AI for New Zealand councils.

Where AI helps a council, and where it must stop

AI helps councils where officers spend hours reading before they can act: checking whether a consent application is complete, working through thousands of submissions, searching district plans and bylaws, or finding documents for an information request. It must stop short of decisions. Granting consents, weighing submissions, withholding information and setting policy belong to delegated officers and elected members, and the legislation is written on that assumption.

Central government is moving quickly, which sets expectations for councils. The Government Digital Delivery Agency’s 2026 cross-agency survey recorded 545 AI use cases across 59 public sector organisations, double 2025’s 272, with about 31% operational. Councils were not part of that survey. Trust is the constraint: in the Privacy Commissioner’s 2026 survey, only 31% of New Zealanders rated their trust in government agencies to handle personal information as high or very high, though that was the highest of any group asked about.

Council workflow map

Each of the first three has a worked guide: consent application completeness checks, public consultation theme analysis and public information search.

The statutes that shape council AI

• LGOIMA 1987. A council must decide on a request and notify the requester “as soon as reasonably practicable, and in no case later than 20 working days” after receiving it (s 13), with extensions allowed in defined cases (s 14), under the Act as at 1 October 2026. AI can find documents faster; the release decision stays with an authorised person.
• Local Government Act 2002. The consultation principles in section 82 require views to be “received by the local authority with an open mind” and given “due consideration” in making a decision, and information to be provided in a format suited to people’s needs. Section 81 requires processes for Māori to contribute to decision-making, so submissions in te reo Māori and from mana whenua need careful handling.
• Public Records Act 2005. Every local authority must “create and maintain full and accurate records of its affairs”, including contracted-out work (s 17). Prompts and outputs that inform decisions belong in the records system.
• Building Act 2004. A building consent authority must grant or refuse a complying application within 20 working days, and the period is suspended while it waits for further information it has required (s 48).
• Frameworks. The Public Service AI Framework is advisory and written for public service agencies, which councils are not. The Ministry for Regulation’s May 2026 guidance says people remain responsible for decisions, with AI as “a support not a substitute”. We found no LGNZ or Taituarā AI guidance.

This is general information; take advice from your legal team or the Ombudsman’s guidance on specific requests.

What other councils are doing

Two public examples are worth reading. Hutt City Council publishes where it uses AI, roots its approach in Tika and Pono, and reports that AI turned handwritten City Summit submissions into a detailed report in two days. It has an AI risk management framework, AI training for new staff and an internal AI governance group. Environment Canterbury re-established a councillor AI working group on 13 May 2026, meeting up to four times a year with workshops generally open to the public.

Most councils run TechnologyOne or similar for property and regulatory work, Objective or SharePoint for records, and a CRM for service requests. AI should read from these with officers’ permissions and write its outputs back into the records system. Our notes on SharePoint and building private AI search cover the pattern.

A first project: consent completeness checks at lodgement

Completeness checking is narrow, measurable and decision-free, and it attacks a delay applicants feel directly.

1. Codify the checklist. Turn the council’s lodgement checklist for two or three common consent types into explicit rules: documents, signatures, producer statements, plan sheets.
2. Test on history. Run the checks on 200 past applications and compare with the further-information requests officers actually issued.
3. Run at lodgement. Give the lodgement officer a list of likely gaps and inconsistencies, each pointing to the page or document.
4. Officer decides. The officer accepts, returns or queries the application. The AI never accepts or refuses.
5. Track the clock. Compare further-information requests and suspended days with the same months last year.

Illustrative example: a council receives 120 building consent applications a month and issues further-information requests on 30% of them, 36 requests, each taking an officer about 45 minutes to write and track. If checks at lodgement prevent half, that is 18 fewer requests, about 13.5 officer hours a month, plus the days applicants no longer spend waiting with the clock suspended. Our AI document processing service runs this kind of test against past applications.

Where AI fails in council work

For councils without an AI policy yet, our AI governance advisory covers the register, approvals and public reporting; public information search itself is a RAG development project, built so every answer links to the council document it came from.

Workflows we map in this sector

• Public information searchCited answers from council documents, with official information requests routed to the people who decide them.
• Consent completeness checksA first pass over building and resource consent lodgements that lists missing items with page references for an officer.
• Consultation theme analysisEvery submission coded against an agreed theme frame, with verbatim quotes and counts that analysts check.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.