Discovery with AI: 40,000 documents and the duty you still owe

By Michael Nadalin, Founder, Market Lead · 28 August 2026 · 8 min read

Forty thousand documents arrive from four custodians. Discovery orders are made, the timetable assumes you will manage, and the client has already asked what the review will cost before you have opened a single file.

Run the arithmetic. If a junior solicitor sustains fifty documents an hour, forty thousand documents is eight hundred hours of reading. That is five months of one person doing nothing else, or a team of four for most of a quarter, before anyone drafts a chronology. On most commercial matters that number does not fit the budget, the timetable or the amount in dispute.

So the question is not whether to put technology across the set. Firms have been doing that for years. The question is narrower and harder: which decisions do you hand to a machine, which ones stay with a solicitor, and what record do you keep so that you can stand up and describe the process if the other side or the court asks.

Run the arithmetic before you run the review

Most of the value in a large discovery exercise comes before anything clever happens. Exact duplicates, near duplicates, email threading down to inclusive messages, date range filters, custodian filters and system file exclusions will strip out a large share of a raw collection without any judgment call being made at all. Those steps are deterministic and reproducible, which is exactly why they are easy to defend.

The mistake is to skip that work and point a model at the raw set. You then pay for the model to read the same attachment forty times, and every validation number you produce is polluted by duplication. Do the mechanical reduction first, record what each step removed, and only then decide what needs a review decision.

The figures below are an illustration of the shape of a cull, not a benchmark. What matters is that you can produce the equivalent numbers for your own matter and explain each drop.

What the model decides and what you decide

The useful distinction is not between tasks a model can perform and tasks it cannot. It performs all of them, with varying reliability. The distinction is between decisions that are reversible and testable, and decisions that produce an irreversible consequence such as a waiver of privilege or a certificate signed by a solicitor.

Rank the work by how much of it you can genuinely hand over, and be honest about the bottom of the list. A model that produces a fluent summary of an email chain is not the same thing as a model that has understood the commercial significance of the chain, and it will not tell you the difference.

The obligations that actually bite

Nothing about technology assisted review changes the duties. It changes how easy they are to breach quietly. The failures that cause trouble are rarely dramatic: a criteria change nobody wrote down, a cloud service nobody read the terms for, a methodology disclosed for the first time in reply submissions.

The Australian courts are not hostile to this. Federal Court practice material and the Supreme Court of Victoria's guidance on technology in civil litigation both contemplate technology assisted review, and judges are generally more comfortable with a method that was put on the record early than one revealed under pressure. The gap is almost always in the record keeping, not in the permission.

Validating the cull is the whole job

Here is the part firms skip. Any ranking system can put plausible documents at the top of a queue. What you have to be able to say is something about the documents it did not put there, because those are the ones you are about to not produce.

That means sampling the discard pile, not just the review pile. Draw a random sample from everything the process is about to exclude, have a solicitor code it against the same criteria, and count how much relevant material was sitting in there. If the answer is uncomfortable, you retrain or you widen the net, and you do it before production rather than after a complaint.

Stop on a measured basis. Stopping because the deadline arrived is a defensible commercial decision and an indefensible discovery one, and the difference shows up in cross examination on the affidavit.

Agree the protocol before you press go

Most disputes about technology assisted review are not really about the technology. They are about a party being told after the fact that thirty thousand documents were excluded by a method they had no chance to comment on. That is an easy application for an opponent to run and an expensive one to defend.

Sequence the conversation properly and it largely disappears. Scope first, protocol second, method third, and all of it before processing costs are sunk.

What to do on the next matter that lands

Start with the mechanical reduction and record it. Fix the relevance criteria in writing before anyone codes a document. Decide, explicitly, which of the tasks on that ladder you are handing over and which you are keeping, and put the privilege review in the keeping column. Sample the discard pile. Write the narrative of the process while you can still remember it.

Do the vendor diligence once, properly, and reuse it. Where the documents sit, who can access them, whether your inputs are used to improve anyone's model, what happens on termination, and what your cross border position is. That review takes an afternoon and it covers every matter you run on that platform for the next two years.

Separate the two kinds of AI work in your head, because they fail differently. A review platform ranks documents you already hold, and its failure mode is a document you never see. A research tool like Legal Brain answers questions about Australian law and rules, and its failure mode is a confident answer that is wrong, which is why the source is always the point rather than the summary. Neither one signs the certificate. You do, and everything above exists so that when you sign it you can say precisely what you did and why it was reasonable.

Frequently asked questions

Can you use AI to review documents for discovery in Australian courts?

Yes. Australian courts have practice material that contemplates technology assisted review, including in the Federal Court and the Supreme Court of Victoria, and it is well established on large commercial matters. The expectation is that you can describe and justify the method, not that you avoid it. Raise it early rather than disclosing it after production.

Do I have to tell the other side that AI was used in the document review?

The safer position is yes, and early. Parties are generally expected to agree a document management protocol and to be transparent about the search performed, and a methodology raised at the first case management hearing rarely becomes a fight. A methodology revealed only after production is where interlocutory applications come from.

Is it a breach of confidentiality to upload client documents to an AI tool?

It depends entirely on the arrangement. Check whether inputs are used for training, where the data is hosted, who at the provider can access it, and what your obligations are under APP 8 if anything is disclosed offshore. Where the material is sensitive or the risk is material, get the client's informed instructions before uploading rather than after.

Can AI handle the privilege review on a large discovery set?

It can build the net. Flagging documents involving legal advisers, counsel and the usual privilege language will surface most candidates quickly. It cannot make the call, because an inadvertent production is a waiver problem rather than a statistical one, and privileged advice repeated inside an attachment or a spreadsheet is exactly what these systems miss.