Faster and cheaper? Maybe: why AI in discovery can cost more before it costs less

Will Atfield, Owen Bourke, Luke Furness, Grace Ness, Anita Burkart and Mackenzie Janes
17 Nov 2026
6 minutes

The case for using AI in litigation, particularly in discovery, usually rests on two promises: faster work and lower fees. Both are achievable when the data suits the technology and the exercise is properly planned. But organisations that bring AI into discovery without first understanding their data, the limits of their chosen tool and the right approach for the matter can get the opposite: higher costs and longer delays. This article explains why that happens and how to avoid it.

In September we compared the generative AI practice notes across six Australian jurisdictions and examined Rogers v McDonald's Australia Ltd (AI-use) [2026] FCA 1264 (Rogers) where Lee J used section 37P(2) of the Federal Court of Australia Act 1976 (Cth) to direct the parties to investigate whether AI could be used, consistently with four requirements: traceability, testability, confidentiality and human supervision.

This article asks "What does it take to deliver the promised savings in practice?"

Key takeaways

  • Following Rogers, parties in Federal Court proceedings should be prepared to explain whether AI could reduce the time and cost of discovery.

  • AI-assisted review shifts costs to the start of a matter. The savings come later, and only if the early work is done properly.

  • AI-assisted review presents two distinct risks: inaccurate outputs and missed material. Each needs its own controls, and missed material is harder to detect.

  • Before committing to an AI approach in court, establish what the tool cannot process; how its results will be validated; how missed material will be identified and handled; and who bears the risk if something important is missed .

The Federal Court is increasingly asking the AI question

Rogers is not an isolated case. A day later, in R&B Investments Pty Ltd (Trustee) v Blue Sky Alternative Investments Limited (in liq) (Security for Costs and AI) [2026] FCA 1265, and a week later, in Matheson Property Group Pty Ltd (Trustee) v Virgin Australia Holdings Limited (NSD 346/2022), the Court made similar orders. In each case, the parties were required to confer with a technology expert on whether AI could reduce the time and cost of high-volume tasks, while maintaining, or improving, consistency and quality.

The potential of AI is easiest to see in class actions and other disputes with large discoveries. However, the efficiencies materialise only if the discovery has been properly mapped, the methodology reflects the nature of the documents, and appropriate validation standards are agreed before the tools are run.

Before applying the four Rogers requirements, a party needs to answer a more basic question: which of its material can AI reliably review, which can it not, and does the party have the right people, processes and technology to deal with both? A party that commits in court to an AI approach before answering that question may find it cannot honour the commitment.

Hurdle one: cost moves to the front

Traditionally, much first-pass review was undertaken by junior lawyers and paralegals. Senior lawyers were involved mainly at second-line review, and the cost of discovery was incurred gradually over many months.

AI-assisted review changes this cost profile. Instructing an AI tool well requires a deep understanding of the matter from the outset, because changing course later means re-running the work. A sound approach also involves testing prompts, sampling results and validating outputs before the tool is run across the full review population. A phased budget that treats discovery as a single line item will therefore show higher costs early, even though the savings arrive in later phases.

Hurdle two: rushing into "just use AI" without recognising the traps

Three common traps are set out below:

  1. Poor scoping: The business says "just run it through AI". No one clearly defines what the tool is being asked to do, the risk assessment is inconclusive, and responsibility for the outcome falls on the general counsel. The results are inconsistent, and the review ends up taking twice as long as a traditional review would have.

  2. Poor data hygiene: A significant proportion of a typical collection may not be suitable for AI review. Specialised file formats, password protection, document corruption, poor-quality scans and image-heavy documents all need to be triaged and dealt with separately. Avoiding this starts with mapping how the organisation creates and stores information across email, messaging platforms, collaboration tools, document management systems, databases and legacy archives. Each source produces data in different forms, with different metadata and gaps. Without that map, problems surface late.

  3. Indefensible discovery: A capable junior lawyer uses an AI tool for retrieval and discovery. The general counsel understands the workflow broadly, but not enough to test its limitations. The other side challenges the discovery and asks how it was done, but no meaningful explanation can be given. If the tool never accessed a source, nothing in its output will reveal that the source was missed. If no record was kept of the tool's inputs, instructions and decisions, its results cannot be meaningfully tested. These are the risks the traceability and testability requirements in Rogers are intended to address.

Consider a company defending a large class action involving hundreds of thousands of documents. It adopts an AI-assisted review process on the assumption that the tool can analyse the entire collection. Several weeks into the review, the legal team discovers that thousands of documents, including attachments and scanned records, were never processed. Those documents must now be reviewed separately, while the team reassesses whether the original review was sufficiently comprehensive. The additional work delays discovery and erodes much of the anticipated saving. The AI's results were accurate; the failure was that no one had established what material it could review before the process began.

Hurdle three: what the tool doesn't see

Discussion of AI risk often focuses on hallucinations and other errors in what the tool produces: an inaccurate summary, a misclassified document or a fact attributed to a document that does not contain it. In discovery, there is a second, quieter risk: the tool may not consider some documents at all. It may be unable to analyse them, documents may have been split in a way that loses context, the instructions may not anticipate how a concept is expressed, or the tool may simply fail to surface them.

The two kinds of error call for different controls. Errors in outputs are managed by tracing outputs back to their source documents, having human reviewers verify key findings, and sampling outputs against the underlying material. Omissions are harder to detect, because there is no output to check. They are managed by reconciling what was collected against what was actually imported into the tool, reporting exceptions, sampling documents the tool classified as non-responsive (an "elusion test"), and estimating recall (the proportion of relevant documents identified) using a validated sample or reference set. A defensible methodology addresses both risks, records the controls applied and documents how the results were validated.

What to do about it

  1. Find out what the tool will miss before committing to an approach. Ask your lawyers and legal technologists to explain the recommended approach, its expected reliability, how material outside the AI process will be handled, and who is accountable if material is missed.

  2. Agree the controls up front. Settle validation measures and acceptable error thresholds before the tool is run at scale.

  3. Settle the methodology early, then keep it under review. Decide the approach in the first weeks of the matter, before the court timetable turns any change into a hurried rebuild. Reassess it as the matter progresses and the technology improves.

  4. Budget for how the work actually runs. Expect higher costs in the opening stages, and ask for a projection of when and where the savings will appear, rather than relying on a single headline figure.

  5. Use Rogers as a guide. The orders identify three tasks: synthesising the facts and issues in dispute, preparing chronologies and summaries of documents, and managing documents more generally. Ask how each would be done in your matter and how the approach would meet the four requirements.

Where to from here?

Rogers is still unfolding. The approach the parties adopt, and Lee J's response to it, are likely to become a benchmark. Because the section 37P(2) power is available in Federal Court proceedings generally, this is unlikely to be the last direction of its kind. The practice notes set a baseline, but individual judges may expect more.

AI can deliver substantial savings in discovery, but those savings depend on decisions made well before the review begins. For general counsel, the real question is whether the proposed approach is reliable and defensible, and when the savings will actually arrive.

Parties should be able to answer that before committing to an AI-assisted process, particularly where the commitment is made to the court. Getting the approach right at the outset is far cheaper than unwinding it mid-matter.

Be the first to know

Legal updates and articles on the topics that matter to you, sent as they happen.

Subscribe

Get in touch

Disclaimer
Clayton Utz communications are intended to provide commentary and general information. They should not be relied upon as legal advice. Formal legal advice should be sought in particular transactions or on matters of interest arising from this communication. Persons listed may not be admitted in all States and Territories.