AI chronology extraction from case documents.

What a source-linked chronology contains

A source-linked chronology is a sortable table in which every row is one event: the date as written in the document, the date as interpreted, what happened, who was involved, the document and page, a short supporting quote, and flags for uncertain dates or disputed facts. Alongside it sit a list of people and organisations with their aliases, a list of undated events, and a gaps report showing periods with no documents at all.

The links are the point. A lawyer who can click from any entry to the page it came from can verify 200 entries in an afternoon; a lawyer handed an unlinked summary has to reread the bundle.

How extraction runs

Inputs

The client bundle or discovery set (native emails with attachments, letters, contracts, invoices, minutes, scanned file notes, chat exports), the list of issues, and the parties with known aliases.

Systems

The firm’s document review platform or document management system, optical character recognition or a vision model for scans, an extraction model under the firm’s data terms, and a database or spreadsheet keyed by document ID.

Steps

Remove duplicates; thread emails and strip quoted history; read each page; extract candidate events with quotes; parse dates using New Zealand day-first conventions and each email’s sent date; merge the same event found in several documents; tag issues; sort and link.

Human review

A solicitor checks entries against sources before use and reviews a sample of pages for missed events. Counsel decides which events matter and how disputed facts are described.

Outputs

The chronology, a people list, an undated-events list, a gaps report, and pages the system could not read.

Measures

Share of entries a solicitor accepts without change, events found in the missed-event sample, and hours to a first usable chronology.

Risks

Wrong dates, invented connections between events, merged events that were separate, unreadable scans, and suppressed or privileged material handled carelessly.

Illustrative example: a Wellington employment dispute

Illustrative example, synthetic numbers: a 4,200-page set of 1,600 documents in a dismissal claim. A paralegal reading at about 40 pages an hour needs roughly 105 hours for a first chronology. The extraction produces 1,900 candidate entries, which merge to about 1,400.

The extraction line assumes 650 tokens of text and 900 tokens of instructions per page, with 150 output tokens, run through Anthropic’s Batch API at half price: Claude Sonnet 5.5 at US$1 per million input and US$5 per million output tokens, as listed on 19 October 2026. The merge pass sends the 1,900 candidate entries in overlapping batches, about 1 million input and 90,000 output tokens at the standard rates of US$2 and US$10 per million. Batch suits this job because nobody needs the chronology within the hour.

What the courts and the Law Society expect

• Keep sensitive material out of public tools. The courts’ guidelines for lawyers say not to enter private, confidential, suppressed or legally privileged information into generative AI chatbots, and to take reasonable steps to ensure the accuracy of information given to the court and to avoid breaching suppression orders. Tag suppressed names and documents before processing.
• Responsibility for what is filed. In December 2025 the Law Society reminded lawyers that New Zealand courts are already dealing with references to cases that do not exist, that responsibility stays with the person filing, and that AI output should be treated like an assistant’s work and checked thoroughly.
• Duty to the court. Rule 13.1 of the conduct rules sets an absolute duty of honesty to the court. A wrong date in a chronology becomes a wrong statement in a submission very easily.
• Confidentiality and privilege. Chapter 8 of the rules and the Law Society’s guidance apply to every provider that processes the bundle.

What counsel keeps

Counsel decides which events are material, how contested facts are characterised, and what the chronology is used for. Witnesses give their own evidence; the chronology helps test it, never writes it. Anything filed is checked against the source page by a lawyer.

Where chronology extraction goes wrong

• Day or month first. 03/04/2025 is 3 April in New Zealand and 4 March in many overseas systems. Keep the date as written, parse by document origin, and flag any date where both readings are valid.
• Relative dates. “Last Tuesday” in an email means nothing without the sent date. Resolve it against the message metadata and mark it as inferred.
• Quoted history in email threads. A reply quotes an earlier message, and the event is attributed to the wrong person and day. Thread messages and strip quoted text before extraction.
• Invented links between events. The model writes “following the meeting, the manager dismissed her” when no document says so. Require each entry to be supported by its own quote and check quotes automatically.
• Unreadable pages. Faint scans and handwriting produce confident nonsense. Route low-confidence pages to a person; our guide to OCR versus vision models explains the trade-off, and document processing covers the build.

The same discipline runs through the other legal workflows: contract clause review with source citations and legal intake and matter triage. See the legal sector guide for the wider picture, and checking AI answers against sources for training your team.

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