Beyond Copilot: building broader Māori AI capability

Microsoft Copilot will be the first substantial organisational encounter with generative AI for many people in Aotearoa. There are understandable reasons for that. Microsoft 365 is already deeply embedded across government and business, identity and permissions already exist, documents are already in SharePoint and OneDrive, and security teams already understand the platform.

There is nothing inherently wrong with using Copilot. The danger is subtler: we may allow one product to define what we think AI is. If Māori organisations want real choice over future AI systems, product literacy needs to grow into broader technical capability.

Useful, but not the whole field

Microsoft currently states that organisational prompts, responses and Graph data used by Microsoft 365 Copilot are not used to train its foundation models. For many ordinary enterprise workloads, that makes Copilot a rational and useful tool. See the current Microsoft 365 Copilot privacy guidance for the detail that applies to the service.

A staff member might use Copilot to summarise a Word document, draft an email, prepare meeting notes or search Microsoft 365. Those are legitimate productivity gains. But they represent only one part of the wider AI landscape.

Artificial intelligence also includes models that run entirely on desktop computers, open-weight reasoning systems, local document retrieval, specialised vision models, geospatial analysis, agents, automated workflows, structured extraction, coding systems, private APIs and systems operating in disconnected environments. If organisational AI education consists entirely of prompt training inside one commercial product, the organisation is developing product competence rather than broad AI capability.

Procurement should not become strategy

There is a familiar technology pattern in large organisations. A platform is purchased for good reasons, then every new requirement starts being expressed as, “How do we do this in the platform?” The original procurement decision gradually becomes the technology strategy.

AI creates the same risk. An organisation may genuinely benefit from Copilot for everyday administrative work while choosing a completely different architecture for sensitive historical research. It may use a commercial frontier model for public-data analysis, a local open-weight model for restricted document search and no AI at all for another collection.

There is no technical reason every AI task must go through the same vendor. The best architecture should follow the information, the purpose and the consequence of the work.

Capability is part of sovereignty

Te Mana Raraunga has long treated capability building as part of Māori Data Sovereignty. The principle is important because authority without practical capability can still leave organisations dependent on somebody else to implement every decision.

Māori organisations do not need to train trillion-parameter foundation models to build meaningful AI capability. Much more immediate skills are available: knowing how models work, understanding the difference between cloud and local inference, running a model locally, building RAG, understanding embeddings, testing output quality, managing permissions, evaluating security and knowing when a task should not use AI at all.

That knowledge changes the relationship with vendors. The organisation becomes a buyer making architectural choices rather than simply a customer accepting whatever AI features arrive inside an existing subscription.

Build translators between governance and engineering

One of the most valuable roles over the next few years will be people who can translate Māori data-sovereignty principles into technical requirements. A governance statement such as “our information must remain under our authority” needs to become practical design decisions about networks, storage, models, logging and access.

That might mean specifying that a model runs locally, that one document collection can never be indexed with another, that embeddings remain within an iwi-controlled environment, that certain users can query only selected collections, or that a system may retrieve evidence but must not make consequential decisions about people.

This is where principles become infrastructure. It also reduces the risk that technical teams and governance teams talk past each other, each assuming the other understands what the system actually does.

Local models should be part of training

For every group learning ordinary AI prompting, it would be useful to develop a smaller technical cohort that goes deeper. Give them suitable hardware and let them install local models. Let them compare model families, understand memory requirements, test quantisation, build a small document RAG system and examine how access controls work.

Let them compare local outputs with commercial models. Let them investigate when a smaller local model is good enough and when a cloud frontier model is materially better. Let them learn where local models fail on te reo Māori, Aotearoa geography, Māori names or specialist subject matter.

This can all begin with public information. Sensitive Māori material is not required to learn the technology.

Do not confuse prompting with expertise

Prompting matters, but it is not the deepest capability organisations will need. Good prompts can improve the usefulness of a system, but prompting cannot fix a poor architecture, inappropriate data access, weak provenance, a model that lacks capability, or a service that should never have received the information in the first place.

The long-term skills are more structural. People need to understand how information reaches the model, how retrieval works, how systems fail, what is logged, how models are evaluated and how to preserve evidence. These are the skills that allow an organisation to change vendors without losing its understanding of AI.

A Māori capability network

There is also little reason for every iwi or Māori organisation to repeat the same technical experiments independently. Shared learning could accelerate capability without requiring shared sensitive data. Organisations can compare model performance using public test sets, share installation knowledge, document hardware requirements, test Māori language behaviour and exchange patterns for secure local deployment.

A Māori AI community of practice could therefore focus not only on policy and ethics but on engineering. It could publish model evaluations, local installation guides, reference architectures and practical security patterns. It could help smaller organisations access knowledge that would otherwise require specialist consultants.

The important point is that collaboration on technical capability does not require pooling all Māori information into one giant system. Shared capability and local authority can coexist.

Avoid a false choice

The choice is not between Microsoft Copilot and rejecting commercial AI. Nor is it between cloud AI and building everything ourselves. A mature organisation will probably use several architectures at once.

Routine office productivity may sit comfortably in an enterprise cloud platform. Public-source research may use a powerful commercial model. Sensitive document retrieval may run locally. Highly restricted information may remain outside AI systems altogether. The architecture can be proportionate to the kaupapa.

That is a stronger position than choosing one platform and forcing every use case into it.

A practical recommendation

Continue ordinary AI training, but expand it. Develop a technical AI capability track alongside user education. Include local models, RAG, open-weight licensing, model evaluation, document extraction, privacy testing, access control, GPUs and AI security. Include GIS and other domain specialists because many valuable AI uses will happen inside existing professional workflows rather than in standalone chatbots.

Within a relatively short period, an organisation can develop people who understand not merely how to use AI but how to make informed choices about AI. That capability strengthens rangatiratanga because it increases the number of realistic options available.

Next in the series: Māori GIS is an excellent place to start.