.jpg?w=1024)
More than half of Australian enterprises say their technology architecture is too rigid, too costly and too slow to support the AI systems they are now trying to build.
Research covering 200 local organizations, conducted by IDC and MongoDB, found that 58% described their existing infrastructure that way — a verdict on the systems businesses were running well before agentic AI arrived.
The finding lands as a global reckoning takes hold over what agentic AI actually demands from enterprise IT. Systems built to handle predictable, human-triggered software requests are now being asked to support AI agents that autonomously call models, query databases, and coordinate with other agents across multiple steps. That’s a job description most corporate technology estates were never designed to fill.
Google Cloud puts a number on the global gap
Google Cloud's 2026 State of AI Infrastructure report, published July 8 based on a survey of 1,402 IT leaders worldwide, found that 83% of organizations need infrastructure upgrades before agentic AI can run in production rather than in a pilot. The report traces the gap to rising inference costs, operational complexity, governance demands, and power constraints as agents take on multistep work across business systems.
52% of respondents now run hybrid multicloud architectures, and 48% prioritize infrastructure with strict data-residency controls. Both are signs that a single cloud region or a single vendor's stack is no longer treated as sufficient. Edge deployment has also become a mainstream requirement rather than a niche one: 90% called it important to their AI plans, with 72% rating it extremely or very important.
Cost is the clearest symptom of the mismatch. 62% of respondents reported a significant "inference tax" — data-egress fees, excess storage, idle specialized hardware — that has nothing to do with model licensing and everything to do with infrastructure built for a different workload. Another 81% cited operational complexity itself as a hidden cost of scaling agents.
Why old architecture and new agents don't fit
The core conflict is structural, not cosmetic. Traditional enterprise systems were designed around discrete, stateless transactions: a user submits a request, a system responds, and the interaction ends.
Agentic AI needs systems that retain context across long-running workflows, recover cleanly from failure partway through a task, and track what an autonomous process did and why, often without a human confirming each step.
That mismatch is showing up in Australian governance data too. Deloitte's State of AI in the Enterprise research found only 12% of Australian leaders say generative AI is already transforming their business, compared with 25% globally. That plays out even as 57% of Australian organizations report deploying robotics, while interest in sovereign, locally accountable AI infrastructure continues to climb.
Deloitte found 72% of Australian companies now weigh a vendor's country of origin in AI decisions, and more than 80% treat sovereignty as a strategic planning priority.
Power and data placement decisions are converging as a result. The federal government's March 2026 expectations for data centers and AI infrastructure developers explicitly link new "AI factory" builds to national security and data sovereignty, not just capacity.
Google's global survey found a similar pattern on the buyer side: 91% of IT leaders now factor power consumption into hardware selection, and 61% rate it as a primary or significant factor.
Governance is the other half of the readiness gap
Security and governance are where the old-versus-new conflict is most acute. Google's survey found that 79% of organizations identify security, governance, and AI operations as their leading infrastructure challenge. This is because agents that can read email, query databases or alter records need permissions scoped to a single task, not the standing access a human employee might hold.
Australian organizations are already living with this gap. Proofpoint's 2026 AI and Human Risk Landscape report found 87% of Australian organizations have deployed AI assistants beyond pilot. Still, 52% describe their AI security posture as catching up, inconsistent, or reactive.
Only 33% say they are fully prepared to investigate an AI- or agent-related incident, and 39% have already had a suspicious or confirmed AI-related security event. Agentic ransomware capable of chaining multiple stages of an attack together adds urgency to that gap: identity and access controls built for human logins don't translate cleanly to software that operates autonomously.
What Australian IT leaders should watch
Google's own recommendations happen to match products it sells. That is reason enough for local IT leaders to weigh competing architectures on portability and cost rather than take any one vendor's roadmap at face value.
For Australian enterprises, the immediate task isn't choosing which AI model to deploy. It's deciding whether aging systems, siloed data lakes, and manual, human-scoped governance processes can be re-engineered fast enough to support agents that never stop running. An alternative is finding whether some of them need to be replaced outright before the next production rollout, not after it stalls.