A Microsoft Copilot Alternative Built for Australian Data Control

Microsoft 365 Copilot runs inside the Microsoft 365 service boundary and calls Azure OpenAI models hosted by Microsoft rather than sending data to OpenAI. Microsoft states that Copilot prompts, responses and data accessed through Microsoft Graph are not used to train the foundation models, and that is a meaningful, contractual commitment we take at face value. Residency is the harder question. Your Exchange, SharePoint and OneDrive data at rest can be committed to the Australian geo, and the Advanced Data Residency add-on extends that across more workloads for an additional per-user fee. Copilot inference is a different matter: Microsoft has published a formal data boundary commitment for the EU, and there is no equivalent Australian data boundary. Microsoft's Product Terms and Copilot data-protection documentation are the authority here and they change regularly, so verify them at each renewal rather than relying on any vendor's summary, including this one.

Yes, and this surprises people who assume Graph access is a Microsoft-only privilege. A private LLM connects through a registered application in Microsoft Entra ID using the Microsoft Graph API, with delta queries to keep the index current as documents change. The critical requirement is permission trimming: retrieval must honour the SharePoint and OneDrive access control lists that already exist, so a user only ever gets answers grounded in content they could open themselves. We implement this at the retrieval layer rather than the response layer, meaning restricted content is never placed in the context window in the first place. Purview sensitivity labels are read as metadata and can be used to exclude classifications entirely. The practical difference is that a private LLM indexes Graph content and your line-of-business systems together, so it can answer questions that span both.

Microsoft 365 Copilot lists at roughly A$45 per user per month on an annual commitment, which is about $539 per user per year — confirm current pricing with your reseller as Microsoft repositions it periodically. That is approximately $27,000 per year at 50 staff, $108,000 at 200 staff, and $269,000 at 500 staff, recurring and rising with headcount. A private deployment is structured differently, and that difference is the whole point: a one-time implementation cost, typically $25,000 to $60,000 for a single well-defined use case, plus a running cost priced on query volume rather than on seats. Our Starter tier is $2,999 per month ($35,988 per year) for a single custom model and up to 100,000 queries per month; Professional is $7,999 for up to 500,000 queries; Enterprise is $14,999 for unlimited queries. Check the pricing page for current tiers before you model anything. Because there is no per-seat licence, adding staff does not by itself add cost — but query volume does, so the honest crossover depends on how heavily people actually use the system rather than on headcount alone. For scale: 100,000 queries a month is roughly 45 queries per person per working day across a 100-person organisation, or roughly 9 per person per working day across 500, so a light-use rollout can sit on Starter at a headcount where Copilot already costs six figures a year, while a heavy-use deployment moves up a tier well before that. Bring us your realistic volumes and we will model both sides against your actual numbers. Where Copilot comes out cheaper, we will say so.

This is the architecture we recommend most often, and it is usually cheaper and less disruptive than a full replacement. Copilot stays for general productivity — drafting in Word, summarising Teams meetings, triaging Outlook — operating on internal and OFFICIAL content where its in-app experience is genuinely unmatched. The private LLM handles regulated, client-confidential or commercially sensitive classifications, plus every question that requires grounding in systems outside Microsoft Graph. Microsoft Purview sensitivity labels become the routing boundary, so the split is enforced by classification metadata rather than by a policy document asking staff to remember what they must not paste. Both authenticate through Entra ID, so there is no second identity store. Most organisations also reduce their Copilot licence count in this model, because they stop buying seats for people who barely touched it.

Australian Privacy Principle 8 requires you to take reasonable steps to ensure an overseas recipient does not breach the APPs before you disclose personal information to them, and section 16C of the Privacy Act 1988 can make you accountable for that recipient's acts as if they were your own. Using an offshore-processed AI service over records containing personal information engages this directly, and the OAIC has published guidance on privacy obligations when deploying commercial AI products. The 2024 amendments raised the stakes: a statutory tort for serious invasions of privacy and new transparency requirements around automated decision-making. None of this makes Copilot non-compliant — Microsoft's contractual protections are substantial and many organisations reasonably conclude their APP 8 steps are satisfied. It does mean the assessment is yours to make and document, and a deployment where processing never leaves Australia is a materially shorter assessment.

Eight to twelve weeks from scoping engagement to production for a typical first deployment. The first two weeks are the licence and usage audit plus data classification design, which is where the coexistence boundary is decided. Weeks three to six cover infrastructure provisioning, model deployment, and building the RAG pipeline over your document corpus and priority line-of-business systems. Weeks seven to nine are integration with Entra ID and permission trimming, followed by systematic evaluation against a test set drawn from real queries your staff actually ask. The final weeks are pilot and staged rollout. Timelines extend when on-premises hardware procurement is involved, since GPU lead times are the constraint rather than the engineering. They compress when you are deploying to an Australian cloud region you already operate in.