Article: Why government AI pilots keep stalling

Authored by Ankur Mahanta – Principal Consultant at xAmplify

I’ve watched more AI pilots die in procurement than in production.

The technology usually worked fine. The agency was serious enough. But every conversation eventually hit the same wall: how do we deploy this in a way that actually meets our security and operating requirements? The platform worked one way. The agency needed it to work another way. And the conversation stalled.

After more than a decade delivering technology inside Australian Government environments, I understand why that question stops things cold. It’s not bureaucratic caution. It’s the right question.

Three reasons AI pilots keep stalling

Most agencies aren’t short on ambition or technical curiosity. The blockers are more structural: deployment inflexibility, governance gaps, and sovereignty requirements that the market hasn’t kept pace with.

Deployment inflexibility. Most agencies already operate in hybrid or multi-cloud environments with data residency agreements in place – that’s not the problem. The problem is that most AI platforms only work well in one environment. When a use case needs on-premise or air-gapped deployment – because of classification, sensitivity, or operational requirements – the options narrow fast.

Governance gaps. Pilots are built to demonstrate capability, not to answer the hard questions: who owns the model? What happens when it’s wrong? How does this interact with existing records obligations? These questions aren’t unreasonable, but most pilots aren’t designed with the answers in mind. So they stall before they become programs.

Sovereignty without good options. Agencies rightly require control over their data, models, and infrastructure. But when it comes to state-of-the-art AI, the market hasn’t offered many solutions that meet sovereign requirements without significant capability trade-offs. Most AI platforms were built for scale first and adapted for regulated environments later – audit trails, data residency controls, and access governance get bolted on rather than designed in. The result is that agencies either compromise on sovereignty or miss out on the capability. Neither is acceptable.

The partnership – and why it matters

xAmplify is an Australian technology delivery firm that has spent seven years helping government agencies and enterprise organisations get platforms into production – ServiceNow, automation programs, digital twins, AI roadmaps. The team combines deep technical expertise across AI, data engineering, and enterprise platforms with the delivery fundamentals that government requires: security clearances, procurement panels, and understanding how agencies actually make decisions. That combination of technical depth and delivery experience is what we bring.

H2O.ai is a global enterprise AI platform company, in market since 2012, covering both predictive and generative AI. Their platform has been adopted by more than 20,000 organisations globally, was named a Visionary in Gartner’s 2025 Magic Quadrant for Cloud AI Developer Services, and their AI agent holds the top position on the GAIA benchmark for real-world accuracy.

What makes this combination worth writing about is that each side builds on the other’s strengths. H2O.ai brings a proven enterprise AI platform. xAmplify brings deep delivery experience in Australian Government environments – the people, the process, and the local context. Together, that covers the full path from platform to production, which is where most government AI initiatives have been falling over.

Here’s how the partnership addresses each of the three problems.

How the partnership addresses these problems

Deployment agnostic. H2O.ai’s platform runs consistently across hyperscalers, private cloud, and on-premise infrastructure – including fully air-gapped environments. It’s the same platform regardless of where it’s deployed, which means agencies aren’t locked into a single environment or forced to start from scratch in a different one. For agencies already running in cloud, AI workloads deploy on the same infrastructure with proper governance. For those with air-gapped requirements, full local deployment is available now, not on a roadmap. For the many sitting in a hybrid middle, the platform works across their entire environment.

Governance by design. xAmplify designs for governance from the beginning rather than retrofitting it after the technology goes live. That means addressing who owns the outputs, how errors are escalated, and how the system is maintained – before deployment. Organisations can also fine-tune smaller, purpose-built language models on their own data, with audit trails, access controls, and explainability built in from day one. Better outcomes for specific use cases, without paying for capabilities you don’t need.

Sovereignty without compromise. H2O.ai’s governance, explainability, and audit controls are part of the base platform architecture, not added later to make it enterprise-friendly. Organisations retain meaningful control over the data, the models trained on it, and the infrastructure they run on – regardless of whether that sits on a hyperscaler, in a private cloud, or on-premise. Critically, this doesn’t come at the expense of capability. Agencies can access and fine-tune enterprise-grade models within sovereign constraints, rather than being forced to choose between the two.

The harder and more important work is still everything around the platform: internal capability uplift, change management, helping teams understand what these tools can and can’t do. The goal isn’t to create another consultant dependency. It’s to make AI capability something agencies own and can operate themselves.

Why this moment matters

Agencies are under genuine pressure to demonstrate AI outcomes – not just innovation intent. The sovereignty and security expectations are getting clearer and more specific, not softer. More organisations are realising they need AI that works across their environments, not technology that forces them into a single deployment model to participate.

I think this partnership is one of the more credible answers to that in the Australian market right now. Not because of the platform alone, or the delivery capability alone, but because both need to be present for this to work – and that combination has been harder to find than it should be.