Australia's Federal Government Uses Six Paid AI Platforms

Six platforms, one government, no single strategy

New reporting on Australian federal government technology spending has surfaced a fairly striking picture. Across the Australian Public Service, at least six different paid AI platforms are currently in active use: Microsoft M365 Copilot, Google Gemini and Notebook, Anthropic's Claude, GitHub Copilot, Figma Make and Otter.ai.

The scale varies enormously between them. Microsoft Copilot dominates, with roughly 29,000 paid licences across 67 responding entities, driven heavily by the Department of Foreign Affairs and Trade with more than 7,000 licences and the Department of Health, Disability and Ageing with more than 5,000. Google's Gemini and Notebook tools sit at 652 licences, concentrated at the Attorney-General's Department. Claude's federal footprint is tiny by comparison, described in the reporting as multiple small deployments, 9 licences at the Digital Transformation Agency, 5 at the Parliamentary Budget Office, 5 at the Australian Fisheries Management Authority. GitHub Copilot sits at 164 combined licences across a handful of agencies, and Figma Make and Otter.ai are smaller still, in single digit to low double digit deployments.

What this actually shows about government AI adoption

The picture that emerges is not one of the federal government executing a single coordinated AI strategy. It is dozens of individual departments and agencies making their own separate purchasing decisions, at wildly different scales, often described in the reporting as limited trials rather than full rollouts.

There is a genuine, unglamorous lesson in that pattern, one that applies well beyond government. Even an organisation with the resources of the Commonwealth is adopting AI unevenly, department by department, tool by tool, rather than picking one platform and standardising overnight. If that is how a national government is actually doing it, in real deployments rather than press release plans, it is a reasonable indication of how most organisations end up adopting AI in practice, messily and incrementally, not through one clean top down decision.

The transparency gap worth naming

The reporting also surfaced a genuine weak spot. At least 2,000 additional licences across government are currently unlabelled in official disclosures, most likely additional Copilot deployments that simply have not been formally tracked yet. The article notes usage data is "often not in the AI transparency statements" departments are required to publish.

This is worth taking seriously rather than dismissing as a rounding error. If a government with formal transparency obligations and dedicated reporting requirements cannot yet produce a clean, complete count of its own AI tool usage, that says something honest about how hard this tracking problem actually is, not just for government, for any organisation of meaningful size. Knowing exactly what AI tools are active across a large organisation, who has access, and what they are actually being used for turns out to be harder than it sounds, even when someone is officially required to know.

Why the early trial finding matters more than the licence counts

One detail buried in the reporting is arguably the most useful part of the whole story. The Digital Transformation Agency's early Copilot trial reportedly struggled to prove real value, specifically because the tool lacked access to actual government data inside Microsoft's applications. The AI was capable. It just did not have anything meaningful to work with yet.

That is a pattern worth remembering whenever an organisation, government or private, evaluates an AI tool and comes away underwhelmed. The problem is very often not that the AI itself is weak. It is that the AI was never properly connected to the real information it needed to be useful, and a trial run under those conditions tells you very little about what the tool can actually do once it is set up properly.

What this means if you run a business, not a department

You do not need a government sized budget to take something useful from this.

Uneven, staged adoption is normal, not a sign you are behind. If your business has one team using an AI tool enthusiastically and another team not using anything yet, that is not dysfunction, it is roughly how the Australian Public Service itself is doing this at national scale.

Track what you are actually using, early, not eventually. The 2,000 unlabelled licences finding is a warning sign every business should take personally. If a government department with reporting obligations does not have a clean read on its own AI usage, an ordinary business will not either, unless it deliberately makes a habit of tracking it from day one.

A disappointing AI trial is often a data access problem, not a capability problem. Before concluding a tool is not good enough, check whether it was actually given access to the real information it needed to be useful. The DTA's own experience is the clearest evidence yet that the two problems get confused constantly.