AI Consulting for Government

Three things, and they compound. First, you cannot choose your customers. A retailer with an AI system that fails 5% of the time has a service problem; an agency with a system that fails 5% of the time has failed 5% of people who had no alternative provider and no ability to walk away. Second, decisions are contestable by design — administrative law expects a person affected by a decision to be able to understand the basis for it and challenge it, which means "the model produced it" is not an answer to a review, an ombudsman or a court. Third, the legitimacy cost is asymmetric: a private company that gets it wrong loses customers, while an agency that gets it wrong damages public trust in government itself, and that damage is slow to repair and not confined to the agency responsible. Robodebt is the reference point every Australian public servant already has, and its central lesson was not technological — it was about what happens when an automated process at scale meets inadequate legal foundation and no meaningful human review. That lesson survives whatever the technology is.

The internal and the informational, which together represent an enormous amount of work. Answering routine public enquiries where the answer is published policy — bin days, permit requirements, opening hours, what form is needed for what. Triaging and routing inbound correspondence. Drafting for officer review. Searching your own policy and procedure corpus so staff can find the answer that already exists. Extracting data from forms and applications for an officer to then assess. Preparing report packs. Meeting minutes. None of that determines an outcome for a person. The line is bright and worth holding: the moment a model influences whether someone gets a permit, a payment, a service or an enforcement action, you have entered automated decision-making, and that carries transparency, contestability, review and record-keeping obligations that must be designed in from the start. Most agencies have years of the first category available and have not touched it.

Australian governments have developed assurance frameworks for AI use in the public sector, including a national framework for the assurance of AI in government and jurisdiction-specific instruments such as New South Wales’ AI assessment framework. They differ in detail but converge on the same demands: know where AI is being used, assess risk proportionately before deployment rather than after, be transparent with the public about it, keep a human accountable for decisions, and be able to explain and review what happened. Which framework binds you depends on your jurisdiction and portfolio, and they are updated — so the current applicable instrument is something to confirm with your own governance area rather than take from a consultant’s page, including this one. What we would add from practice: the frameworks are less onerous than people fear when applied at the start, and close to unworkable when retrofitted to something already deployed. The cost of assessing early is a fraction of the cost of unwinding late.

This is the part most often forgotten and it is genuinely important. Public sector records obligations — under the Archives Act at the Commonwealth level and equivalent state legislation such as Victoria’s Public Records Act — do not have an exception for automated processes. If an AI system contributes to the conduct of public business, the record of what it did is a public record, subject to the same creation, capture, retention and disposal duties as anything else. That has concrete design consequences: prompts and outputs may need to be captured, the version of the system that produced a given output may need to be identifiable, and the whole trail needs to be discoverable under FOI. A vendor whose architecture makes that impossible has disqualified themselves regardless of how good the model is. It is also worth thinking about before deployment because retrofitting record capture to a system never designed for it is one of the more expensive things we see agencies attempt.

It works, but the sequencing matters and it is where councils in particular get stuck. Agencies must procure through the applicable rules and panels, with value-for-money and probity obligations attached, and that is not a barrier to good AI work — it is a constraint on how it is scoped. The practical difficulty we see is that agencies often go to market before they have diagnosed the problem, which produces a specification written from a vendor’s marketing rather than the agency’s need, and then a panel of vendors who all answer that specification faithfully. Separating the diagnosis from the delivery is usually the fix: a small, properly scoped piece of independent assessment work produces a specification that describes your actual problem, and the subsequent procurement gets you something you can use. We are a small independent firm, not a panel behemoth, and there are engagements we are the wrong size for — we will say so rather than shape ourselves to fit a procurement we should not be in.

A free initial consultation to start — a conversation about where the administrative load actually sits and what your governance posture permits, not a demonstration. If there is a real opportunity, the next step is generally the AI Opportunity Audit at around $3,000: a written map of your workflows, systems and data, the privacy, records and assurance constraints attached to each, and a ranked shortlist with realistic effort and payback, separated explicitly into work that does and does not affect outcomes for a person. That separation is the most valuable page in it for a public sector body. You keep the audit either way, including the recommendations to do nothing. To be explicit about our limits: we are AI consultants. We do not certify your compliance with any assurance framework, privacy legislation or records duty, and we would be suspicious of anyone who offered to.