Skip to main content

Before a Council Team Uses Generative AI: Choose the Task, Set the Guardrails

A practical guide for council teams to select a suitable generative AI task, check information and privacy risks, keep human accountability, and measure whether a controlled trial is useful.

Author avatar
Steve Agi 13 November 2026 · 7 min read
Before a Council Team Uses Generative AI: Choose the Task, Set the Guardrails

Start with the work, not the tool

A council team may have access to a generative AI tool and still not know where it belongs in day-to-day work. Starting with “What can this tool do?” often produces a long list of possibilities and little clarity about which one is safe, useful or worth keeping.
A more practical starting point is a specific, repeated task. For example: turning a publicly available policy document into a first-pass list of themes for a staff member to check. That is easier to assess than a broad ambition to “use AI across the council.”
This guide offers a way to screen one task before a team tries it. It does not replace a council’s policies, privacy assessment, procurement process, records requirements or security advice. The Australian Government and Victorian Government have published AI guidance for their own personnel and entities; local councils should treat it as useful reference material, not assume that every government policy applies to them.  

1. Describe the task in one sentence

Write down who does the work, what starts it, what input they use, what output they need and what happens after the task is complete. Keep the description narrow enough that a colleague could recognise the task when they see it.
For instance: “A governance officer uses a council-approved tool to draft a plain-language summary of a public agenda report for internal review; the officer checks every claim against the source before the summary is circulated.”
This description makes boundaries visible. It says the source is public, the output is a draft, the user is identified and a human checks the work. It does not imply that the system decides what council should do or publishes material on its own.
If a team cannot describe the task and the output clearly, it is too early to test a tool. First clarify the business need and current workflow. If a manual step exists because it provides a necessary control, do not remove it simply because AI can produce text quickly.

2. Check the information before writing a prompt

Identify the information the task would require. Ask whether it is already public, internal, confidential, personal, sensitive or otherwise restricted. Also check copyright, contractual terms, legal privilege and any other handling rules that apply.
The Victorian Government’s staff guidance says public generative AI tools should only be used with publicly available information, and that staff should use agency-approved tools only within their authorised purpose and information-protection settings. It warns against trusting outputs as accurate, unbiased or aware of local context.  The OAIC likewise recommends, as a matter of best practice, not entering personal information—especially sensitive information—into publicly available generative AI tools. It also says organisations should do due diligence on commercially available products, including privacy and security risks, human oversight and who can access information. 
For a council team, a safe screening question is: “Could this input identify a resident, staff member, supplier or individual case, or reveal non-public council information?” If yes, stop and check the council’s authorised-tool rules and relevant privacy, security and records advice before proceeding. Removing a name may not be enough if other details can identify someone.
A trial should not begin with live complaint files, personnel material, procurement responses or case records simply because those are the most interesting examples. Use public material or synthetic examples to test whether the task is worth exploring. Synthetic examples must be clearly labelled and must not be mistaken for real council records.

3. Decide what the tool may—and may not—do

Set the tool’s role before the first use. A low-risk support role might include creating a draft outline, grouping themes in a public document or suggesting clearer wording. The employee remains responsible for checking the result and deciding whether to use it.
Do not give a trial authority to make or communicate decisions that affect people, determine eligibility, assess a complaint or application, rank individuals, or take administrative action. Victorian guidance cautions public-sector personnel not to use generative AI to make decisions or undertake assessments with consequences for people, groups or organisations.  Even where a tool only informs a decision, the team needs to understand how its output is reviewed and how errors can be challenged.
Agree who can use the tool, which account and settings are approved, what information may be entered, what the reviewer checks and who owns the final material. If the team cannot identify an accountable person who can validate the output, the task is not ready for a trial.

4. Make human checking concrete

“Human in the loop” is not a control unless the reviewer knows what to check and has time and authority to reject the output. For a draft summary, the reviewer might check that every factual statement is supported by the source, names and dates are correct, qualifications are preserved, missing context is restored and the wording does not change the meaning.
The reviewer should compare the output to authoritative source material rather than ask the same AI tool whether it is correct. If a claim cannot be verified independently, remove it or rewrite it as a question for further research. Keep human judgement over publication, advice and decisions.
Consider who may be affected if an output is wrong or incomplete. A draft for internal brainstorming has a different risk from information used in a public notice, regulatory response or service decision. Increase review and approval controls as the consequence of error increases. Where rights, safety, privacy or access could be affected, seek the appropriate specialist review before any use.

5. Decide what needs to be recorded

Before trialling the task, agree where any final work product will be stored and what evidence is needed to explain how it was produced. The National Archives of Australia’s current guidance applies to Commonwealth records, not councils, but its risk-based approach is a useful prompt: retain AI outputs when they are final records, materially contribute to a significant record, evidence a sensitive or high-risk activity, or are relied on for a recommendation or decision. It also says that prompts, inputs and metadata may be needed to provide context, while not every prompt or draft must be kept. Local councils must apply their own records authorities, retention schedules and policies. 
For a council task, ask records staff to determine what constitutes the official record and capture it in the authorised system. Do not rely on a chat history or a vendor platform as the only record of a decision. Record the tool, task purpose, reviewer and material changes when that information is necessary for accountability or to understand the final work.

6. Test whether the task is actually better

Set a baseline before trialling. Depending on the task, this could be the time required to produce a first draft, the number of corrections needed, whether key points are missed, or how much review time is added. Choose one or two measures that reflect the purpose; avoid claiming productivity gains based only on how quickly the tool returns text.
Compare like with like. Use the same task boundary and quality criteria for the manual and AI-supported approaches. Record errors, rework and reviewer effort. A faster first draft may not save time if staff spend longer checking it or repairing omissions. Do not use one team’s limited trial to claim that an approach will work across all council services.
Agree a stop rule in advance. Pause if the tool receives information it was not approved to handle, produces an error that cannot be reliably caught, creates extra work, or changes the nature of the task in a way the team did not assess. After the trial, decide whether to stop, change the task, seek formal assessment or proceed under the council’s normal approval process.
The Australian Government’s current responsible-use policy requires covered Commonwealth entities to establish AI governance and accountability, train staff and assess AI use cases; it was updated in October 2026. Those requirements do not automatically extend to local councils, but they show the direction of mature public-sector practice: define ownership, understand the use case and consider risk before deployment. 

A short task-screening checklist

Before a council team tries generative AI, confirm that it can answer these questions:
•What exact task are we testing, and what is the business purpose?
•What information will be used, and is the tool approved for that information?
•Who owns the work and checks the result against authoritative sources?
•Could an incorrect output affect a person, a decision, a service or public trust?
•What is the official record, and where will it be captured?
•What baseline, review criteria and stop rule will determine whether to continue?
If any answer is unknown, pause and resolve it before using real council work. A small, clearly bounded test can help a team learn. It is not a shortcut around governance, professional judgement or existing obligations.
The goal is not to use AI wherever possible. It is to identify whether a specific task can be supported in a way that is useful, authorised, reviewable and consistent with the council’s responsibilities.

References

[1] Victorian Government, “Guidance for the safe and responsible use of generative AI in the VPS.”

[2] Office of the Australian Information Commissioner, “Guidance on privacy and the use of commercially available AI products.”

[3] National Archives of Australia, “Information management for records created using artificial intelligence (AI) technologies.”

Published by

Steve Agi Autonomous Automation Architect, C51 Consulting

About our partner

C51 Consulting

C51 Consulting helps Australian councils and public sector teams improve reporting, strengthen evidence traceability and reduce avoidable administrative effort. Council Reporting & Evidence Assurance — Review a defined corporate reporting process to identify evidence gaps, manual follow-up and unclear ownership. Indicative diagnostic: A$7,500 + GST, subject to agreed scope. Public Sector Workflow Reviews — Map one operational workflow, pinpoint bottlenecks and duplicated work, and identify practical improvements and appropriate automation options. Indicative diagnostic: A$7,500 + GST, subject to agreed scope. Practical AI Productivity Clinics — Build staff capability through a practical session on useful tasks, responsible use and internal policy constraints. Indicative clinic: A$1,500 + GST, subject to format and agreed scope. We start with a defined business problem, agree deliverables and focus on measurable improvement opportunities. We work alongside existing systems where practical; implementation is separately scoped. Clinics respect each organisation’s security, privacy, records-management and AI-use requirements. Request a scoping conversation about a council reporting process, government workflow or team clinic. https://c51consulting.com

Learn more