Most writing about agentic AI in law is written by people selling it, and it describes capability. This is one of the few serious treatments written from the other direction — starting from what a solicitor is obliged to do and asking whether these systems permit it.
The finding that matters is a structural mismatch rather than a technical shortcoming. A solicitor remains professionally responsible for work produced under their name. An agentic system produces work through a long chain of intermediate decisions — which sources to consult, what to discard, when the task is complete — most of which are not surfaced and some of which are not recorded. The responsibility has not moved; the ability to discharge it by inspection has. The Law Society names this as a widening regulatory and liability gap, and does not suggest it closes on its own.
The practical consequence for a firm is a design rule rather than a prohibition. Systems that escalate consequential decisions to a human and keep a readable record of what they did are compatible with professional obligations as they currently stand. Systems that impress by needing nobody are, on this analysis, exactly the ones a firm cannot supervise. That inverts the usual procurement instinct, which rewards autonomy.
The optimistic counter-position is that this is transitional — that auditability tooling will mature and the gap will close. The pessimistic one is that a system complex enough to be worth deploying is complex enough that meaningful human inspection of every step is impossible in principle, and that the profession will have to change what supervision means. This is genuinely unresolved. Note that the jurisdictional frame here is England and Wales; the structural argument travels, the regulatory specifics do not.