AI on Trial: Justice Lee's McDonald's Ruling and What It Means for Your Firm's Technology

When a Federal Court judge tells lawyers to get imaginative, it is worth listening

Justice Michael Lee has turned the $100 million wage underpayment class action against McDonald’s into a live test of whether artificial intelligence can bring runaway class action costs back under control. In late August 2026 he ordered McDonald’s Australia and the Shop, Distributive and Allied Employees Association (SDA) to confer, with the assistance of a “suitably qualified technology expert,” on using AI to collate payroll data, build chronologies and prepare for a ten week trial listed for 2028.¹

Similar orders followed the same week in the Virgin Australia investor class action, run by Corrs Chambers Westgarth, and the Blue Sky class action, run by Shine Lawyers and Banton Group, where document review costs had already climbed towards $20 million.²

What the McDonald’s orders actually require

The underlying claim alleges McDonald’s underpaid tens of thousands of store managers by not paying them for time spent opening and closing stores before and after rostered shifts.³ The proceeding, formally known in an earlier case management decision as Rogers v McDonald’s Australia, concerns managers who worked at corporate owned and franchisee owned restaurants between December 2017 and February 2020, and the Court has already made orders narrowing the initial trial to sample employees, sample employers and specified common questions.⁴ Before mediation, scheduled for May 2027, the parties must exchange rostering data, pay rates and classification records and turn that into a reliable chronology, described by Justice Lee as “dispersed throughout a miscellany of business records, apparently maintained in different repositories and in different forms.”¹

Justice Lee drew a careful line. He is “not asking the parties to become software developers,” and none of this assumes AI is a complete answer, “still less that the court should entrust to a machine any substantive adjudicative task.”¹ What he is asking for is a joint report and a draft protocol showing how AI could reliably perform the mechanical, high volume work, so that solicitors and the teams working under them are not billing enormous hours for tasks a properly supervised system can do at a fraction of the cost.

The timing is pointed. A BHP shareholder class action run by Maurice Blackburn and Phi Finney McDonald spent $48 million in legal costs without a trial ever being held, more than triple the amount originally projected to group members.⁵ That is precisely the kind of cost blowout Justice Lee wants case management to prevent in future matters.

Australia is catching up to overseas courts, not inventing this from scratch

Domestic commentary has understandably focused on how novel these orders feel in Australia. Courts in the United States and England and Wales, along with newer international commercial courts, have been building precedent and formal rules on exactly this question for well over a decade. Firms advising on the McDonald’s, Virgin and Blue Sky matters, or on any large document-heavy dispute in commercial litigation , do not need to start from first principles.

United States, February 2012. In Da Silva Moore v Publicis Groupe, Magistrate Judge Andrew Peck held that computer assisted review, including predictive coding, “is an acceptable way to search for relevant [electronically stored information] in appropriate cases,” approving a protocol built on a documented seed set, transparency with the opposing side and random sampling to verify accuracy.⁶ It was the first American judicial decision of its kind and it set a template still followed today: technology assisted review is approved because it is applied with quality control and proportionality, not because the technology merely exists.

England and Wales, February 2016. In Pyrrho Investments Ltd v MWB Property Ltd, the High Court approved predictive coding for the first time in England and Wales, finding that with more than three million documents identified for review, manual review would be prohibitively expensive and predictive coding was a proportionate alternative that was at least as accurate.⁷ That reasoning was later formalised in Practice Direction 57AD, which governs disclosure in the Business and Property Courts and defines Technology Assisted Review to expressly include predictive coding, requiring parties to justify not using it in any case involving more than 50,000 documents.⁸ That flips the presumption Justice Lee is nudging Australian class actions toward: technology assisted collation and review is the expected starting point in large document sets, and lawyers who skip it need a reason.

United Kingdom judiciary, refreshed October 2025. The Courts and Tribunals Judiciary has issued and repeatedly refreshed formal AI guidance for judicial office holders, covering accuracy checks, device security and how AI tools are used by litigants in person, and is separately consulting on rules for AI use in preparing court documents.⁹

Qatar International Court and Dispute Resolution Centre, 2026. After a lawyer was found in contempt of court for citing fabricated AI generated cases in Jonathan David Sheppard v Jillion LLC [2025] QIC (E) 3, the QICDRC released Practice Direction No. 1 of 2026 on the Use of Artificial Intelligence, requiring Court Users to verify all AI generated material, banning confidential or privileged information from unregulated AI tools, and setting out cost, evidentiary and contempt consequences for misuse.¹⁰

Read together, these sources show Justice Lee’s approach is consistent with, and arguably more cautious than, where overseas courts have already landed: use the technology to cut mechanical cost, but build in verification, transparency and human oversight from day one.

The liability question: what happens if AI review gets it wrong

This is the part that should concern every firm before it agrees to any AI protocol, whether ordered by a court or adopted voluntarily. The professional standard of care does not lower because software did the first pass. A recent briefing on solicitor liability in AI assisted due diligence put it plainly: delegating document review to an AI system is comparable to delegating it to a junior lawyer, and supervision and verification remain the solicitor’s responsibility. If review is defective because of an AI oversight, a court will ask whether relying on the output was reasonable and whether adequate human review took place. Simply pointing to “industry standard use” of the tool will not, by itself, be a defence.¹¹

The scope of that duty follows ordinary principles. In Manchester Building Society v Grant Thornton, the UK Supreme Court confirmed that a professional adviser is liable for losses falling within the scope of the responsibility it assumed.¹² A solicitor who certifies a chronology or an entitlement calculation as reliable has assumed responsibility for that output, however it was generated.

The most vivid illustration of what goes wrong without verification is Mata v Avianca. In June 2023, Judge P. Kevin Castel of the United States District Court for the Southern District of New York sanctioned two lawyers and their firm $5,000 after they filed a brief citing six entirely fabricated cases generated by ChatGPT, complete with invented docket numbers and fake quotations, and then continued to stand by them after the court questioned their existence.¹³ The court found the lawyers had acted with subjective bad faith and ordered them to notify their client and the judges falsely named as authors of the fake decisions. We have explored similar dynamics in our analysis of Judge Humphreys’ ChatGPT decision in Sydney and autonomous AI agent liability under Australian cyber and tort law . While generative drafting carries hallucination hazards that differ from technology assisted review, the fundamental legal principle is identical: the legal practitioner, not the vendor, answers for the output.

Is there still a case for manual review?

Yes, and every jurisdiction discussed above says so expressly. None of the courts approving predictive coding or AI assisted collation have dispensed with the underlying legal standard, which in Australia and England is a “reasonable search” carried out proportionately, not a perfect one.¹⁴ Under United States discovery rules, counsel must still make a “reasonable inquiry” into the effectiveness of any technology assisted process before certifying the result, because some fraction of relevant material will always be left behind and there is no universal recall or precision threshold that satisfies that obligation on its own.¹⁵ In practice this means:

• Sampling and validation of AI output against a human reviewed baseline, at the outset and at intervals through the review.
• Manual review of anything flagged as borderline, privileged or high risk, rather than trusting an automated confidence score.
• A documented, defensible protocol agreed with the other side before the technology is relied on, exactly what Justice Lee has ordered the McDonald’s, Virgin and Blue Sky parties to produce.
• A named solicitor who takes responsibility for sign off, because that duty cannot be delegated to a vendor or a machine learning model.

The Federal Court’s own Technology and the Court Practice Note (GPN-TECH) has permitted technology assisted approaches to document review since 2016, but it has never suggested a lawyer can certify accuracy without having tested it.¹⁶

Who counts as a suitably qualified technology expert?

Justice Lee’s orders lean on a specific phrase: the parties must confer “with the assistance of a suitably qualified technology expert.”¹ That is not a throwaway line. Engaging a generalist consultant who understands neither the litigation nor the underlying data architecture is exactly how firms end up back in court explaining why their AI protocol did not work, or worse, why privileged material leaked into a public model.

It is worth being precise about what “suitably qualified” means in Australia today, because no state or territory law society currently recognises technology law or data law as a formal accredited specialisation. The Queensland Law Society, for example, recognises ten areas of specialist accreditation: business law, commercial litigation, criminal law, family law, immigration law, personal injuries, property law, succession law, taxation law and workplace relations.¹⁷ There is no equivalent pathway to become an “Accredited Specialist” in technology or data law, the way a solicitor can in family law or commercial litigation. That means any claim of technology expertise in this space, including this firm’s, rests on genuine background and demonstrated experience rather than a formal credential the profession has agreed to recognise. It is a fair and important distinction to draw, and one the profession may need to address as these orders become more common.

With that caveat clearly stated, this is a genuinely rare combination of skills: someone who has actually built and run data systems at scale, who understands how enterprise information ends up “dispersed throughout a miscellany of business records… maintained in different repositories and in different forms” (Justice Lee’s description, and an accurate one of most large organisations’ payroll and HR systems), and who also understands privilege, discovery obligations and evidentiary reliability.

Andrew Bell, principal of Bell & Senior Lawyers, sits in that overlap. Before moving into full time legal practice, Andrew spent a sustained career as an IT architect, software developer, solution architect and head of data, including enterprise data leadership roles before founding Bell & Senior. He now practises in technology law and sits on the Australian Computer Society’s technical committee for AI and the Queensland Law Society’s data and technology committee. Again, that is relevant experience and current professional engagement with the subject matter, not a formally accredited specialisation, because no such accreditation yet exists in Australia. For firms already working with Bell & Senior facing a case management order in this mould, or any large dispute with a miscellany of records to synthesise, that combination of software architecture and litigation practice is worth having in the room before drafting a response to the court.

If your practice management or discovery software does not do this natively

Most firms’ existing case management, document management or eDiscovery platforms were not built with generative AI collation, chronology building or entitlement calculation as native features. If a firm is looking at bolting AI onto its existing stack, or evaluating a vendor’s new “AI powered” module, there are specific governance and technical questions to work through before any client or class member data goes near it.

None of this is exotic. It is the same discipline Justice Lee is asking the McDonald’s, Virgin and Blue Sky parties to document in their court ordered reports: do not simply assert that AI was used, show the protocol, show the controls and be ready to defend the reliability of the output. A firm that cannot answer the table above for its own systems is not ready to represent to a court that its AI assisted process is sound.

Why this matters beyond class actions

Justice Lee’s framing, that “our lodestar should be the use of AI to implement reforms which allow lawyers to spend more time exercising professional judgment and judges to spend more time identifying what really matters,” is not confined to representative proceedings.¹ Shine Lawyers’ head of class actions, Craig Allsopp, made a similar point, noting that the pressure to analyse “significant volumes of data” and calculate loss “across complex class actions involving thousands of group members” is only going to grow.¹⁸

Whether it is a $100 million wage claim, a commercial dispute or a startup’s due diligence dataset, the underlying problem is the same: useful information trapped in scattered, inconsistently structured records, and a court or client that needs a reliable answer quickly and at proportionate cost. The overseas precedent shows this problem is solvable, but only with the right combination of legal judgment, data architecture and documented process, and only if a qualified human remains accountable for the result. For wider analysis of emerging digital compliance issues, see our guide to ambient privacy and smart device regulations .

Frequently asked questions

Does Justice Lee’s order mean AI generated evidence or AI drafted documents can now be filed without review?

No. The judge was explicit that “none of this assumes AI provides the complete answer, still less that the court should entrust to a machine any substantive adjudicative task.”¹ Every jurisdiction discussed above—the UK, the US and the QICDRC—requires human verification of AI output before it is relied on or filed in court.⁹¹⁰

Is manual review dead once a court approves AI assisted review?

No. Approval of predictive coding or AI collation in Da Silva Moore, Pyrrho and the Federal Court’s own practice note has always sat alongside, not instead of, sampling, validation and manual review of high risk and privileged material.⁶⁷¹⁶

Is this only relevant to class actions?

No. The same cost and reliability problems arise in any matter with large, dispersed datasets, including commercial litigation, employment disputes, regulatory investigations and technology due diligence. The Federal Court’s willingness to order this kind of protocol in a representative proceeding is a strong signal for case management more broadly across Australian courts.

What happens if a firm gets its AI use wrong?

The contempt finding and Practice Direction from the QICDRC, and the $5,000 sanction and client notification orders in Mata v Avianca, show the downside clearly.¹⁰¹³ Getting the governance, confidentiality and audit trail right from the outset is far cheaper than remediating after the fact.

This article is general information about a developing area of legal practice and technology law and is not legal advice. If your firm has received a case management order requiring you to explore AI use, or you are evaluating AI tools for litigation or discovery, contact Bell & Senior Lawyers on (07) 5532 8777 to discuss your specific obligations and options.