"Transformation" is a word that has been asked to do a great deal of work in government technology, usually just before a programme that does not transform anything. Being specific about what AI changes in the operations of an agency matters, because the answer is narrower than the market suggests and larger than the sceptics allow.
The three patterns that repeat
Across the engagements we have run, useful AI work in government tends to fall into one of three shapes.
Finding things in a corpus nobody can hold in their head
Agencies sit on enormous volumes of unstructured material: guidance, precedent decisions, correspondence, submissions, legislation and its explanatory material, and decades of internal advice. Institutional knowledge about what is in there tends to live with a handful of long-serving staff.
Retrieval systems that answer questions with citations back to source are the most reliably valuable AI application we deploy. They are also the least risky, because the human reads the source. The productivity effect is real and it compounds: new staff become useful faster, and the dependency on a few individuals reduces.
Triage and routing at volume
Where an agency receives a high volume of inbound work such as correspondence, applications, complaints and requests, a large share of the effort goes into classification and routing rather than into resolution. Getting an item to the right person with the relevant context attached suits current capability well, particularly when the system is designed to escalate anything it is uncertain about rather than to guess.
The important design decision here is the confidence threshold, and the important cultural decision is that the system is measured on the cost of its errors rather than on the proportion of work it handles automatically.
Drafting a first version of something a human will finish
Summarising a long submission, drafting a routine response, preparing a briefing pack from source material. The output is a starting point, the human is the author, and the accountability does not move. This works well and is widely useful, provided the system is honest about its sources and the organisation is honest that the human is still responsible for the content.
Where it does not work, and where it should not be used
One pattern keeps being sold to agencies and deserves a flat answer, which is AI as a replacement for a decision that has legal effect on a person. Even where the accuracy might be adequate, the requirements around review, reasons, transparency and appeal are difficult to satisfy and the reputational exposure is severe. The public conversation about automated decision-making in Australian government over 2023 and 2024 has made the boundaries clearer, and the sensible position is that AI assists the decision-maker with analysis and does not make the determination.
Equally, if a process is deterministic and the rule is knowable, a rules engine is a better answer than a language model. It is cheaper, faster, testable, and defensible in a way a probabilistic system is not. A large share of the "AI opportunities" we assess turn out to be this.
How to tell the difference before spending money
Three questions filter most of it.
Can you define what a correct output looks like, specifically enough to label a few hundred examples? If not, you cannot evaluate the system, which means you cannot deploy it responsibly or maintain it. Start there rather than with the build.
Does the value survive contact with the workflow? A system that saves ten minutes of drafting but adds fifteen minutes of verification has made things worse. Trace the whole path, including the checking.
Is the data actually available, in the form you assumed? More AI initiatives are killed by data access, quality and ownership than by anything to do with models. Confirm this before design, not during build.
The realistic picture
The agencies making real progress are rarely the ones with the most ambitious strategies. They picked two or three unglamorous problems, built the evaluation before the system, involved their assurance functions early, and then re-used the resulting platform for the next use case.
That approach looks slower for about six months and is dramatically faster after eighteen. Transformation, if the word means anything here, is what accumulates from that work rather than something a programme title can confer.
- Strategy
- Government
- Operations