Australian governments at every level have published AI strategies. The federal AI in Government taskforce has produced reports. State governments have announced pilots. The 2026 question is which of this has actually produced operational AI in government, versus PowerPoint AI.

The honest answer is that operational AI in Australian government is real but narrower than the strategies imply. The areas where it’s working in production: document classification and routing in several agencies, FOI redaction support, transcription of administrative recordings, customer-facing chatbots for routine queries, and some translation work in service delivery agencies.

The areas where pilots have struggled to move past pilot: predictive analytics for case management (consistently flagged as unacceptable risk by integrity bodies after several past controversies), generative AI for drafting policy documents (works technically, fails on the political risk dimension), and most decision-support tooling that touches benefit determination or compliance enforcement.

The Robodebt aftermath continues to shape the risk appetite for any AI deployment that touches benefit determination, eligibility, or compliance enforcement. That’s appropriate. The political and human cost of automating decisions about citizens’ entitlements went so badly that any AI proposal in adjacent territory now faces a far higher bar than equivalent commercial deployments. This is one of the genuinely good outcomes of an otherwise terrible episode.

For agencies looking at AI adoption in 2026, the working pattern has stabilised. Start with internal-facing tools where the failure mode is internal inefficiency, not citizen harm. Build assurance and audit capability before scaling to citizen-facing applications. Treat AI as a category that needs explicit governance, not a technology to deploy quickly.

The integrity backdrop is mature. The Office of the Information Commissioner, integrity bodies, and the ANAO have all developed substantial AI auditing capability. Agencies that deploy AI without engaging early with these bodies are setting themselves up for the same trajectory as Robodebt.

The vendor ecosystem in 2026 has consolidated around a small group of incumbents (Microsoft, AWS, Salesforce, the major consultancies) plus a slowly growing set of Australian specialist firms. The boutique market has been harder than expected because the procurement gravity continues to favour established suppliers.

For the next phase, the test will be whether agencies can build genuine AI delivery capability in-house, alongside vendor relationships, rather than relying entirely on outsourced delivery. The DTA’s recent direction suggests federal government is taking this seriously. Whether state governments follow remains to be seen.