Sell to Government? Your AI Use Is Now a Contract Term

If your business sells to government — a council, a state department, a federal agency — the next contract you sign may ask a question you have not had to answer before: do you use AI to deliver this work, and can you prove you control it? Answer it badly and it costs you the bid. Australia's Digital Transformation Agency has rebuilt the rules for AI in government, and the obligations do not stop at the agency's door. They flow down the contract to you.

The DTA overhauled the Policy for the responsible use of AI in government — now at version 2.0 — and paired it with new procurement guidance and an AI impact assessment tool. The timeline has teeth. The first mandatory requirements bind agencies from June 2026. From December 2026, an agency must run a mandatory AI impact assessment before it deploys any in-scope AI system, put a formal approval process around each use case, and report AI incidents.

Alongside the policy, the DTA published a set of Model AI Clauses — contract terms an agency can drop into a supplier agreement, available through the government's BuyICT panel. They are optional and tailorable, not a blanket law. But an agency that now has to account for every AI use case internally cannot do that while its suppliers quietly run AI in delivery. So the demand flows down, in the one place a supplier cannot ignore it: the contract.

Read the Model AI Clauses through a supplier's eyes — as law firm King & Wood Mallesons did in its analysis — and four obligations stand out. You need written approval before using an AI system to deliver the goods or services. You have to run quality-assurance checks on what the AI produces. You must keep records of which AI systems you used, how, and what data they collected, processed and stored. And you cannot use banned systems — the clauses point to tools the government has ruled out, such as DeepSeek, which Home Affairs prohibited across government systems.

The trap is the word "embedded". This does not only catch a purpose-built AI product you sell. It catches the AI features baked into the software you already run — the assistant in your document tools, the summariser in your practice software, the automation in your scheduling app. If it touches the government's work, it is in scope. Most suppliers have never inventoried that, which is exactly why a tender question about it is so easy to fumble.

Want to know what this means for your business? We'll map it in one call.

Government is a large, steady customer for exactly the SMEs we work with. Councils put trades on maintenance panels. Departments hire accountants, engineers, consultants and lawyers by the hundred. If you are on one of those panels, or want to be, this is now part of the price of entry — and it is an admin leverage problem, not a technology one. The firms that lose here will not be the ones using AI. They will be the ones who cannot say, on paper, where they use it.

Think about how a bid actually gets scored. A procurement team following the DTA's guidance will ask for evidence: where AI touches your delivery, what data it sees, whether that data stays onshore — Australian data residency matters to government — and how you check its outputs. A rival who can hand that over in a page looks like the safer contract. You, reconstructing it after the request for tender has closed, do not. The cost of getting caught flat-footed is not a fine. It is the work walking to someone more prepared.

Worth doing before a tender forces it: map where AI already touches your delivery, including the features you did not choose but your software switched on for you.

The DTA's policy, procurement guidance and Model AI Clauses are public and free to read. Use them. But notice who they are written for: the agency buyer. They tell the government what to demand from you. They do not build your side of the conversation — the inventory of where AI touches your delivery, the data-residency position, the QA process you can point to, the one-page disclosure a buyer can accept without a follow-up meeting. That is the half no template does for you, and it is the half a tender is actually scored on.

Build your AI disclosure position now, before a tender asks. One page: every place AI touches your government-facing delivery, including embedded features; what personal or sensitive data each system sees and where it sits; how you check outputs before they reach the client; and the tools you will not use. It is a half-day exercise done calmly this quarter, or a scramble mid-bid later. And it is not wasted if you never bid — it is the same map you will need when you have to tell your own customers where AI makes decisions about them from December, and the same groundwork behind Australia's shift from voluntary AI guidance to legislated standards.

Key takeaways

Common questions

Related notes

Bidding for government work and not sure what your AI answer is?

Book a call and we will inventory where AI touches your delivery — including the embedded features you did not switch on yourself — sort out your data-residency and QA position, and turn it into the one-page disclosure a procurement team can accept without a second meeting.