# AI public consultation theme analysis.

URL: https://levystreet.com/industries/local-government/public-consultation-theme-analysis/
Updated: 2026-10-19

AI public consultation theme analysis for NZ councils: code every submission against agreed themes, with quotes and counts, so analysts report in days.

## Short answer

AI public consultation theme analysis reads every submission to a council consultation, tags each one against an agreed set of themes, pulls verbatim quotes and counts how views divide, so analysts can report to elected members in days rather than weeks. In New Zealand the Local Government Act 2002 still requires views to be considered with an open mind, so the AI organises submissions and people weigh them.

- Analysts set the theme frame; the AI applies it to every submission and quotes its evidence.
- Section 82 of the Local Government Act 2002 requires due consideration of views by the council.
- Counts describe who wrote in, not what the whole community thinks.
- Handwritten, organisational and te reo Māori submissions need their own checking step.

## What AI theme analysis produces

AI theme analysis turns a pile of submissions into a coded dataset: for each submission, the themes it raises, its position on each consultation question, and a verbatim quote supporting each tag. From that come theme counts, a quote bank, a list of new points nobody anticipated, and a list of submissions a person must read.

Councils are already doing this. [Hutt City Council](https://www.huttcity.govt.nz/council/our-projects/ai-at-council) reports that AI turned handwritten City Summit submissions into a detailed report in two days, a job it says could have taken weeks. It also keeps a public AI Register of its use cases.

## How submissions are coded, end to end


- **Inputs**: Online form responses, emails, PDF and handwritten submissions, hearing notes, the consultation document and its questions.
- **Systems**: The consultation platform export, a transcription step for scans, a language model run against a fixed theme frame, and a spreadsheet or BI tool for the coded dataset.
- **Steps**: De-duplicate and group template submissions; transcribe scans; analysts build the theme frame from a sample; the model codes every submission with quotes; quotes are checked against source text automatically; analysts review a sample and every flagged item; staff draft the report.
- **Human review**: Analysts own the frame, check a stratified sample, read all flagged, organisational and iwi submissions in full, and write the conclusions.
- **Outputs**: Coded dataset, theme summary with counts, quote bank, flagged list, and a methods note for the council report.
- **Measures**: Agreement between analyst and AI coding on the sample, quote match rate, days from close to report, flagged items resolved.
- **Risks**: Lost minority views, campaign submissions inflating counts, misread handwriting, invented quotes, and members relying on summaries alone.


## Illustrative example: a long-term plan consultation

Illustrative example: a mid-sized council (synthetic figures) receives 4,200 submissions: 3,100 online, 900 email or PDF, 200 handwritten. Assumptions: manual coding averages 5 minutes a submission; 150 are from organisations; the AI flags 12% for human reading; each submission uses about 2,500 input and 400 output tokens per pass, over two passes.

| Task | Manual | AI-assisted |
|---|---|---|
| Build theme frame from 300 submissions | Included in coding | 25 hours |
| Code all submissions | 4,200 × 5 min = 350 hours | Model run, hours not days |
| Check 10% stratified sample | Not done | 420 × 3 min = 21 hours |
| Read flagged submissions | Not separated | 504 × 4 min = 34 hours |
| Read organisational submissions in full | Included in coding | 150 × 15 min = 38 hours |
| Check handwritten transcriptions | Not applicable | 200 × 3 min = 10 hours |
| **Staff hours** | **About 350** | **About 128** |

Model cost: 21 million input tokens and 3.4 million output tokens comes to about US$76 at Claude Sonnet 5.5's [listed price](https://platform.claude.com/docs/en/about-claude/pricing) of US$2 and US$10 per million (19 October 2026). Staff time dominates; test your own figures in the [AI cost calculator](/resources/ai-cost-calculator/).

## The rules that frame the analysis

[Section 82 of the Local Government Act 2002](https://www.legislation.govt.nz/act/public/2002/84/en/latest/) sets the principles: views must be received with an open mind and given due consideration in making the decision, and submitters should get a clear record of the decisions and the material behind them. Section 82(2) requires processes for consulting with Māori. Where the special consultative procedure in section 83 applies, people must also get a chance to present views in person, so hearing notes belong in the dataset.

The Ombudsman's [October 2026 advice](https://www.ombudsman.parliament.nz/resources/advice-agencies-information-requests-and-ai) treats AI used to identify themes in large volumes of information as official information, prompts included. Write the methods note as if a submitter will request it.

## What stays with analysts and members

The theme frame is a judgement about what the consultation is about, so analysts write it and the model applies it. Staff read organisational, iwi and hapū submissions in full and summarise them by hand, and they check model accuracy on te reo Māori text before relying on it. Recommendations and the weight given to any view belong to staff and elected members.

## Where theme analysis goes wrong

| Failure | Why it happens | Control |
|---|---|---|
| Campaign inflation | 900 copies of one template counted as 900 arguments | Near-duplicate grouping, reported separately |
| Lost minority view | Rare points vanish into "other" | A "new point" flag that a person reads |
| Invented quote | Model paraphrases inside quotation marks | Exact string match against source, or reject |
| Stance errors | Sarcasm, conditional support, mixed answers | Sample checking by question; see [monitoring AI quality](/engineering/monitoring-ai-quality/) |
| Frame drift | Model invents themes mid-run | Fixed frame; new themes go to analysts |

The [AI quality review checklist](/resources/ai-quality-review-checklist/) is a useful template for the sampling step, and [summarising documents with citations](/learn/summarising-documents-with-citations/) covers the quote discipline.

## Running a first consultation

Pilot on a closed consultation your team already coded by hand, and compare agreement before using it live. Add the use to your AI register; our [AI governance](/services/ai-governance/) work covers that record and its review. The extraction pipeline overlaps with [AI document processing](/services/ai-document-processing/), [public information search](/industries/local-government/public-information-search/) and [consent application completeness checks](/industries/local-government/consent-application-completeness-checks/); the [local government guide](/industries/local-government/) shows how they fit together.

## Questions people ask

### Is it acceptable for a council to use AI to analyse submissions?

Section 82(3) of the Local Government Act 2002 leaves councils discretion over how they observe the consultation principles, and some councils already use AI to organise submissions. The test that matters is whether elected members still give the views due consideration with an open mind. That means people set the themes, check the coding, read the submissions that matter in full, and members can still see the original submissions.

### How do you stop form submissions from a campaign skewing the results?

Detect near-identical text before coding and report campaign submissions as a separate, clearly counted group. They are legitimate views and should be considered, but presenting 900 copies of one template as 900 independent arguments misleads the reader. The report should say how many submissions used each template and quote the template once.

### Can elected members just read the AI summary instead of the submissions?

They can use it as a map, not a substitute. The summary should link every theme to the submissions behind it so members can read the originals, and the report should say plainly how the analysis was done, what the AI did, and what staff checked. Organisational and iwi submissions should be summarised by staff who have read them in full.

### Is the AI's coding of submissions subject to LGOIMA?

Yes. The Ombudsman's October 2026 advice says information generated when an agency uses AI for official purposes, including prompts and responses, is official information. Keep the theme frame, prompts, coded dataset and checking records together, because a submitter who feels misrepresented may ask for exactly that.


## Sources

- [Local Government Act 2002 (version as at 1 October 2026)](https://www.legislation.govt.nz/act/public/2002/84/en/latest/)
- [AI at Council](https://www.huttcity.govt.nz/council/our-projects/ai-at-council)
- [Advice for agencies on information requests and AI](https://www.ombudsman.parliament.nz/resources/advice-agencies-information-requests-and-ai)
- [Pricing](https://platform.claude.com/docs/en/about-claude/pricing)

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