Discovery is the part of litigation that eats budgets and calendars. A mid-sized commercial dispute in Sydney or Melbourne can generate tens of thousands of documents across email, chat logs and shared drives, and every one of them needs a first look before anyone can talk about relevance or privilege. Claude is now doing meaningful work in that first pass for a handful of Australian firms and in-house legal teams, and it is worth being precise about what it does well, what it doesn't touch, and what the numbers actually look like.
The discovery bottleneck in Australian litigation
Most litigation support teams still run first-pass document review the same way they did a decade ago: graduate lawyers or contract paralegals working through a document review platform at a set hourly rate, applying tags for relevance, privilege and confidentiality one document at a time. It is careful work, and it has to be, because a missed privilege call or a mishandled confidential document creates real professional risk under the Legal Profession Uniform Law and the firm's own duty of care to the client.
The problem is volume and deadlines. Federal Court of Australia practice notes and case management timetables give parties a fixed window to complete discovery, and that window does not stretch just because a matter has 60,000 documents instead of 6,000. Firms respond by throwing more reviewers at the problem, which is where the cost blows out.
Document volume that scales faster than headcount, especially once chat platforms and shared drives are added to the custodian list
Privilege calls that require careful, contextual reading rather than keyword matching
Multiple custodians with overlapping but not identical document sets, which creates deduplication headaches
Format variety: PDFs, scanned images, spreadsheets, and export files from half a dozen internal systems
Court-imposed deadlines that don't move for the size of the document set
Where Claude fits in the discovery workflow
The useful role for Claude sits in the first pass, before a lawyer's eyes are needed on every document. Fed a batch of discovery documents through a connector or a Files API workflow, Claude can flag likely duplicates, group documents by topic and chronology, and surface a probability estimate for relevance and privilege risk based on the content and context of each document, not just keyword hits.
That matters because Boolean search misses a lot. A document that discusses a settlement figure without ever using the word "settlement" won't get caught by a keyword list, but Claude reading for meaning will usually flag it for a closer look. The output is not a decision, it's a prioritised queue: documents most likely to need a lawyer's attention first, documents most likely to be genuinely irrelevant pushed to the back.
What Claude does not replace
The privilege determination stays with the solicitor. That is not a workaround, it is the correct design. Legal professional privilege is a substantive legal judgment tied to the specific dominant purpose of a document, and Australian courts expect that call to be made, and be defensible, by a lawyer who understands the matter. Claude-assisted triage narrows what a lawyer has to look at closely; it does not sign the privilege log.
There is also a confidentiality dimension under the Privacy Act that firms need to think through before any client documents touch a third-party tool: data handling terms, where processing happens, retention settings, and whether the matter itself involves data too sensitive to run through any AI system at all. For some family law and criminal matters, the answer is still that it doesn't go near a model, and that's a reasonable line to hold.
A realistic cost picture
Take a mid-sized commercial matter with 40,000 documents after deduplication. At a blended paralegal and junior lawyer rate, a full manual first-pass review of that set can run close to $120,000 before a single senior lawyer hour is billed on top. Firms running Claude-assisted triage on comparable matters report cutting the number of documents that need full manual review by roughly 30 to 40 percent, mostly by correctly identifying near-duplicates and clearly irrelevant system-generated material. On a $120,000 first pass, that is a genuine five-figure saving, not a rounding error, and it shows up before the client even asks about the invoice.
The saving is real, but it depends on the firm keeping its own quality controls in place. A batch of triaged output is only useful if someone audits a sample of the "likely irrelevant" pile before it is excluded, and that audit step should be built into the workflow from day one rather than added after a client asks how the process actually worked.
Keep a full audit trail of what Claude flagged and why, in case it needs to be explained to opposing counsel or the court
Have a partner or senior associate sign off on the privilege log, every time, with no exceptions for AI-assisted batches
Confirm data residency and retention terms before any client documents are processed
Sample-audit the excluded pile, not just the included one
Getting started without disrupting an existing workflow
The firms that have had the smoothest rollout didn't switch their whole discovery process over on day one. They ran Claude-assisted triage in parallel with the existing manual process on one closed matter, compared the two sets of tags, and measured the agreement rate before trusting it on a live file. That gave the partners a number to point to internally, and gave the review team confidence that the tool was catching what mattered rather than quietly missing it.
If your firm is weighing this up, the sensible next step is a scoped pilot on a matter with a known outcome, not a live discovery deadline. We help Australian firms set that pilot up, including the data handling review, in a short paid engagement. You can book a time through our booking page to talk through what a pilot would look like for your review volumes.
