputs and outputs, where processing occurs and what retention applies. Clarify whether data may be used to improve or train models and whether relevant settings are contractually supported.

The OAIC’s commercial AI guidance discusses due diligence and privacy risks. Use it to assess your actual data flow rather than assuming a trial account is exempt from ordinary obligations.

Step three: choose evaluation material and success measures

Prepare a representative set of tasks before the trial begins. Include ordinary cases, awkward cases and situations where the system should decline or escalate. Synthetic or appropriately de-identified material may help some tests, but it still needs to represent the intended work.

Agree how errors will be identified and who can judge the result. A fluent answer is not evidence of accuracy. If reviewing every output takes substantial time, include that workload in the comparison.

  • What is the current way of completing the task?
  • Which errors matter most?
  • What evidence will establish whether an output is usable?
  • How much review and correction time is required?
  • Which results would stop the trial?
  • Who decides whether the next stage is justified?

Make success criteria specific without pretending that a small test proves performance in every future situation.

Step four: clarify human control and responsibility

Identify who approves access, checks outputs and handles incidents. Ask whether the system can act on external services and which actions can be disabled. A writing trial should not silently become permission to publish, contact customers or change records.

Check how staff can pause the service, remove an integration or revert to the existing process. Responsibilities should be operationally realistic rather than assigned to somebody who lacks time or access to carry them out.

The National AI Centre implementation guidance provides further detail about governance and supplier information. Use it to prepare questions proportionate to the trial’s consequences.

Step five: read the price and duration together

Ask what the trial fee covers, how usage is measured and whether spending can be capped. Include licences, implementation, integrations, support and the staff time required to evaluate the system.

Check when a free or discounted period ends and whether it converts into a paid commitment. Confirm cancellation, minimum terms and what happens if usage exceeds the expected amount. Do not assume removing users cancels the business’s agreement.

Discuss changes during the trial. A supplier may update the system, model or features, so the agreement should make clear how relevant changes are communicated and how the evaluation will remain interpretable.

Keep a record of the version, configuration and connections being tested. If a meaningful change occurs, note it beside the results rather than combining different conditions into one score. Ask who will provide the technical information needed for that record and whether access to it continues after the trial ends.

Step six: plan the exit before starting

Ask how information can be exported, returned or deleted, and what evidence of completion is available. Clarify ownership and permitted use of outputs, existing materials and trial results without relying on a broad marketing statement.

Keep the original workflow available while the experiment is assessed. Identify who will remove access and integrations at the end, and how any retained records will be protected.

Finish with a short written trial plan attached to the commercial understanding where appropriate. It should state the task, duration, data, evaluation, owner, cost limit and exit decision. A useful AI trial produces evidence for a next step, including the valid decision not to proceed.