AI Document Chronology for Australian Lawyers: Beyond Technology Assisted Review
by Sandlabs Team, Founder, Sandlabs
In 2023, Clayton Utz published a measured explainer on Technology Assisted Review — the AI-assisted approach to document sorting that became standard in large Australian eDiscovery matters. The framing was cautious, appropriate for the time: TAR can help rank documents by relevance, but treat it carefully.
Three years on, the technology has moved. The question for Australian legal practice is no longer whether AI can help sort a document set. It's whether AI can extract structured, verifiable facts from a scanned bundle and build a chronology a lawyer can actually rely on.
That's a different capability, and it's where the practical opportunity sits in 2026.
What TAR did — and what it didn't
Technology Assisted Review, as used in Australian eDiscovery, was built for a specific problem: large document sets in commercial litigation where the task is finding which documents are relevant. TAR uses machine learning to rank documents, reducing the volume a human reviewer needs to read.
What it doesn't do is tell you what happened. It surfaces relevant documents; it doesn't extract the facts from them, sequence them into a timeline, or tell you who was involved and when. That reconstruction still happened manually.
For litigation involving factual disputes — personal injury, employment disputes, AHPRA investigations, contract claims — the chronology is often the entire case. When did the injury occur? What treatment did the practitioner provide and when? What was communicated, to whom, and in what order? Getting this wrong is not a minor error.
Manual chronology work on a 600-page bundle takes a senior paralegal or junior solicitor one to three days. On a matter with multiple volumes, it can take a week. That's before the bundle gets updated.
What AI document chronology does
A properly built AI chronology system takes a scanned document bundle — PDFs, Word files, mixed formats — and produces a structured, cited timeline of events. Not a summary. A timeline where every entry has a date, a description, and a reference to the specific document and page it came from.
The process has distinct stages.
Digital text detection. Many PDFs contain embedded digital text — typed letters, electronically generated reports. These are read directly, without OCR, which is faster and more accurate. The system identifies which pages already have digital text and which are genuine scanned images, then processes them accordingly.
OCR for scanned pages. Pages without embedded text go through optical character recognition to generate a text layer. Modern vision-language models handle reasonable scan quality well. Pages with poor image quality — faded, rotated, heavily degraded — are flagged rather than silently processed at low accuracy.
Fact extraction. With text available, the AI reasoning layer extracts structured facts: dates, parties, organisations, events, and — critically — the source page for each. Large language models are well-suited to this. They understand professional language, legal and medical terminology, and the conventions of correspondence, reports, and formal notices.
Verification. A verification step checks each extracted fact against the source text. If the AI has summarised or inferred something that can't be directly cited in the original, it's flagged for human review. Facts that can't be verified don't appear in the output as established — they appear as items requiring review.
Parallel processing and stitching. Large bundles are split into sections and processed concurrently, then merged into a single timeline. Cross-document references — a letter referencing an earlier report, a medical note updating a previous diagnosis — are connected where identifiable.
The result is a searchable chronology covering the entire bundle, each entry linked to its source.
Where this applies in Australian legal practice
Personal injury and CTP
PI litigation in Australia — particularly under NSW CTP and equivalent state schemes — generates dense document bundles: medical records from multiple practitioners across years, workplace documents, insurer correspondence, expert reports. The factual chronology of injury, treatment, and consequences is often the substance of the dispute. Building that chronology from a 1,200-page bundle is exactly the task this technology addresses.
AHPRA complaints
A practitioner facing an AHPRA investigation may have years of clinical records across multiple facilities, correspondence with the Board, and responses from colleagues. Their legal representative needs to reconstruct a complete factual picture quickly. The bundle is typically scanned, mixed-quality, and voluminous. A verified AI chronology reduces the time to first legal advice from days to hours.
Regulatory investigations and ASIC notices
Commercial disputes involving regulatory notice responses — ASIC examinations, APRA inquiries, ATO investigations — frequently involve large bundles of internal corporate records: emails, board papers, contracts, financial records. Building a factual timeline of what was known, decided, and communicated — and when — is the core reconstruction task. The same AI chronology architecture applies.
Coronial inquests
Coronial matters often involve years of records across healthcare providers, emergency services, and government agencies. Counsel assisting and legal representatives for interested parties both need rapid factual reconstruction from voluminous, mixed-quality records.
The traceability requirement
For legal use, the most important property of an AI-generated output is not accuracy in the abstract. It's traceability. A lawyer cannot rely on a summary they can't verify. They can rely on an extracted fact that cites a specific document and page number.
This distinction drives the architecture of any system built for legal use. The output isn't "the injury occurred in March 2019." It's "injury occurred March 2019 (Volume 2, p.47, treating practitioner report, Dr A Smith)." The lawyer can verify that in seconds. They can cross-examine on it. They can cite it in submissions.
An AI system that produces unverified summaries is a liability. One that produces cited, verifiable facts is a tool.
The verification step also catches hallucination — the tendency of large language models to fill gaps with plausible-sounding content. If an extracted fact can't be located in the source text, it doesn't appear in the chronology. That's the difference between AI assistance and AI risk.
Australian data considerations
Most AI chronology tools with meaningful capability are US products (Mary Technology, Legalyze, Superinsight, DigitalOwl). They're built for US personal injury law firms and hosted on US infrastructure.
For Australian practitioners handling documents subject to client legal privilege, AHPRA regulatory matters, or personal health information, sending documents to a US-hosted SaaS product raises Privacy Act and confidentiality considerations that most firms aren't prepared to work through.
A custom-built system hosted on Australian infrastructure — AWS Sydney, Azure Australia East — avoids this entirely. The documents don't leave the country. The client's privilege is not exposed to a foreign service agreement.
Human review remains the standard
AI document chronology is a reading and extraction tool, not a decision-making tool. The practitioner reviews the chronology, identifies inconsistencies, determines what matters, and makes the professional judgement. The AI handles the volume. The lawyer handles the analysis.
This framing also matters for professional obligations. A solicitor who has reviewed an AI-generated chronology and can speak to its contents is meeting their professional standard. One who has relied on unreviewed AI output is not.
Frequently Asked Questions
Getting started
If your practice handles voluminous document bundles and chronology work is taking days per matter, this is a practical problem with a practical solution. The first step is understanding your document types and current workflow — from there we can scope what's achievable.