The AI demo ends before the contract risk does

An impressive AI demonstration does not establish what a buyer may do with the service, what the supplier may do with buyer data or how acceptable performance will be demonstrated after delivery. Australian procurement teams should turn those questions into a clear specification and an evidence-based acceptance process before treating a successful demo as a purchasing decision.

There is a useful Australian reference point. Commonwealth AI model clauses published on 17 March 2025 address services involving supplier use of AI and bespoke AI systems. Australian Government Solicitor’s Legal Update 327 explains that the clauses require selection and tailoring and do not form a complete contract. They are not universal mandatory terms for every Australian private-sector purchase. 1

A demonstration is usually designed to make a capability visible in a short time. Procurement must establish whether that capability fits the buyer’s intended work under the conditions that will actually apply. A clean sample, an expert presenter and a temporary trial account may conceal the effort needed to operate the service reliably.

The practical takeaways are to define the permitted use, make data boundaries explicit and agree what evidence supports acceptance. Keep commercial, technical, privacy and legal questions connected. A vendor’s answer to one should not be treated as an answer to all the others.

Begin with the work the buyer expects to perform. Describe the users, information, decisions and consequences involved. Include what the service must not do. If an assistant may draft a response but may not send it, that boundary belongs in the requirements and the acceptance demonstration, not only in an informal conversation with the sales team.

This proposed question set is an original procurement aid, not model contractual wording.

For each material answer, distinguish a product feature from a supplier commitment. A button visible in a trial may not be included in the purchased tier. A presentation may describe a roadmap rather than a current capability. Ask for the version, scope and conditions behind the claim and have the appropriate specialist check the resulting agreement.

Do not let a broad assurance such as “your data are secure” stand in for an explanation of allowed uses. Security, retention and permission to use information for further training are different questions. The buyer needs to know which information enters the service, which organisations process it and which settings or contract terms govern those uses.

The Commonwealth model clauses offer one example of making this explicit: AGS describes restrictions on using buyer data beyond the contract and on ingesting it into AI systems unless the contract permits that use. Whether those terms are included and suitable for a particular purchase requires review. 1

Use approved, appropriately protected material during evaluation. A need to understand product behaviour does not itself authorise uploading customer records or confidential documents. Where representative information cannot safely be used, document the limitation and ask the technical and privacy reviewers how acceptance can be supported.

Agree the acceptance approach before the final presentation. Choose scenarios that represent the actual workflow, including ambiguous inputs, missing information and cases that should be referred to a person. Document who evaluates the output and which uncertainties require further work. Avoid a universal percentage target disconnected from the consequences of error.

Keep the test configuration and observed limitations with the decision. A pass for one language, document type or user group should not silently approve a broader deployment. The purchaser also needs a practical route for dealing with unresolved defects: an acceptance decision should say what remains open and who is responsible for the next action.

For Commonwealth agencies, the Digital Transformation Agency’s responsible-AI policy materials provide additional procurement context. Private organisations can learn from those materials while identifying their own applicable obligations and approval arrangements. 2

Consider a fictional buyer seeking an assistant that drafts supplier replies. The demonstrated product can also send messages. The intended purchase permits drafting only. The following is a proposed acceptance record, not a report of completed testing or suggested legal wording.

The resulting purchasing decision is to resolve those four items before accepting the service for the stated workflow. Procurement assigns the contractual questions to its commercial and legal reviewers; the technical owner demonstrates the actual permissions. A supplier presentation remains supporting material rather than acceptance evidence.

If the supplier cannot disable sending in the purchased configuration, the buyer must change the design, select another configuration or reconsider the purchase. Describing the feature as optional does not remove the capability. These conditions apply to this invented purchase; a different use requires its own acceptance decision.

Instead of asking only whether the tool worked, ask whether the proposed purchase includes the capabilities, permissions and evidence needed for the intended use. If the answer remains uncertain, identify the missing decision rather than assigning a reassuring overall label.

Related work on supplier evidence and its coverage can help teams organise that discussion. That article uses a UK and Ireland cyber-renewal context; it does not establish Australian procurement obligations. The common discipline is to keep every assurance tied to the service and scope it actually covers.

AI assisted research and drafting. General procurement information, not contract drafting or legal advice. Obtain competent technical, privacy and legal review for the actual purchase.

1. Australian Government Solicitor, Legal Update 327 — 4 April 2025, explaining the Commonwealth AI model clauses published 17 March 2025. Checked 9 October 2026.

2. Digital Transformation Agency, Policy for the responsible use of AI in government: preparedness and operations — current version 2.0 page checked 9 October 2026; government-policy scope, not a blanket private-sector mandate.

This article is general information. Verify current requirements and seek qualified advice for your specific circumstances.

Spotted an issue? Contact our editorial desk.