Why a question list
Most AI agent proposals read alike. They promise speed and scale, and few say who is accountable or what happens when the agent is wrong. Australian buyers are already being given better tools than the pitch deck. The Digital Transformation Agency's policy for the responsible use of AI in government, version 2.0 from December 2025, asks agencies to name accountable officials and an accountable owner for each use case. The Department of Industry's Guidance for AI Adoption, published in October 2025, lists six practices, including an accountable person, risk management, testing and monitoring, and human control. We have read the practice wording through law-firm summaries, so check the source.
The ten questions below are built from those expectations. We answer each as a vendor, because a list that exempts its author is not worth much. Where our answer is standing policy, we say so. Where we cannot yet show evidence, we say that too.
The questions
1. Who is the named, accountable person?
Why it matters. When an agent makes an error, a person has to own the outcome. Government policy already expects named owners, and your own governance will too.
A good answer. A name and role for each engagement, who has authority to sign deliverables and who you can call. Not a team alias.
Our answer. Standing policy: every engagement has a named human lead who is accountable for the work and signs it off.
2. Is AI use disclosed on every deliverable?
Why it matters. You cannot assess what you cannot see. Reviewers need to know which parts an agent produced.
A good answer. A statement on each deliverable of what AI was used for, and which person reviewed it.
Our answer. Standing policy: AI use is disclosed on every deliverable.
3. Will you get my written approval before using AI on my work?
Why it matters. The DTA's AI model clauses, as reported by Holding Redlich and the Australian Government Solicitor, require a Commonwealth supplier planning to use AI to deliver services to notify the buyer first, with details of the system and its functionality, and obtain prior written approval. Even outside government, it is a sound default.
A good answer. Yes, in writing, before work starts, with the tools and models described.
Our answer. Standing policy: we seek your prior written approval before we use AI on your work.
4. Which models and tools do you use, and where does my data go?
Why it matters. Data handling cannot be judged in general terms. The same AGS note reports that Commonwealth protective security directions in 2025 treated certain products as an unacceptable risk, so agencies need to know exactly which models are in play.
A good answer. An answer for each component: model provider, hosting region, who can access it, what is retained.
Our answer. We will tell you the models and tools for your engagement as part of the approval in question 3. We have not yet published a per-component data handling document, and we would rather say that than offer a general assurance.
5. What will you not give an agent?
Why it matters. A vendor who has no limits has not thought about risk. Joint guidance from the Australian Signals Directorate and partner agencies, as reported by New Zealand's NCSC, advises limiting agents to low-risk, non-sensitive tasks until practices mature.
A good answer. A written list of exclusions, with reasons.
Our answer. We will not hand decisions about people, final legal, financial or regulatory sign-off, unreviewed production changes, or unsupervised customer communication to an agent. We use sensitive or personal information only where scoped and approved. Our reasoning is in When not to use an AI agent.
6. How is the work reviewed, and by whom?
Why it matters. Agents produce plausible errors. Review is the control, so its quality is the quality of the service.
A good answer. Described review steps, named reviewers, tests required before handover, sources cited so claims can be checked. Be wary of any promise of "zero errors".
Our answer. Work is reviewed by the named lead before it reaches you. We do not claim agents never err. We have not yet published sample deliverables that show our review in practice, and we can walk you through a live example under an agreement instead.
7. What is the fixed scope, and how will I know it is done?
Why it matters. Gartner, in a June 2025 forecast reported by CRN, expects more than 40 per cent of agentic AI projects to be cancelled by the end of 2027, citing cost, unclear value and weak risk controls. We have seen this only in secondary reporting, so treat it as indicative. Unclear scope is where cost and value problems begin.
A good answer. Written scope, written acceptance criteria, and a clear statement of what falls outside.
Our answer. Standing policy: fixed scope and stated acceptance criteria, agreed before work begins.
8. What stops the bill growing?
Why it matters. Agent work consumes compute and review time, and open-ended usage is open-ended cost.
A good answer. A cap that the vendor cannot exceed without your agreement, and pricing that shows its drivers.
Our answer. Standing policy: a hard spend cap on every engagement. We do not exceed it without your approval.
9. What do I get at the end, and can I take it elsewhere?
Why it matters. Third-party dependency is a live concern for regulated buyers. APRA's April 2026 letter to industry, as summarised by the Compliance Council, says governance, risk management, assurance and resilience are not keeping pace with AI adoption, and existing standards such as CPS 230 still apply.
A good answer. Everything delivered in standard formats you can open, run and maintain without the vendor.
Our answer. Standing policy: handover in standard formats. We do not make you dependent on a proprietary runtime to use what we deliver.
10. What independent assurance do you hold?
Why it matters. Self-description is cheap. Buyers in government and financial services reasonably look for independent evidence.
A good answer. A plain list of the certifications and assessments held, with scope and dates, and honesty about what is in progress.
Our answer. We do not currently hold or claim any independent assurance, certification or panel status. We would rather you start from that and test our controls on a small, fixed-scope piece of work than rely on a badge we do not have.
Using the list
Ask all ten of every vendor and compare the answers side by side. Treat a vague answer as a cost, and a candid "not yet" as better than a confident claim you cannot verify. If you are in a regulated sector, map the answers to your own third-party risk process before you sign.
How this article was written
Drafted with AI agents and reviewed by xShopper. Facts are linked to their sources; anything we could not confirm at a primary source is marked as such.