Some documents cannot be pasted into a public chatbot: client files, board papers, health records, contract terms, plant data. We design and build AI environments that run inside your network or a private Australian cloud region, so your staff get the tools and the data stays where you can account for it.
Staff already use AI. When the approved tools cannot touch sensitive work, people paste it into whatever is open in their browser. Blocking public chatbots can simply push the habit onto personal phones. Our post on staff leaking business data into AI tools covers how that happens.
The alternative is to give people an AI tool that is allowed to see the sensitive material because it runs somewhere you control. Open-weight language models can now be run on a single server or a private cloud instance using inference software such as Ollama or vLLM, with a search index over your own documents so answers cite your files. Our local LLM guide shows the basic idea on a laptop. Doing it for an organisation adds access control, logging, capacity planning and an owner.
For a working example, SupportAgent is a self-hosted investigation tool our founder built on Ollama. It is the founder's own product, and it shows the same architecture we design for clients.
The reason to go private is the data and the obligations attached to it.
Law, accounting, advisory and engineering firms whose engagement letters limit where client material can be processed.
Records covered by the Privacy Act where a cross-border disclosure would need justifying under APP 8.
Banks, insurers and super funds assessing AI suppliers against APRA CPS 230 and CPS 234.
Agencies and contractors with data handling clauses, security classifications or the NSW AI Assessment Framework to satisfy.
Finance and executive teams who want AI summaries of board packs, mergers and forecasts without the material leaving the business.
One specialty: sites with segregated OT networks or SOCI Act obligations, where some AI must run without an internet connection.
Sector detail is on our pages for financial services, healthcare and government.
The right one depends on how sensitive the data is, how much you will use it and who will run it.
| Metric | On-premises | Private cloud in an Australian region | Improvement |
|---|---|---|---|
| Where data is processed | Your own servers and network | Your cloud account, Australian region | Both keep data out of public AI services |
| Internet needed | No, can run fully disconnected | Yes, within your private network | On-premises suits segregated sites |
| Upfront cost | Hardware purchase | Little or none | Cloud is easier to trial |
| Scaling | Limited by hardware bought | Add capacity as needed | Cloud handles uneven demand |
| Who maintains it | Your team or a managed service | Shared with the cloud provider | Both need a named owner |
A hybrid design sends sensitive requests to the private model and routes everyday work to a hosted service or Microsoft Copilot. That keeps hardware small while protecting the material that matters. Our comparison of hosted and local models and the build versus buy cost guide help with the trade-off.
A private model on its own is only half the system. The parts that take the most care are the permissions and the logs. If the search index ignores folder permissions, a private AI can still show a junior staff member the payroll file. We carry your existing access groups through to retrieval, so the model only sees what the person asking could open, and every request is logged for review.
Model choice comes last. We shortlist open-weight models, test them on a sample of your real documents and questions, and pick the smallest one that does the job well, because smaller models need less hardware. A technical guide to LLMs covers the terms if your IT team wants the detail.
Under the Privacy Act, APP 8 governs sending personal information to an overseas recipient, and APP 11 requires reasonable steps to protect it. Keeping processing in Australia does not remove either duty, though it makes the cross-border question simpler to answer. Our guides to Privacy Act compliance and AI and data sovereignty in Australia go further.
APRA-regulated entities assess material service providers under CPS 230 and information security under CPS 234. A private deployment changes which provider you are relying on and what you can evidence. See our post on CPS 230 and AI service providers. Critical infrastructure owners have SOCI Act risk management programs to consider, covered in our SOCI Act post.
If your sensitive content already lives in Microsoft 365 and the main risk is who can see what, a private model will not fix that. Copilot readiness work on permissions and labels usually comes first.
If usage will be light, a hosted model in an Australian region with the right contract terms is often cheaper and simpler. If nobody will own patching and model updates, the environment will age badly unless it sits under managed AI services. And if you have not yet decided what AI should do for the business, start with AI strategy. Our Australian versus offshore comparison is a useful read before choosing.
Solve8 is based at 9/86 City Rd, Beenleigh QLD 4207. Cloud builds are delivered remotely anywhere in Australia, and on-premises installs in South East Queensland can be done on site. See the Brisbane AI consulting page, or read about our system integration work for connecting the environment to line-of-business systems.
Our principal consultant brings 18 years of enterprise delivery experience from projects for BHP, Rio Tinto and Senex Energy, delivered with previous consulting employers. These companies are not Solve8 clients. You work directly with senior people, not a sales team handing you off to juniors.
Book a 30-minute scoping call. Tell us what data you need AI to work with and what rules apply to it, and we will tell you which deployment model makes sense.