AI in government: which work should AI assist, augment or act on?

Helix's public-sector work includes Port Adelaide Enfield, the SA Department for Education and NSW Government (Policy Lab).

With the City of Mitcham, Helix narrowed 51 use cases to 3 pilots and ran a six-week AI Pod.

Most government AI use still sits in Assist × structured work (Q1 Operator): retrieval, drafting, summarisation and known workflows where a public servant stays at the helm. That is a description of the common pattern, not a league table and not a statistic about every agency. Some programmes already reach Augment on structured or dynamic work; fewer have earned Adapt on complex, rights-affecting work — and that is often the right answer.

Do government agencies use AI? Yes — across service delivery, policy support, education and local government. How is the Australian government using AI? Follow programme and policy updates on digital.gov.au and related Commonwealth channels; this page answers the design question beside them. How are local councils using AI? Often as Assist tools and bounded pilots first; see Helix’s government and enterprise work. For APS and council ethnography contexts, see Ethnobot for the APS and Ethnobot for councils.

When a council or agency lists dozens of ideas, the first decision is placement, not build. Assist on structured work can be the coherent first cell. Jumping to Adapt without the trust budget is how portfolios mislabel automation as transformation.

Education and policy work fit the same grammar. The SA Department for Education and NSW Government (Policy Lab) show design and practice inside government systems; they are not proof that Adapt-row agents are already the default. Start where the work and the trust actually sit, then move only when the override path and the public story can keep pace with the agency you are granting.

As agency rises, so must trust already earned. The book’s line is exact: don’t buy a Q9 outcome with a Q1 trust budget. In public service that gradient is sharper because the public holds the downside — fairness, explainability and the right to a human decision when rights are at stake.

• Q1 — people stay at the helm; AI helps inside known procedures.
• Q5 — people and AI iterate together; a person remains accountable on each step.
• Q9 — AI acts across complex systems; humans govern purpose, thresholds and legitimacy, on call for exceptions.

Between those markers, Augment on structured work (Q2 Tactician) is risk scores, triage flags and decision support with a human clearing each call. Adapt on structured work (Q3 Automator) is where invoice and workflow automation belongs — machines run the process; people govern exceptions. Neither cell is “more advanced” than Assist; each asks for a different trust budget.

The path Q1 → Q5 → Q9 is a capability path, not a deployment strategy. Off-diagonal moves need an explicit reason. Board-facing trust language lives on the 9Q Grid for boards. Agency design detail lives on AI autonomy levels. For who should decide on a single call, see the decision-making framework.

Public-evidence readings from the 9Q Case Library — not Helix clients — show what happens when Adapt-row automation outruns matching governance. See Robodebt and Dutch SyRI. They are labelled that way on purpose: the grid reads the published record, not a client engagement. The failure pattern is diagnosis first — the machine’s role was sold ahead of the oversight the work demanded.

Mislabelling in government often sounds like service improvement while the system is already acting on structured eligibility or risk scoring. If humans are not in the loop on each step, and the work touches rights or money, you are not in a drafting assistant cell. Name the quadrant honestly before you scale. The cost of a wrong label is not a bad slide — it is a public decision made at the wrong agency level.

Public-evidence readings also show careful placement. See Singapore GovTech in the library — again public evidence, not a Helix client — for a reading of how a government technology agency places agency against work. Good looks like Assist where people must stay at the helm, Augment where iteration is the product, and Adapt only where boundaries, exception paths and public legitimacy already exist.

Australian government AI guidance lives on .gov.au — including digital.gov.au policy, assurance frameworks and APS AI plans. Who in government is responsible for AI? Follow the accountable-official and policy wording as published on digital.gov.au; this page does not restate statute. The 9Q Grid answers the delivery and design question beside that policy: which role may AI take on this work, and what oversight does that role need?

Policy as a journey — purpose, authority, design and delivery — is the subject of The Policy Playbook by Catherine Althaus, Suhit Anantula and Pia Andrews. Use the book for the policy craft; use 9Q for the AI-role map inside each stage. Helix delivery for agencies and councils starts at government and enterprise.

Practically: open the policy page for the controls you must meet, then open the grid for the role AI is actually taking. If the role is Assist, keep a person at the helm and measure speed and quality. If the role is Augment, keep a person in the loop on each step. If the role is Adapt, publish the boundaries, the exception path and the review cadence before the system acts at scale.