AI Can Make the Factory Faster. It Cannot Make the Bottleneck Bigger.
The machine got faster but the lawyer's mind did not. Legal technology is racing to accelerate the work professionals produce, while almost nobody is asking how much the professional has to hold in their head to produce it. If firms mistake faster components for greater human capacity, AI will not lighten the load on the profession. It will add to it.
What the Legal AI Race Is Optimising For
The direction of legal AI is no longer ambiguous. In May 2026, Harvey announced more than 500 pre-built agents across practice areas, alongside a tool that lets firms build their own. Its chief executive described the industry as "well past AI as an assistant". Harvey's own product pages describe agents that run legal work end to end in the background and return review-ready output, while the lawyer keeps the judgment.
This is not a criticism. Research, drafting, document review and due diligence are real work, and making them faster is valuable. But every one of these advances is aimed at the same question: how do we produce more, faster? Almost none is aimed at a different one:
How much mental load does a professional have to carry to exercise judgment reliably, and what happens to that load when everything around them speeds up?
The Threading Machine Problem
Picture a factory line. Machine A used to produce 10 components an hour and now produces 100. Machine B has been upgraded the same way. At the end of the line sits a threading machine that can finish 20 products an hour, and nobody has touched it.
The factory has not gone from 10 finished products an hour to 100. It has gone to 20, and it now has a growing pile of half-finished work in front of the threading machine. Every component that arrives faster than the final step can take it becomes inventory, delay and pressure.
Manufacturing learned this decades ago. Eliyahu Goldratt's The Goal (1984), which introduced the Theory of Constraints, made the point that a system produces only as fast as its constraint. Improving any step that is not the constraint does not increase output. It increases work in progress.
The Legal Version
AI accelerates the components: research, drafting, review, searches and summaries. But every matter still passes through a human mind that must understand it, weigh it, verify it, put it in context and take professional responsibility for it. That mind is the threading machine. Its capacity is set by attention and working memory, and nothing upstream changes either.
The Bottleneck Thesis
Legal AI accelerates the production of material that still requires legally accountable cognition. Where the arrival rate of that material exceeds the available capacity for competent lawyer review and judgment, the result is not limitless productivity. It is a growing queue at the lawyer-review stage.
AI relieves lawyer workload only if firms deliberately preserve some of the released capacity, rather than automatically converting it into additional demand.
The Bottleneck Has a Measurable Size
This is not a metaphor about stamina. Human cognitive capacity has hard limits that psychology has been measuring for decades.
Research Finding: Working Memory
In an influential 2001 review in Behavioral and Brain Sciences, Nelson Cowan argued that the mind can actively hold only about 4 chunks of information at once, somewhere in the range of three to five. That is the working space a lawyer uses to connect a client's facts, the relevant law, the risk and the advice. It does not grow because the research arrived faster.
Research Finding: Attention Residue
Sophie Leroy's 2009 study in Organizational Behavior and Human Decision Processes found that when people switch from one task to another, part of their attention stays behind on the first task, and performance on the next task suffers. She called this attention residue. Simply finishing the first task was not enough to clear it. Every switch carries a cost the timesheet never records.
Put those two findings together and the bottleneck comes into focus. A professional has a small working space, and every interruption, open question and task switch leaves some of that space occupied. Faster tools that increase the number of threads running at once do not expand the space. They crowd it.
| Part of the work | What AI does to it | Where the load lands |
|---|---|---|
| Research and retrieval | Much faster | Increasingly the machine |
| First drafts and summaries | Much faster | Increasingly the machine |
| Checking what AI produced | New work that did not exist before | The professional |
| Judgment and synthesis | Largely unchanged | The professional |
| Keeping track of everything in motion | Unchanged, often increased | The professional, in their head |
Where the Saved Time Actually Goes
This is not a new pattern. Think about property searches. When they took days to come back, a lawyer might run three, consider each one carefully and move on. Now that results arrive in minutes, running ten is easy, and ten quietly becomes what careful practice looks like. Retrieval got faster. The lawyer's capacity to absorb, weigh and take responsibility for what ten searches show did not.
The cycle repeats with every productivity gain:
- Technology reduces friction on a task.
- More work becomes economically viable, so more of it enters the system.
- Professional expectations expand to match what is now possible.
- The same mind now has more arriving at it.
There is early evidence that AI follows the same path, and that it adds a new kind of load of its own. In February 2026, Aruna Ranganathan and Xingqi Maggie Ye of Berkeley Haas published findings in Harvard Business Review from an 8-month study of how generative AI changed work at a US technology company of about 200 employees. AI did not reduce work. It intensified it: people worked faster, took on a broader range of tasks and extended work into more hours of the day, often without being asked.
The mechanism they describe is a cognitive one. People ran several threads at once, checked AI output repeatedly, and revived deferred tasks because AI could supposedly handle them in the background. The result was constant attention switching and a growing number of open tasks. The researchers warned that this can lead to cognitive fatigue, burnout and weaker decision-making.
The Irony
AI is sold as a way to take work off the professional. But every agent running in the background is another open thread, and every output is something that must be checked by the one person who carries the responsibility. Delegating the work does not delegate the attention required to trust it.
An honest caveat: the HBR study examined one technology company, not a law firm, and the authors describe it as in progress. It is a signal, not proof. But it points straight at the bottleneck.
The Load Nobody Is Automating
Much of what fills a professional's working memory is not the work the client pays for. It is the overhead that surrounds it:
- Remembering what is in motion and what is waiting on someone else
- Wondering whether a colleague actually did what they said they would
- Holding a follow-up in mind until tomorrow
- Rebuilding context every time you return to a matter
- Verifying output you did not produce yourself
- Deciding whether something is urgent enough to interrupt someone
- Worrying, in the background, about what might have been missed
None of this requires professional judgment. All of it competes with judgment for the same four or so slots of working memory. Much of it is what we have previously called meta-work: the work about the work. It is not tiring because it is difficult. It is tiring because it never closes. Did I send that? Did Sarah check it? What if she thought I was doing it? Should I chase her now or tomorrow? Each question is small. Carried across dozens of live matters, they become a constant background tax on attention.
Why Open Loops Are So Expensive
There is a well-studied reason this overhead is so draining. In a 2011 paper in the Journal of Personality and Social Psychology, E. J. Masicampo and Roy Baumeister showed that unfinished goals persist in the mind. Across several studies, open goals produced intrusive thoughts during unrelated reading, kept goal-related words unusually accessible, and reduced performance on an unrelated problem-solving task.
The more useful finding was the remedy. When participants made a specific plan for the unfinished goal, those interference effects disappeared. The mind let go once it trusted that the goal was being handled, and that effect depended on the plan being a real commitment rather than a gesture.
What This Means for Practice
An open loop does not only cost attention while you work on it. It costs attention while you are trying to work on everything else. The research suggests the loop can close not only when the task is finished, but when something credible is holding it. An obligation held by a system you trust costs far less attention than the same obligation held in your head.
These were laboratory studies, not studies of law firms. But the mechanism applies to anyone carrying dozens of live matters at once.
Two Different Uses of Technology
Almost all of the energy in legal technology is going into one of two possible goals, the same productivity-first pattern we described in Why Professional Services Is a Juggling Act. Both goals are legitimate. Only one of them protects the bottleneck.
| Question | Capacity multiplication | Cognitive load reduction |
|---|---|---|
| The promise | You can now do twice as much | You can do the same with far less held in your head |
| What it targets | Task throughput | Attention and working memory |
| Effect on the bottleneck | More work arrives at it | Less noise competes for it |
| Economic story | Obvious and easy to sell | Quieter, but it compounds |
The goal is not to take professionals from 100% effort down to 70%. It is to stop them spending their 100% on things a system could carry for them. The case for doing so rests on quality and risk. Judgment draws on a small, finite working space, and anything that occupies that space without adding value is a cost the firm pays in the quality of its advice. As we argued in The 150% Trap, pushing past cognitive capacity does not fulfil the duty of care. It erodes the judgment that duty depends on.
How to Protect the Bottleneck
1. Measure Load, Not Just Output
Firms measure hours, matters and billings. Almost none measure how many threads a person is holding at once, or how often they are interrupted. If the constraint is attention, then attention is what needs to be visible.
2. Preserve Released Capacity Deliberately
When AI saves a lawyer three hours, the instinct is to treat that as three hours of new capacity. Decide in advance how much goes to more work, and how much is held back for review, recovery and thinking time. Released capacity that nobody protects is converted into demand by default, and the queue at the review stage grows. If nobody decides, the pile decides.
3. Move Open Loops Out of Heads
Every commitment, follow-up and dependency should live in a shared structure with a named owner, visible to the people who would need to act if the owner cannot. The point is not reminders. It is making obligations trustworthy enough that nobody needs to carry them mentally. Memory under pressure is unreliable for neurological reasons, not moral ones, which is why systems beat willpower, and why accountability works best as a system rather than a culture.
4. Let Technology Watch the System, Not Occupy the Person
The most useful role for AI around professional work is ambient, not conversational. It should not be one more thread demanding attention. It should notice when a matter has gone quiet, when expected progress is missing, or when one person is carrying far more than others, and surface that to a human only when a human needs to decide something.
This is the thinking behind Duetiful. Deadlines are one of the most persistent open loops in professional practice, so Duetiful gives each one a clear owner and a four-layer Backstop: the owner first, then a colleague, then the team manager, then the business owner. The aim is not more alerts. It is fewer things any one person has to hold in their head, because the system is holding them.
Take the Remembering Off Your Team's Minds
Give your people back the attention they need for judgment.
- 14-day free trial to see it with your own matters
- Four-layer Backstop so no one carries it alone
- Cancel anytime, no lock-in contracts
The Question Firms Should Be Asking
Tools like Harvey may well help lawyers finish the work faster. That is worth having. But the profession has spent decades asking how to make professionals more productive, and far less time asking how to make professional practice cognitively sustainable.
AI can make the factory faster. It cannot make the bottleneck bigger. The firms that do well in the next few years will not be the ones that push the most work through their people's minds. They will be the ones that recognise attention as the scarce resource, and build the infrastructure to protect it.
About the Author: Matt is the founder of Duetiful, a registered migration agent, and a non-practising Australian lawyer who has run a professional services practice since 2007.
Sources
- Harvey. Harvey Launches Purpose-Built Legal Agents Across Every Major Practice Area (5 May 2026).
- Artificial Lawyer. Harvey Drives Forward Use of Legal Agents (5 May 2026).
- Harvey. Harvey Agents: Delegate the Work. Own the Judgment.
- Goldratt, E. M. and Cox, J. The Goal: A Process of Ongoing Improvement. North River Press (1984).
- Cowan, N. The Magical Number 4 in Short-Term Memory: A Reconsideration of Mental Storage Capacity. Behavioral and Brain Sciences, 24(1), 87 to 114 (2001).
- Leroy, S. Why Is It So Hard to Do My Work? The Challenge of Attention Residue When Switching Between Work Tasks. Organizational Behavior and Human Decision Processes, 109, 168 to 181 (2009).
- Ranganathan, A. and Ye, X. M. AI Doesn't Reduce Work, It Intensifies It. Harvard Business Review (9 February 2026).
- Masicampo, E. J. and Baumeister, R. F. Consider It Done! Plan Making Can Eliminate the Cognitive Effects of Unfulfilled Goals. Journal of Personality and Social Psychology, 101(4), 667 to 683 (2011). doi:10.1037/a0024192
Comments