When AI Wears Lawyers Down
There is a kind of tiredness that did not exist three years ago. It does not come from the usual sources: the long files, the late nights, the difficult client, the urgent deal, the draft that comes back covered in comments. We know that kind of tiredness. We have lived with it for years. This one is different, because it comes from working next to something that never stops.
In our team, we call it the supervision tax.
AI may produce words at almost no marginal cost, but someone still has to read them, doubt them, and check whether they are accurate, useful, lawful, appropriate, consistent with the client’s position, consistent with the firm’s standards, and consistent with professional judgment. In legal work, that someone is always a human being. The model can generate the draft, but it cannot carry the responsibility. We can. And that is where the fatigue begins.
Why the Supersion Tax Is Different
Burnout, stress, overload, lack of control, poor recognition, unfair reward, FOMO. These are familiar categories. They are real, and they are well mapped. The supervision tax is something else.
When a junior lawyer drafts a memo, we are not only reviewing the words on the page. We are also relying on a relationship, a training process, a professional hierarchy, and a shared understanding of how the work was done. We may still check carefully, but there is a human context around the draft. When a model drafts something, there is no such context. We trust nothing by default.
Every sentence carries a small question: is this accurate, is this invented, is this too confident, is this the firm’s position or just a plausible sentence, is this the law or a linguistic approximation of the law, is this useful or merely fluent? That kind of reading is suspicious reading, defensive reading, reading with one hand on the emergency brake.
Do that for one document and it is manageable. Do it hour after hour, across contracts, emails, memos, summaries, policies, clauses, presentations, and client notes, and it becomes a form of labour in itself. The output looks effortless, but the checking never is. That is the tax: the machine produces at speed, and we pay in attention.
Two other charges sit on top. The first is that the finish line keeps moving. Before AI, a memo was usually finished when we ran out of time, budget, energy, or useful ideas. Now there is always another pass available: one more angle, one more version, one more rewrite, one more “make it sharper,” one more “make it more concise.” The work expands to fill the model’s willingness, and the model’s willingness has no limit.
The second charge is ownership. AI can produce the words, but it cannot answer for them. Responsibility has to land somewhere, and it lands on the person who approves the final output. That person may not have written the first draft. They may not even fully know how the draft was generated. But they are still the one who carries the professional risk. Fast production combined with slow, anxious ownership is what wears people down.
How It Shows Up in Legal Teams
The symptoms are quiet. We delay starting a task because we can already imagine the hours we will spend editing the machine’s draft. We stop writing from scratch and spend more time reacting to text produced by something else. Our own voice begins to flatten because we are constantly correcting, trimming, polishing, and normalising AI-generated language. We produce more, but we trust less. That is the paradox.
AI is supposed to reduce friction, and often it does. But when the workflow is badly designed, it creates a low, constant resistance. Not dramatic exhaustion. Not collapse. Just a persistent sense that every task now comes with an invisible layer of supervision.
In law firms, the tax shows up in one way. Associates use AI to draft, summarise, compare, or structure. Partners then review work that nobody on the team may have truly authored in the traditional sense. Meanwhile, the billable model can quietly reward iteration. If the draft can always be improved, refined, expanded, or tested again, the temptation is to keep going long after the point of real value.
In corporate legal departments, the tax shows up differently. Teams are smaller, the volume is higher, and the business expects speed because it assumes the tool has already made everything instant. Often, one in-house lawyer becomes the only real checkpoint between an AI-assisted draft and the business decision that follows. No second pair of eyes, no long review chain, no margin for theatrical experimentation. Just speed, pressure, and accountability.
In both settings, the legal function risks producing more than ever while trusting its own output less than ever.
Five Habits That Lower the Tax
The supervision tax is real, but it is not inevitable. We can lower it by changing the way we work with AI.
The first habit is to define “finished” before we prompt. Before asking the model to produce anything, we should define the stopping condition. What is the document for? How long should it be? What standard does it need to meet? What question must it answer? What would make it good enough? Without a stopping condition, the work has no natural end. The model will always offer another version, and we will always be tempted to ask for one. A clear definition of “finished” protects judgment from endless iteration. When the draft meets the standard, we stop, even if a better version is theoretically one click away.
The second habit is to split making from checking. Drafting and verifying are different mental modes. When we draft, we need flow. When we check, we need discipline. Mixing the two every few lines is exhausting. A better workflow separates the phases: first we produce, then we verify, then we edit. Not all at once, not paragraph by paragraph, not in a constant loop of generation, correction, regeneration, doubt, and correction again. The model may be fast, but our attention is not infinite. We should spend it deliberately.
The third habit is to reduce the number of assistants. Many teams now keep several AI tools open at once: one for drafting, one for research, one for translation, one for document review, one inside Word, one inside the browser, one inside the knowledge system. Each tool has its own interface, tone, strengths, weaknesses, and failure modes, and that creates switching fatigue. The answer, often, is fewer tools with clearer use cases: one tool for one kind of task, one workflow, one review standard, one place where accountability sits. The fatigue often lives less in the technology itself than in the constant movement between systems.
The fourth habit is to keep one task the model never touches. This may sound inefficient, but it matters. Every lawyer should keep at least one meaningful task fully human: a clause we always draft by hand, an argument we reason through on paper, a negotiation position we structure without assistance, a memo outline we build ourselves before asking the model to help. Not because AI cannot help, but because judgment needs a place to stand. If everything begins with the machine, we risk becoming permanent reviewers of someone else’s first draft. Over time, that changes how we think. It also changes how much confidence we have in our own professional instincts. Keeping one task fully ours preserves muscle. It keeps authorship alive.
The fifth habit is to name the owner before the model starts. Ownership should not be discovered at the end. Before using AI on a task, we should decide who is accountable for the final output. Who reviews it? Who approves it? Who checks the legal position? Who decides whether it can be sent? This sounds basic, but many AI workflows skip it. The result is predictable: the draft moves quickly until the final stage, where everyone suddenly becomes nervous. Then the real work begins late, under pressure, with unclear responsibility. Clear ownership at the start reduces anxious re-reading at the end. It also reminds the team of a simple truth: AI can support the work, but it cannot absorb the duty of care.
The Point
The supervision tax is real. But we do not lower it by rejecting AI, and we do not lower it by pretending that every AI-assisted task is automatically more efficient. We lower it by designing better workflows.
When AI leaves a legal team more drained than before, we should look at the process before blaming the tool. Where does the draft begin? Where does review happen? Who owns the output? When do we stop? What do we still do ourselves?
At Better Ipsum, we help law firms and corporate legal departments build these habits into the way they actually work, so that AI adds capacity instead of quietly draining it.
If your team is faster than ever and more tired than ever, that is worth a conversation..