Generative AI is an easy way to produce a paragraph. Policy work is harder: it means weighing evidence, understanding who is affected and explaining why a recommendation follows. Research with senior public servants gives a useful early view of where the technology fits, and where it does not.

FIELD GUIDE

A bounded policy workflow

01 Frame Define the policy question and audience
02 Explore Use AI to organise or test supplied material
03 Judge A policy professional checks evidence and consequences
The system can help structure thinking. The policy judgement stays with the accountable person.

What the research covered

UNSW Canberra and the Australian National University report research involving 37 senior public servants from 22 state, territory and federal agencies. The fieldwork ran from November 2024 to January 2025. It explored perceived strengths, weaknesses and conditions for safe use in policy work.

This is exploratory research, not a controlled productivity trial. It is valuable because it asks practitioners about the work itself, rather than assuming that a fluent answer is a useful policy answer.

Where a drafting assistant can help

A policy team can use a model to organise a long pack of supplied documents, produce a first outline, identify questions that need evidence, or rewrite a technical explanation for a different audience. These are support tasks. They do not establish that a proposal is lawful, fair, effective or politically workable.

The useful boundary is visible in the workflow: provide a defined evidence set, ask for a structured working draft, then verify every material point against the source. Keep the source list beside the draft so reviewers can trace claims quickly.

Try this with one live brief

Choose a document-heavy task with a clear reviewer. Ask the model to group evidence under the brief's existing headings and list gaps separately. Do not ask it to invent evidence or decide which community impact matters most. Compare the draft with the current manual process and record what still required expert judgement.

  • Keep confidential or Cabinet-sensitive material out until approved tools and handling rules are confirmed.
  • Ask reviewers to mark unsupported statements, omissions and useful structure separately.
  • Stop if the source material is incomplete or the task requires a decision the model cannot own.

Want to explore how this could apply to your organisation?

Explore Private AI ↗