Every Australian business deploying AI in 2026 hits the same fork in the road: where should the AI actually run. Microsoft Azure for Copilot. AWS for Anthropic’s Claude. Google Cloud for Gemini. The honest answer is that the question changed in the last twelve months, and most of the advice still floating around is out of date.
Two things moved. The major AI models now live on specific clouds, so the platform you pick decides which AI you can deploy with full enterprise controls. And from 10 December 2026, the Privacy Act makes you accountable for how automated decisions are made, which turns a technical choice into a governance one. Here is how the three platforms actually compare for Australian businesses right now, where each one is strong, and how to make the call without getting it wrong.
Three years ago you chose a cloud on cost, regional presence, and the skills your team already had. That logic is gone. In 2026 the platform decides which AI models you get reliable, governed access to, because the model vendors have aligned with clouds.
Microsoft 365 Copilot only runs on Azure. Anthropic’s Claude has its primary enterprise home on AWS Bedrock. Google’s Gemini sits inside Google Cloud and Workspace. Pick a platform and you have, by default, picked a lane for your AI tooling. The catch for Australian businesses is that a model being available globally does not mean it is available onshore, and that gap is where most platform decisions go wrong.
Here is the reality almost nobody plans for: most Australian SMBs end up multi-cloud whether they intended to or not. Copilot for productivity on Azure, Claude for technical and agent work on AWS, the occasional Gemini workload on Google. Single-platform purity is harder to hold than it used to be, and chasing it usually costs more than it saves.
The default for the large number of Australian businesses already running on Microsoft 365. Azure is where Copilot, Azure OpenAI, and the Foundry model catalogue live, and the integration with M365, Entra ID, Microsoft Defender, and Purview is tighter than anything the other two can match. If your business runs on Microsoft and you want Copilot deployed properly with governance, Azure is the anchor.
The residency story is solid. Azure has Australia East (Sydney) and Australia Central (Canberra), and many AI services run in-region. The honest weakness is at the frontier. As at mid-2026, Australia East standard deployments still centre on GPT-4o, that model is on a retirement path, and the newest frontier models (the GPT-5 family and equivalent reasoning models) are not yet generally available onshore. If your requirement is the very latest model running on Australian soil under a standard deployment, Azure has gaps you need to check before you commit.
The default for custom AI applications and agent work, because AWS Bedrock gives you the broadest model selection through one API. As at 2026, Australian Bedrock customers can reach Anthropic’s Claude Sonnet 4.5 and 4.6, Haiku 4.5, and Opus 4.5 and 4.6 through Australian geographic cross-region inference, which keeps the request inside the Sydney and Melbourne regions. In February 2026 AWS also launched Project Mantle, bringing open-weight and third-party models from providers including DeepSeek, Mistral, Meta, and others to Sydney through an OpenAI-compatible endpoint.
AWS has Sydney and Melbourne regions, and the cross-region inference design routes traffic over the AWS network without crossing the public internet, with logging in your source region. If you are building AI agents, custom workflows, or industry-specific applications, this is the most capable place to do it. The trade-off is that AWS does not integrate with M365 the way Azure does, identity alignment with your existing Microsoft directory takes more work, and the learning curve is steeper for teams raised on the Microsoft stack.
The default if you already run Google Workspace, or for the specific workloads where Gemini is the right model. Google runs Vertex AI, now the Gemini Enterprise Agent Platform, in australia-southeast1 (Sydney) and australia-southeast2 (Melbourne), with GPU support and configurable data residency for at-rest storage and machine-learning processing. Vertex also carries Claude and a large model garden, so it is not a Gemini-only platform.
Google is strong on analytics integration through BigQuery and on specific AI workloads where its models lead, and for a Workspace business the integration story mirrors what Azure offers a Microsoft shop. The weaknesses are smaller Australian SMB market share, fewer local deployment partners, and the same frontier-lag pattern as Azure: the newest Gemini releases often land on global or multi-region endpoints first, with single-region Australian availability following later.
| Factor | Microsoft Azure | AWS | Google Cloud |
|---|---|---|---|
| Flagship AI | Copilot, Azure OpenAI (GPT family) | Claude via Bedrock, plus 100+ models | Gemini via Vertex |
| Australian regions | Sydney, Canberra | Sydney, Melbourne | Sydney, Melbourne |
| Onshore model strength (2026) | GPT-4o in-region; frontier models lag onshore | Claude 4.5/4.6 via AU cross-region inference | Gemini available; newest releases lag onshore |
| M365 / Workspace fit | Best for Microsoft 365 | Weak office-suite integration | Best for Google Workspace |
| Pricing model | Per-user licence (Copilot) + Azure consumption | Consumption, per token | Consumption, per token |
| Best for | Copilot and productivity AI | Custom AI agents and multi-model work | Workspace businesses, analytics-heavy AI |
This is the part that catches people, and it is worth slowing down for. An Australian region existing is not the same as your data being processed in Australia. What controls residency is the deployment or routing type, not the postcode of the data centre.
On AWS Bedrock you choose between three modes. In-Region keeps requests inside one nominated region. Geographic routing keeps them inside a defined geography, so the Australian profile routes only between Sydney and Melbourne. Global routing sends requests to any commercial region worldwide for the best price and throughput, with no residency guarantee. Pick Global to save money and your data can leave the country, which may quietly breach a residency commitment you made to a client.
Azure has the same shape of problem. A model available under a Global Standard deployment is not the same as one available under a regional or Data Zone deployment that keeps processing in Australia. The newest models often appear on global endpoints first, so the residency-safe option can be a generation behind the headline model. On Google Cloud, Vertex residency depends on using regional endpoints and, for stricter cases, Assured Workloads, because the global endpoint does not honour residency.
The practical implication: confirm AI-specific data residency for the exact workloads you intend to run, not the platform in general. This is the due-diligence step that separates a governed deployment from a compliance incident waiting to happen, and it is the work a capable provider should be doing for you. You can read the primary detail on each vendor’s own documentation, including Azure Foundry model and region availability and Google’s Vertex data residency commitments.
From 10 December 2026, the Privacy and Other Legislation Amendment Act 2024 introduces transparency obligations for automated decision-making. If your business uses personal information in a substantially automated decision that could significantly affect someone’s rights or interests, you must disclose it in your privacy policy: the data used, the kinds of decisions made, and where computer programs do the deciding. Serious breaches of the Privacy Act carry penalties up to $50 million.
This matters for platform choice because the obligation is yours, not the cloud vendor’s. The OAIC began its first privacy compliance sweeps in early 2026, so this is enforcement, not theory. Whichever platform you choose, you need the same governance scaffolding: data residency set to Australian processing, zero-retention configured so your data is not used to train models, access controls aligned to your identity provider, audit logging of AI usage, and a clear policy on which tools your team may use. The platforms differ in how they expose those controls, but the requirement does not change. The Australian frameworks worth mapping to are ISO 42001 and the NIST AI Risk Management Framework, both platform-agnostic. You can confirm the automated-decision obligations directly with the OAIC.
This is where most SMBs are exposed. Deploying AI tools is easy. Deploying them with governance that survives a regulator’s question is the part that needs expertise, and it is the core of our AI governance work.
The decision tree for an Australian business in 2026 is not complicated once you strip out the marketing.
If you run on Microsoft 365 and want Copilot deployed properly, Azure is your anchor. Most of your AI value comes from Copilot working against your existing M365 data, and that only happens on Azure. Other tools can sit alongside on other platforms. Our Azure cloud services team handles this pattern most weeks.
If you are building custom AI applications or want access to multiple models, AWS Bedrock belongs in the mix. The breadth of models through one API and the maturity of the deployment tooling make it the right home for custom agent work, even when your productivity AI lives on Azure. This is the territory our managed AI service is built for.
If you are deep in Google Workspace, Google Cloud and Gemini make sense as your primary AI platform. Switching to Microsoft 365 purely to use Copilot rarely pays off if Workspace already works for you.
And if you run specialised industry workloads, follow the vertical-specific AI tools to whichever platform they prefer, because aligning with that platform cuts integration overhead. For most businesses, the honest end state is Azure for productivity AI, AWS for custom agent work, and Google only where Workspace or a specific Gemini strength makes it the right call. The deployment quality matters more than the platform badge. A poorly governed Copilot rollout is worse than no Copilot, because it exposes data across the business that staff were never meant to see.
The pricing models are not comparable on a single number, which is why platform cost comparisons are usually misleading. Copilot is a per-user licence. Bedrock and Vertex are consumption-based, billed per token, so cost tracks usage rather than headcount.
As a directional guide, Microsoft 365 Copilot Business sits around AUD $27 to $31 per user per month on the promotional pricing running to 30 June 2026, and Copilot Enterprise around AUD $45 per user per month. Worth flagging now: Microsoft’s base Microsoft 365 plans rise between 5 and 33 percent from 1 July 2026, so if your renewal falls after that date your AI and licensing costs are both moving. A custom AI agent on Bedrock or Vertex can run anywhere from a couple of hundred to a few thousand dollars a month depending on workload. For a 50-staff business deploying a sensible mix, total AI platform spend, governance included, typically lands in the tens of thousands per year rather than the small numbers vendors quote per seat.
Run an AI inventory. Find out which AI tools your team already uses, on which platforms, and whether any are processing personal information offshore. Most businesses discover shadow AI they did not know about, and that is the first thing a December 2026 privacy review will surface.
Confirm residency on the workloads that matter. For each AI workload, check the actual deployment or routing type, not just the region. Make sure the model you rely on is available under a configuration that keeps processing in Australia, and document it.
Map your platform choice to a governance framework before December. Talk to us about a free AI deployment review. We will assess your platform mix, residency, and readiness for the 10 December 2026 automated-decision obligations, and give you a written plan you can act on. Start with a conversation on 1300 EPIC IT or our IT strategy team.
There is no single best. For Australian businesses on Microsoft 365, Azure is the default because Copilot only runs there. For custom AI agents and multi-model work, AWS Bedrock has the broadest selection, including Claude running onshore via Australian cross-region inference. Google Cloud suits Workspace businesses and analytics-heavy workloads. Most businesses end up using more than one.
Yes. Most Australian businesses running serious AI use more than one provider, each suited to different work. The real question is governance: making sure data labelling, access controls, and audit logging are consistent across every platform. That consistency is the work a capable managed AI provider should handle for you.
Not automatically. Residency depends on the deployment or routing type, not just the region. On AWS, global routing can send data offshore while geographic routing keeps it in Sydney and Melbourne. On Azure and Google Cloud, the newest models often appear on global endpoints first, so the residency-safe option can be a generation behind. Confirm AI-specific residency for each workload.
From 10 December 2026, businesses must disclose in their privacy policy when personal information is used in substantially automated decisions that significantly affect individuals. The obligation sits with your business, not the cloud vendor, and serious Privacy Act breaches carry penalties up to $50 million. Reviewing your AI deployment and privacy policy now is the sensible move.
It varies by model. Microsoft 365 Copilot is a per-user licence, roughly AUD $27 to $45 per user per month depending on tier. AWS Bedrock and Google Vertex are consumption-based, so a custom agent might run from a few hundred to a few thousand dollars a month. Total spend for a 50-staff business with a sensible strategy usually lands in the tens of thousands per year, governance included. Note that Microsoft base plan prices rise from 1 July 2026.
Not strictly, but the governance is where most businesses come unstuck. Configuring data residency, zero-retention, access controls, and audit logging consistently across Azure, AWS, and Google takes expertise most internal teams do not have spare. A managed provider maps your deployment to a recognised framework so it holds up under regulatory scrutiny, rather than handing you a feature list.
