Four tests that make AI “private”.
“Private AI” gets used loosely. We hold it to four concrete tests — if any one fails, what you have is a public tool with your data in it:
Your data stays inside your controls.
Documents, records and knowledge bases live in an environment governed by your access rules and data-residency requirements — not scattered across consumer accounts.
Nothing trains public models.
Prompts, documents and outputs are contractually and technically excluded from model training. Your competitive knowledge never becomes everyone's model.
Access is controlled and logged.
Who used AI on what, when, with which data — answerable from the record. That's what makes AI defensible to a board, an auditor or a regulator.
The AI works to your policies.
Safeguards, house standards and human-approval gates are built into the system itself — so quality and compliance don't depend on every user prompting carefully.
Your team is already using AI. The only question is on whose terms.
In most organisations, staff quietly paste work into consumer AI tools because the productivity gain is real. The knowledge in those prompts — client details, pricing, unreleased work — leaves your boundary with them. Banning AI doesn't stop this; it just keeps it invisible.
Private AI is the grown-up answer: give your people AI that's better than the consumer tools — grounded in your own knowledge, integrated with your systems — inside a boundary you control. Adoption goes up, exposure goes down, and the capability compounds as yours.
Who insists on private AI
Organisations whose data is the business or whose sector is regulated: professional services carrying client confidentiality, healthcare, education providers under quality frameworks, financial services, government and iwi organisations with data-sovereignty obligations, and any firm whose bid library, formulas or research is its moat.
Common questions
Is private AI the same as self-hosting a model?
No — self-hosting is one implementation, and usually the wrong first one: you inherit weaker models, GPU bills and an ops burden. Private AI is about the boundary, not the plumbing: frontier models consumed under enterprise no-training terms, inside an environment you govern, can pass all four tests above while staying at the frontier of capability.
Isn't the enterprise tier of a chatbot enough?
It fixes the training-leakage problem and little else. It doesn't ground answers in your knowledge, encode your policies, integrate your systems or log activity in a form your auditor can use — and it leaves quality dependent on each employee's prompting. Private AI as we build it is an application layer, not a subscription tier.
Does private AI meet New Zealand and Australian data requirements?
That's the design goal: ANZ data residency on AWS Australia, ISO 27001 / NZISM-aligned controls, and contractual terms documented in our legal centre. Specific obligations (health data, government classifications, iwi data sovereignty) are assessed per engagement — bring yours to the first conversation.
What does private AI cost?
Engagements are structured as a build fee in small committed blocks, then a monthly subscription that covers hosting, monitoring, support and improvement — sized to the system, not per seat. Our approach explains the model; talk to us for numbers against your use case.
Related: AI agents for business · AI for governance, risk & compliance · Security Statement · All case studies