The Jeff Bleich Centre at Flinders University just told Parliament what most vendors won't say: sovereignty has two dimensions. Capacity (the compute, models and data centres), and Autonomy (who sets the rules).

Australia is only funding the first, and the numbers are stark.

AWS: $20B. Microsoft: $25B. The new AI Safety Institute: $29.9M. Capacity is pouring in on terms Australia didn't set.

Every enterprise buyer should consider the following.

Research published by the Jeff Bleich Centre put a thin governance layer (policy routing, enforced citations, approved sources, human escalation) over ten models from five vendors. Out of the box the models disagreed, wildly. On the same regulated questions, refusal rates ranged from 1.4% to 97.3%.. With the governance layer in place, that spread shrank more than fourfold. And every model improved.

Here's the fun part: an 8B open-weight model, self-hosted onshore, scored higher on refusal accuracy than every ungoverned frontier model in the test. Just read that again.

Accountable AI comes from the architecture, and not the model.

You certainly don't need the biggest model, or even the latest model for the workloads you’re currently designing. You just need to own the layer that governs it.

Which largely defines New Dialogue. Your rules, your sources, your audit trail, with models you can swap underneath.

Boards are continually monitoring and asking, "which model should we choose?" And it’s probably the wrong question. Instead "who controls the rules on which our AI runs?"

If the answer isn't you, you've bought capability. Not sovereignty.

‍

Author: Tod Pedler