Augmented intelligence is software that helps a person do a task, rather than software that replaces the person. In legal work the distinction matters. Fully autonomous artificial intelligence would take a document, decide what to do with it and act, with no human in the loop. Augmented intelligence, sometimes called intelligence augmentation to keep it distinct from AI, keeps the human in the loop: the software does the high-volume pattern work, and the lawyer sets the standard, trains the system and answers for the result.
For an in-house legal team the two places this shows up most often are discovery in litigation, through software called predictive coding, and contract review. Both follow the same underlying design: a machine that processes documents at a scale no team of lawyers could match, supervised by a lawyer who remains responsible for the outcome. This article explains who does what in that loop, how each type of software actually operates, the conduct rules that still apply, the pitfalls that most often cause problems, and when external help is worth getting.
The players and what each one does
The first step in understanding augmented intelligence is mapping the roles. There are five:
- The in-house lawyer: sets the definition of what is relevant or acceptable for the matter, trains the software by example, reviews its output and carries the professional responsibility for the final work product.
- The software: applies the rules it has learned across thousands of documents without tiring, but it only reproduces what it was taught.
- The software provider: builds the tool, supplies pre-loaded rules and may run the training; its claims about accuracy need to be tested rather than assumed.
- The courts: where augmented intelligence is used in litigation discovery, the court controls the process and may require a validation protocol before technology-assisted review is used.
- Professional bodies: the Law Council of Australia and the state and territory law societies publish guidance confirming that the existing conduct rules apply to automated work.
Predictive coding in discovery
Discovery is the stage of litigation where each party must identify the documents relevant to the dispute and make them available to the other side. In large commercial and intellectual property matters the document set can run to millions of records. In McConnell Dowell Constructors v Santam (No 1) [2016] VSC 734, the Supreme Court of Victoria dealt with a discovery exercise that began with at least 1.4 million documents. Reviewing a set like that manually, document by document, is slow and expensive, and it is exactly the job augmented intelligence was built for.
Predictive coding operates in a loop. First, the lawyer reviews a sample of the documents and codes each one as relevant or not relevant to the dispute. The software analyses that coded sample, looking at the content and metadata of the documents, and works out which characteristics tend to indicate relevance: particular words and phrases, relationships between names and dates, document types. It then scores the whole document set by predicted relevance and presents the lawyer with the documents most likely to be relevant, together with the borderline ones where its confidence is low.
The lawyer reviews those documents, corrects the coding where the software has it wrong, and feeds the corrected decisions back in. Each round sharpens the model. This is what is meant by supervised machine learning: the lawyer supervises the software's learning, and the software gets more accurate as the rounds continue. Once the model's performance stabilises within an acceptable range, tested against documents the lawyer has already reviewed, the review is complete and the relevant documents can be produced.
Australian courts have accepted this process when it is run transparently and validated. McConnell Dowell was the first decision in which an Australian court ordered the use of technology-assisted review, and a follow-up decision in the same proceeding, McConnell Dowell Constructors (Aust) Pty Ltd v Santam Ltd [2017] VSC 640, considered how the validation of the process should be conducted. The Federal Court has since dealt with discovery applications involving predictive coding with continuous active learning technology, for example in ViiV Healthcare Company v Gilead Sciences Pty Ltd (No 2) [2020] FCA 1455. The consistent thread is that the court is not reviewing the software's output blindly. It is reviewing the process: what the software was trained on, how it was validated and who supervised it.
Contract review software
The other common use of augmented intelligence in an in-house team is contract review. Contract review software typically does three things. It classifies clauses, so it can identify a change of control clause or an assignment clause across a large contract portfolio. It extracts information, pulling out dates, renewal periods, termination rights and price terms into a structured summary. And it flags clauses that are inconsistent with the organisation's standard position, so the lawyer knows which clauses need attention.
The software can work in two ways. Some products come with pre-loaded rules for certain types of clauses with consistent phrasing. Others learn from the organisation's own documents: the lawyer marks up what an acceptable clause looks like, and the software identifies patterns across the portfolio. The distinction matters because it determines what the team is actually relying on. With pre-loaded rules, the team inherits the vendor's assumptions about what a standard clause looks like, which may not match the organisation's risk appetite. With documents the team trains itself, the output reflects the quality of what the organisation has actually negotiated over the years. Either way, a missed inconsistency is still the lawyer's risk, not the software's.
The conduct rules that still apply
Using software does not change the standard of professional conduct. The Australian Solicitors' Conduct Rules, which apply to solicitors in the Uniform Law jurisdictions and in equivalent form across Australia, require a solicitor to deliver legal services competently, diligently and as promptly as reasonably possible (rule 4.1.3). They also require a solicitor with designated responsibility for a matter to exercise reasonable supervision over the solicitors and other employees engaged in providing the legal services for that matter (rule 37.1). An automated review tool sits squarely inside that supervision duty: the lawyer must understand what the software does, ensure it has been properly trained and validated, and review enough of its output to be satisfied the work is right.
The profession's guidance bodies say the same thing. The Law Council of Australia maintains a central portal of guidance for lawyers using AI and warns that improper use may carry ethical, professional and legal risks. The Law Society of New South Wales has published a guide to the responsible use of AI that works through the accuracy, bias, intellectual property, privacy and data security issues that arise, and confirms that the existing conduct rules, including rule 4, still apply. The Uniform Law regulators, together with the Legal Practice Board of Western Australia and the Victorian Legal Services Board and Commissioner, have issued a joint statement on the use of AI in Australian legal practice. The consistent message is that automation does not lower the bar.
There is a second, less obvious reason the supervision duty matters. Software never tires and never gets distracted, but it can compound human error. If the training sample is coded wrongly, or the software's limitations are misunderstood, the mistake is reproduced across every document in the set at a speed no human could match. The efficiency gain and the risk come from the same source: scale.
Quality control and the common pitfalls
Because the software works at scale, quality control has to be built into the workflow rather than bolted on at the end. The accepted approach, and the approach the courts look for in the discovery context, is a validation loop: spot check the input the software is trained on, spot check the output, and keep testing the model against documents the lawyer has already reviewed until its accuracy is within an acceptable range. In the 2017 McConnell Dowell decision the court examined exactly this kind of validation process before approving the technology-assisted review.
The pitfalls that most often cause problems are:
- An unrepresentative training sample: If the documents used to train the model do not reflect the wider document population, the software learns the wrong patterns and confidently reproduces them.
- Unsupervised learning on unstructured documents: Some newer tools identify relationships in documents without guidance. That is useful for exploring a set the team knows little about, but the patterns they find still need a human to test whether they mean anything.
- A black-box provider: If the vendor cannot explain its assumptions, methodology and the software's limitations in plain language, the lawyer cannot supervise the process or manage the risk.
- Confidentiality and data security: Feeding client documents and privileged material into third-party software means the material leaves the organisation's control. It is worth checking where the data is stored, who can access it and what happens to it if the arrangement ends.
- No review at all: Reviewing every output defeats the purpose of the software, but reviewing none of it is how errors at scale go unnoticed. Spot checks are the middle path.
Where a lawyer can help
Most of the work described above can be done by an in-house team, but there are points where external legal help is worth getting early. Vendor selection is one. An external lawyer can help develop the selection criteria, test the vendor's claims and review the contract: the service levels, the data security commitments, ownership of any trained models, and what happens to the organisation's data if the arrangement ends.
In litigation, a technology-assisted review protocol may need to be negotiated with the other side and approved by the court, and a lawyer who has run this process before knows what the court will want to see in the way of training and validation. At the governance level, a lawyer can help put the policy framework in place: an AI use policy for the team, a record of how each tool is trained and validated, and a documented supervision loop that shows the professional obligations have been met if the work is ever challenged.
The question that decides whether the automation is defensible
Everything in this article comes back to one question, and it is worth asking it of every tool before it is adopted: if the software makes the call, who can explain and justify it? Augmented intelligence delivers its gains precisely because it moves decisions from the lawyer to the machine. The machine is fast, consistent and tireless, but it is only as good as the training and supervision behind it, and the professional responsibility never moves with the task.
The most expensive mistake is treating augmented intelligence as a technology implementation when it is really a change in how legal work is supervised. The cheapest way to find out where an organisation stands is to run a small pilot on a completed matter, document how the software was trained and validated, and test whether the team can explain every automated decision it produced. If it can, the tool is genuine leverage. If it cannot, the tool is a liability, and no vendor warranty fixes that.