# AI public information search.

URL: https://levystreet.com/industries/local-government/public-information-search/
Updated: 2026-10-19

AI public information search for NZ councils: cited answers from plans, bylaws and minutes, with LGOIMA requests routed to people and every search logged.

## Short answer

AI public information search lets council staff, and later residents, ask questions in plain English and get answers quoted from the council's own plans, bylaws, minutes and reports, with a link to each source. In New Zealand it has to work inside LGOIMA: any question can be a request for information, prompts and outputs are official information, and decisions on requests stay with people.

- Answer only from council documents, and cite the clause or page every time.
- A question typed into a search box can be an LGOIMA request; route it, do not answer it.
- AI prompts and outputs used for council business are official information and public records.
- Start staff-facing on published material before opening a resident-facing version.

## What AI public information search does for a council

AI public information search answers questions from the council's own documents and shows the paragraph it used. A planner asks "what is the maximum building height in the Medium Density Residential Zone?" and gets the rule, the plan chapter and a link, instead of a list of 40 PDFs.

The useful corpus is larger than most people expect: the operative and proposed district plans, bylaws, the long-term plan and annual plans, adopted policies, open agendas and minutes, and LGOIMA responses already released. The hard part is not the search. It is knowing which version is current, which documents are public, and which questions the system must not answer itself.

## The workflow at a glance


- **Inputs**: A question from a staff member (later, a resident), plus an indexed set of published council documents with version dates and status (operative, proposed, superseded).
- **Systems**: The council's EDRMS or [SharePoint](/integrations/sharepoint/), the website's document library, the meetings system, a retrieval index and a language model that answers only from retrieved text.
- **Steps**: Classify the question; route anything that requests unpublished information to the LGOIMA queue; retrieve passages; draft an answer with citations; show sources and dates; log the exchange.
- **Human review**: Staff check answers before they reach a resident. The LGOIMA team reviews every routed request. A records officer samples logs weekly.
- **Outputs**: A cited answer, or a handover note that starts an official information request with the received date recorded.
- **Measures**: Share of answers judged correct against a test set, citation accuracy, time to locate information, LGOIMA requests answered by pointing to published material.
- **Risks**: Answering from a superseded plan, exposing public-excluded papers, treating a request as a chat, unlogged answers.


## Illustrative example: one LGOIMA team, one month

Illustrative example: a provincial district council (synthetic figures) receives 50 LGOIMA requests a month. Assumptions: a third ask for information that is already published, locating material takes an average of 3 hours per request by hand, and AI-assisted search with officer review brings that to 1 hour.

| Item | Before | With AI search |
|---|---|---|
| Requests per month | 50 | 50 |
| Hours locating information | 150 | 50 |
| Requests answered by pointing to published material | Found late or missed | About 17 identified on day one |
| Officer review of AI-found material | Not applicable | Every request |

The arithmetic: 100 officer hours a month back for decision-making and drafting. Those hours only appear if the test set shows the search is finding what a careful officer would find; the evaluation method is in our guide to [RAG retrieval evaluation](/engineering/rag-retrieval-evaluation/).

## Which New Zealand rules shape the design

Three rules change how this system is built.

- **Any question can be a request.** Under [section 10 of LGOIMA](https://www.legislation.govt.nz/act/public/1987/174/en/latest/), a request may be made in any form, including orally, and need not mention the Act. Section 13 then gives the council no more than 20 working days to decide it. A resident-facing search box must recognise "can I see the report on..." and start the clock, not answer from memory.
- **The AI's own outputs are official information.** The Ombudsman's [advice for agencies on information requests and AI](https://www.ombudsman.parliament.nz/resources/advice-agencies-information-requests-and-ai) (6 October 2026) says prompts and responses generated for official purposes are subject to LGOIMA. It also says that when AI helps collate information, the agency should record the search parameters, have an official review the results, and tell the requester that AI was used.
- **Logs are records.** [Section 17 of the Public Records Act 2005](https://www.legislation.govt.nz/act/public/2005/40/en/latest/) requires every local authority to create and maintain full and accurate records of its affairs. Design the [audit log](/engineering/ai-audit-logs/) with your records team before launch.

## What stays with a person

People decide anything that carries legal weight. That includes whether a withholding ground under section 7 applies and whether the public interest outweighs it, charging, extensions, and refusals. It also includes any statement about a specific property's consent history or compliance, and any answer the system cannot support with a cited source. The Ombudsman's guidance is explicit that decisions on requests should still be made by people.

## Where public information search goes wrong

| Failure | What it looks like | Control |
|---|---|---|
| Version drift | Quotes a rule from a superseded plan change | Index status and effective date; show both on every answer |
| Leaked confidential papers | Public-excluded agenda items (section 48 of LGOIMA) appear in results | Separate indexes by access level; test with known confidential items |
| Unsupported answer | Fluent reply with no source | Refuse to answer without a retrieved passage |
| Missed request | Resident asks for an unpublished report and gets a chat reply | Request classifier tested on last year's LGOIMA log |
| Scanned documents | Older minutes are image-only PDFs and never match | OCR pass and a coverage report before launch |

Teach staff to [check AI answers against sources](/learn/checking-ai-answers-against-sources/) during the pilot; the habit matters more than the interface.

## What it costs and how to start

Cost is driven by document volume and quality, permissions work, and integration with the EDRMS and meetings system, not model fees. Start staff-facing with published material, build a test set of 150 real questions from last year's enquiries and LGOIMA log, and only open a resident version once accuracy holds. Our [RAG development](/services/rag-development/) service covers this build, and [building private AI search](/engineering/building-private-ai-search/) explains the architecture.

The same document pipeline supports [consent application completeness checks](/industries/local-government/consent-application-completeness-checks/) and [public consultation theme analysis](/industries/local-government/public-consultation-theme-analysis/). The [local government guide](/industries/local-government/) maps where each fits.

## Questions people ask

### Can a council put an AI chatbot on its website to answer resident questions?

Yes, but start narrower than most vendors suggest. Limit it to published documents, make it cite every answer, and have it hand over anything that asks for information the council holds but has not published, because that is likely to be an LGOIMA request with a 20 working day clock. Run it with staff for a few months first and measure how often its answers are right.

### Are the prompts and answers from council AI tools subject to LGOIMA?

Yes. The Ombudsman's October 2026 advice for agencies says information generated and retained when AI is used for official purposes is held by the agency and subject to the OIA or LGOIMA, including the prompts and the responses. Councils should keep search logs under their records policy and expect requests for them.

### Can AI decide whether to release or withhold information?

No. AI can find documents that may be in scope, suggest where redactions might be needed and draft plain-English replies. Deciding whether a withholding ground applies, whether the public interest outweighs it, and whether to charge or extend are decisions for an authorised officer, and the Ombudsman expects a person to read and review what an AI search has found.

### What documents should go into a council AI search index first?

Start with material that is already public and stable: the operative district plan, bylaws, the long-term plan and annual plans, adopted policies, open meeting agendas and minutes, and previously released LGOIMA responses. Leave property files, personnel records and public-excluded meeting papers out of the first version, because permissions and privacy controls take longer to get right.


## Sources

- [Local Government Official Information and Meetings Act 1987 (version as at 1 October 2026)](https://www.legislation.govt.nz/act/public/1987/174/en/latest/)
- [Advice for agencies on information requests and AI](https://www.ombudsman.parliament.nz/resources/advice-agencies-information-requests-and-ai)
- [Public Records Act 2005 (version as at 6 June 2026)](https://www.legislation.govt.nz/act/public/2005/40/en/latest/)

Published by Levy Street (https://levystreet.com), Wellington, New Zealand.