AI in legal practice: how to protect your professional judgement

By Sonja Cilliers and Maryke Swarts, PMRI · 7 October 2026

There is a particular quality to the hour before a founding affidavit is due, and most legal professionals will recognise it without needing it described: the coffee going cold at the elbow, the client sending voice notes instead of instructions, the sense of a day already spent before it has properly begun. What is new is the AI tool on the desk that will produce eleven pages of competent legal prose in rather less time than it takes to reread the notice of motion. A great many legal professionals now use it, and there is nothing in that to apologise for.

The draft comes back and it is good: fluent, well organised, entirely reasonable. It is also, somewhere around paragraph nine, quietly and confidently wrong, though not in the obvious way that would be a mercy, since obvious errors announce themselves and get corrected. It is wrong in the manner that reads perfectly well until it is reread the night before the hearing, at which point the argument shifts underneath the paragraph and the paragraph stops holding.

Whether the error is caught depends on something the profession has never really given a name to, that no firm has ever put a line item against, and that all of us are now relying on rather more heavily than we were three years ago. It is the legal professional’s accumulated pattern store, the consolidated record of every matter previously worked through, held in long-term memory and matched against each new set of facts faster than deliberate reasoning can operate. It is what allows an experienced legal professional to feel that something is off well before they can explain why. This capacity is now the most valuable thing a legal professional can own because it is the one part of the work AI cannot supply, and it is also what is most quietly threatened by the way the tools are being used.

What actually became scarce when AI arrived in legal practice

For most of the history of this profession, legal knowledge was the capacity limit. Finding the authority took time and reading it properly took considerably longer, so a firm’s competitive advantage was substantially a function of who had read what. The commodity was legal expertise, and the profession built its hierarchies, its fees and its self-image around accumulated knowledge.

Information is now commonplace and becoming more effortless to obtain by the month. The machine has read more than any of us will and will summarise it on request, with citations, in whichever register is asked for. This is genuinely useful and there is no honourable version of pretending otherwise.

What has grown scarcer as everything around it has multiplied is judgement. Judgement is the capacity to see what a proposition will actually do once it meets a particular set of facts, a particular opponent and a particular bench. It is noticing that an answer is technically correct and practically useless. It is weighing two defensible positions, neither of which is wrong, and knowing which one will hold. It is seeing which of the client’s eleven complaints the matter turns on, and which three should never be put in writing at all. None of that is technical legal knowledge. All of it is the residue of experience, and none of it is available for download.

This distinction is already written into the profession’s response to the technology. Courts and regulators have converged on the same formulation, that the tool may assist but may not substitute for professional judgement. What that formulation also does, and what has attracted markedly less attention, is allocate risk. It places the liability of a powerful technology onto a single human capacity, which the profession has never had cause to examine until now.

Where professional judgement actually comes from

Judgement feels like intuition, and it behaves in practice a great deal more like a filing cabinet. Every matter a legal professional has worked through is somewhere in there, and the reason they can look at a set of facts and know at once where the difficulty will be is that the cabinet answers first.

The research on expertise is remarkably consistent across every field in which it has been examined, from chess to firefighting to emergency medicine. Experts do not reason through more possibilities than novices. They recognise more. Decades of exposure lay down an enormous store of situational patterns, and the closest match surfaces well before deliberate reasoning has got under way.

In legal work this happens constantly and mostly below conscious notice. A legal professional reads an opposing affidavit, and something in the sequencing of the annexures says the deponent is concealing facts. They cannot yet say where, but they know to go looking, and they are usually right. That is a pattern match against several thousand prior encounters, most of which they could not now describe.

The useful question is how that store gets built, because once the mechanism is understood it becomes possible to say when it is being interfered with. Here the memory research is unusually settled. Material a person generates is retained substantially better than material the same person merely reads (J Schindler, S Schindler and M-A Reinhard ‘Effectiveness of self-generation during learning is dependent on individual differences in need for cognition’ (2019) 7(2) Frontline Learning Research 23).

The difficulty a legal professional experiences while drafting complex papers is the part doing the encoding. Effortful engagement obliges the brain to construct the connection rather than recognise one that has been handed over, and constructed connections are the ones that survive. The struggle, in other words, is the filing mechanism rather than a sign that the legal professional does not have the legal knowledge.

Which is why the question was never whether to offload but what to offload. Handing over work that was never building anything is a straight gain: formatting, first pass summarising, turning ninety pages of transcript into a chronology. It only becomes a liability when it removes the work that was building the pattern, meaning forming a view, choosing the framing, deciding what matters and what is noise. Both arrive through the same interface at the same speed, so that summarise this regulation and tell me how to frame this application are typed with the same fingers thirty seconds apart. The first costs nothing but the time it saves. The second costs a repetition, and repetitions are what the store is made of.

In isolation the cost is negligible. Accumulated over a career it produces a legal professional whose store has not grown at the rate their experience implies.

The fluency problem: why a polished AI draft slips past review

A second effect is doing damage now rather than in ten years. Reading a fluent piece of writing produces the experience of fluency itself as comprehension. Nothing snags, and the brain interprets that absence of friction as evidence that the reasoning underneath is sound. This gap between the feeling of understanding and actual understanding is indifferent to seniority. It is exactly why a confidently written wrong paragraph can survive three rounds of review inside a competent firm and then die in twenty seconds in court. The reviewers were reading for friction, and the draft had been engineered to have none.

An AI draft supplies a polished surface over a thought that may never have been had at all, so the reviewing skill required is not the same skill, and very few of us have adjusted how we review. This is measurable. A survey of 319 knowledge workers drawing on 936 first-hand accounts of using generative tools at work found that confidence in the tool predicts less critical thinking, while confidence in oneself predicts more. It also found that the tool shifts the nature of the thinking towards verification, integration and oversight of what has been supplied (H-P Lee, A Sarkar, L Tankelevitch, I Drosos, S Rintel, R Banks and N Wilson ‘The impact of generative AI on critical thinking’ (2025) Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems Article 1121).

All of which quietly assumes a rested reader, an assumption that is difficult to make in this profession. Scrutiny is metabolically expensive, and a depleted system conserves energy by taking the fluent option. A tired brain is a fluent draft’s ideal customer, and the draft does not have to be persuasive to win. It only has to be smooth. Scrutiny under load is a trainable capacity, and the legal professionals who can still hold it at 16h40 are better conditioned rather than more conscientious. (We wrote about what sustained load does to the legal mind in What the profession will not name.)

Take a position first: a three-step method for using AI in legal work

As these tools advance, the ability to hold attention, memory and scrutiny under load becomes more valuable, and it can be trained deliberately. The smallest place to begin, and the one that survives a full day in practice, is to take a position first.

Position. Before the tool is opened, the legal professional writes their own view in three sentences. Rough is fine, half formed is fine, and wrong is arguably best of all, because forming a wrong view and discovering exactly where it failed is one of the most efficient forms of encoding available to anybody.

Prompt. Now the tool is used, properly rather than apologetically. The legal professional is no longer reading into a vacuum but against a position already held, which is a different cognitive operation and produces a different quality of attention. Where there is a divergence between the view and the draft, the divergence surfaces immediately, because there is something for it to collide with.

Pressure test. Before reading the draft closely, the legal professional writes down the three questions the matter actually turns on. Then they go looking for the answers in what the machine produced. Questions formed in advance survive fatigue in a way that open-ended scrutiny does not, because they do not depend on anything snagging, and the answers are either there or they are not. Where a question cannot be answered from the draft, the answer is in the primary sources.

What is actually at stake

The profession spent two years worrying that AI would take the work. The more precise worry is that it will take the practice. Not practice in the sense of the firm or the office, but practice in the older and plainer sense: the repetitions through which competence is made and then maintained. Every other demanding discipline accepts that its central capacity must be trained deliberately and kept in condition. Law never had to think that way, because the repetitions arrived on their own, built into how the work was done. They no longer arrive on their own, and what was a by-product of practice now has to be chosen.

Which is why, at 07h50 on a Tuesday with a founding affidavit due at noon, the ninety seconds spent writing down your own view before opening the tool matter more than anything else in the morning. It is the one part of the day still building the legal professional you are becoming.

How PMRI works with firms on this

Every firm, chambers and legal department is now deciding, mostly without noticing, how its people will use AI, and those decisions shape the next generation of judgement far more than any single policy document. PMRI explains the mechanism first, because legal professionals do not change how they work until they understand why their own mind behaves as it does. The most valuable step is a conversation with us about your practice: how the tools are being used, by whom, and where the review is quietly thinning. From there PMRI spends time at the firm, in whatever form fits, from a single session to a full programme, works with the leaders who decide how juniors are trained when the first draft writes itself, and speaks at meetings, forums and year-end functions. Start a conversation with us. There is no charge for the first one.

For a legal professional who would rather test the water first, The Complete PMRI Webinar Series is the strongest place to begin: four recorded two-hour sessions with their workbooks, covering burnout and its progression, productivity under load, goal setting for the legal year and the leadership dilemma in law, which together build the attention, memory and scrutiny under load that this article describes.

Frequently asked questions

Should legal professionals use AI tools at all?

Yes. Handing over work that was never building anything, such as formatting, first pass summaries and chronologies, is a straight gain. The risk lies in handing over the work that builds judgement: forming a view, choosing the framing and deciding what matters.

Why do AI errors survive review in competent firms?

Fluent writing produces the feeling of understanding. Reviewers read for friction, and a polished AI draft has none, so a confidently wrong paragraph can pass several rounds of review. Fatigue makes this worse, because a depleted brain takes the fluent option.

What does “take a position first” mean in practice?

Before opening the tool, write your own view in three sentences. Use the tool. Then, before reading the draft closely, write down the three questions the matter turns on and look for the answers in the draft. Where an answer is missing, go to the primary sources.

Does AI use weaken a junior legal professional’s development?

It can. Material a person generates is retained far better than material they merely read, so a junior who never drafts the framing never builds the pattern store that judgement depends on. The answer lies in what is offloaded rather than whether the tool is used.