The bottleneck is the point.

The objection arrives on cue in every conversation about supervision, and it always arrives in the same words. The human is the bottleneck.

It is meant as a criticism. Machine output flows at machine speed; a lawyer who insists on reading it, testing it and disagreeing with it slows the pipeline to the speed of one tired person on a Thursday afternoon. Everything upstream can be scaled and nothing about that person can. In any other domain this would be a straightforward engineering problem with an obvious direction of travel, which is to widen the constriction until it stops constricting.

The observation is completely correct. The conclusion is confused, and the confusion is worth taking seriously rather than dismissing, because it comes from people who are usually right about systems.

A profession is a bottleneck. That is not an unfortunate by-product of how professions grew up; it is the mechanism. A profession is a deliberate narrowing of who may advise, of what may be promised, and of how quickly a conclusion may be reached when somebody's life or business depends on it. We restricted entry on purpose. We attached duties on purpose. We put a person in the narrow place on purpose, and we made that person answerable so that the narrowness would mean something. Reserved legal activities exist under the Legal Services Act 2007 precisely so that certain things cannot be done at scale by whoever fancies it. The constriction is the safety feature.

The distinction I keep coming back to is this. In most domains where AI regulation has landed on a human in the loop, the human is a checkpoint. A moderator approving flagged content, an operator confirming an automated decision, a clinician signing off a triage recommendation. The human adds a verification step to a process whose accountability is carried by the surrounding institution. Remove the human and the institution still exists, still answers, still pays. So the checkpoint really can be optimised, and honestly should be.

In law the human is not the checkpoint. The human is the product. There is no surrounding institution holding the accountability, because the accountability was always vested in the individual; that is what the roll is for. Take the professed human out of a legal workflow and you have not removed a verification step from a product. You have removed the product and left the packaging, which is fast, cheap, fluent and binding on nobody.

Once you see it that way, the bottleneck complaint reads oddly. It is a bit like complaining that the slowest part of a bridge inspection is the inspection.

Nor are the codes silent on this, which is the part that tends to end the argument in a professional audience. Paragraph 3.6 of the SRA Code requires you to ensure that the individuals you manage are competent to carry out their role. Paragraph 2.1 of the Code of Conduct for Firms requires effective governance structures, arrangements, systems and controls. Core Duty 10 at the Bar requires you to take reasonable steps to manage your practice competently and in such a way as to achieve compliance with your legal and regulatory obligations. Supervision is not caution that a bold firm may choose to dispense with. It is mandated architecture. A firm that engineers its friction away at the point of judgement has not been brave; it has quietly stopped complying, and nobody will notice until something goes wrong in public.

There is a habit of mind underneath the complaint that is worth naming, because it is the thing the friction actually protects against. AI is not a religion. Its outputs are not revelations to be received on faith, however fluent the prose, and fluency is precisely what makes misplaced faith easy; nothing in a well formed paragraph tells you whether the authority in it exists. A system that produces plausible words deserves the treatment we give any persuasive witness, which is verification. That is not hostility to the technology. It is the technology taken seriously enough to check.

Where I have landed in practice is a rule of thumb rather than a principle, and I offer it as such. Remove friction everywhere the machine serves, aggressively, without sentiment. Intake, assembly, drafting, checking, research triage, all of it, and be embarrassed about every hour still being spent on work that never needed a vow. Then keep the friction at the point of judgement and make it real. Not a review screen with an approve button, which is friction theatre; a professional actually engaging with the material, close enough to the matter to disagree with it.

And be honest about what that costs, because there is a real tension here and pretending otherwise is how firms drift. Proper supervision does cap throughput. A firm that supervises properly will process fewer matters per lawyer than one that does not, and it will look worse on a spreadsheet that measures only volume. The answer is not to hide from that. It is to notice what the spreadsheet is not measuring, which is the entire liability position, the insurability of the practice, and whether the thing being sold is legal advice at all.

We built the narrow place on purpose. Anyone who widens it until it no longer constricts has not solved the bottleneck. They have removed the reason a client came to a lawyer instead of a search box.