Skip to main content
assembl

Compliance · Model Use

AI use, plainly.

assembl uses language models as drafting and retrieval tools. The product promise is not autonomy. It is a better first pass, grounded in sources, reviewed by people, and sealed with proof.

No external action is sent automatically. Every material output remains draft-only until a named human reviews and approves it.

No customer workflow input is used to train public foundation models.

Personally identifiable information is minimised or masked before model calls where the workflow permits it.

Every evidence pack records sources, assumptions, reviewer decisions, timestamps, and audit metadata.

Users can see when output is grounded in live data, static policy, uploaded files, or model reasoning.

Models we may call

Current production workflows may call Claude Sonnet 4.6, Gemini 2.5 Flash, Gemini embedding models, OpenAI models where configured, or deterministic retrieval and scoring code. Model choice depends on task risk, latency, and whether citations are required.

The four pou

Rangatiratanga means people keep agency. Kaitiakitanga means data is cared for. Manaakitanga means tools are useful without being pushy. Whanaungatanga means handoffs are visible and accountable.

What we do not do

assembl does not make final legal, financial, employment, health, or entitlement decisions. It does not auto-send messages, auto-file documents, or replace professional judgement. It prepares drafts for review and records the reasoning trail.

Public-sector posture

Our AI use is shaped by New Zealand public-service expectations for safe, transparent, and responsible use of generative AI. We design for review, source visibility, record keeping, and clear user accountability.