How we turn a thousand pages of council documents into plain English

A council Long Term Plan routinely runs to 500–900 pages. An Annual Plan is shorter but still dense. A Fees and Charges schedule can be 40 pages of tariffs with no plain-English note on what changed. A District Plan — the rulebook for what you can build on your own property — stretches across thousands of pages of policies, rules, and overlays that cross-reference each other.

These documents are written for statutory compliance, not for the people they affect. So almost no ratepayer reads them. The decisions that change what you pay and what you can do are real, public, and adopted — they are simply buried.

This site exists to dig them out. Below is exactly how we do it, start to finish, so you can judge whether to trust what you read here.

What we ingest — and what we leave out

The first decision in the pipeline is also the most consequential: which documents are allowed in. We ingest only adopted council documents that describe what happens to you — what you pay, what you can or can't do, and what's coming to your area.

The exclusions matter as much as the inclusions. Annual Reports describe what a council already spent in a past year — useful history, but not a guide to what you'll pay next. Funding Impact Statements are framed around council revenue and income, not around the dollar figure that lands on a ratepayer's bill, so they answer a different question than the one this site asks. Submissions and consultant reports are opinions and advice, not decisions. None of them produce impact cards. That boundary is enforced in the ingestion code, not left to a reviewer to catch later.

The pipeline, step by step

Once a document is approved for ingestion, it moves through a fixed sequence. Each step is designed to remove a specific way the process could go wrong.

1. Source the official document

We start from the council's own published file — the adopted PDF on the council website or its consultation portal. No secondhand summaries, no news coverage. The source is always the primary document.

2. Extract the text

The document is converted to machine-readable text. Many council PDFs are scanned images rather than digital text, so the pipeline falls back to optical character recognition (OCR) for those files, page by page, until usable text is recovered.

3. Strip the boilerplate

Before any AI sees the text, we remove the material that carries no impact: mayoral and chief-executive forewords, vision and mission statements, ceremonial front matter, acknowledgements, and consultation preambles. This keeps the model focused on the substantive content and keeps processing costs down.

4. Extract into single-issue impact cards

The cleaned text is passed to an AI model with a strict framework: pull out one decision at a time — a fee that changes, a new targeted rate, a rule that restricts or allows something — and record what it is, who it affects, the dollar figure or rule, and when it takes effect. One issue per card. No bundling, no narrative.

5. Verify every number against the source

This is the step that makes the cards trustworthy. Every figure on a card must trace back to verbatim text in the source document. A card whose numbers can't be matched to a contiguous quote in the original is flagged and held — it does not go live. The model is not allowed to round, infer, or fill gaps.

6. Human review before publishing

Nothing publishes automatically. A person reviews the extracted cards, checks the flagged items, removes duplicates, and routes anything that isn't a genuine personal impact out of the feed. AI surfaces; people decide what goes live.

The editorial standard

One test governs every card: how does this affect me? A card earns its place only if it answers what you pay, what you can or can't do, or what's coming to your area. Everything is stated as fact — the figure, the rule, the date — with no spin and no editorializing about whether a change is good or bad.

That rule also decides what is not a card. Council strategy, direction, budget context, and the reasoning behind a decision are worth knowing, but they don't change your individual bill or your rights. That material becomes blog posts like this one, not cards. Keeping the two separate is what lets the card feed stay tight, factual, and personal.

A card is not a summary of a document. It is a single answer to a single question: what changes for you.

Why we built it this way

AI is genuinely good at reading a thousand pages in seconds and pulling structured data out of them — the part that defeats a human reader. It is also capable of inventing a plausible figure when a prompt is loose or a table is messy, which would be worse than useless on a document where a single wrong number misleads someone about their rates. The verification step and the human review exist precisely to contain that risk.

Other governments are applying AI to civic documents along similar lines — the OECD has documented how AI is used in civic participation and open government to lower the barrier to information rather than replace democratic engagement. The decisions don't change. The access does.

The promise we hold ourselves to is narrow and concrete: plain English, every number traceable to the adopted source document, and independent of any council. We summarize what councils have decided; we don't lobby, and we're not affiliated with any of them. You can read a card in ten seconds, and if you want to check it, the figure came straight from the council's own published plan.