Moats and Missiles: How the Castles Will Starve
Let us start with the punchline to this blog post:
Under a narrow but plausible set of assumptions, most law firms do not survive the next decade.
We saw a post on LinkedIn last week depicting the battle for legal services as a shift from castles with moats to armies launching projectiles. It made us think about the battle that is raging right now — between law firms as the castles with moats, and everyone else trying to claim a slice of the $1 trillion legal services market.
The Moat Was Real
Law firms built moats over centuries. Bar licensing. Malpractice liability. Institutional trust. Relationships cultivated over decades. And a billing model that bundled commodity work with premium work. The moat was legal, economic, relational, and structural — all at once.
AI has been attacking the castle for a while. Document review, research synthesis, first drafts — Gen AI tools were doing legal work since 2023.
Now, there is a step change with AI agents.
The difference with AI agents is not quality of work, it is autonomy of work.
A Gen AI tool requires a human to initiate every task, review every output, and decide every next step. An AI agent can be given an objective and pursue it across multiple steps without hand-holding, and only stop when a certain quality threshold has been met — pulling the relevant documents, running the research, flagging the risk clauses, drafting the response, routing the output for sign-off. The human is no longer the operator of every step. They become the supervisor of the outcome. One lawyer can supervise many agents running simultaneously.
The economics of legal service delivery — which were built entirely around the cost of human time per task — breaks with autonomous agents. And let us be clear — agents do not have to be as good or better than humans to break the law firms’ model. The pyramid leverage model of the law firm breaks because it was never designed to compete with machine-speed parallelism.
The Starvation Strategy
Every previous wave of legal technology made the law firms’ castles stronger. Document management, e-discovery platforms, contract tools — faster workflows, better organized files, smarter search, etc. The incumbents benefited most, because they had the most workflow to optimize.
Agents are different in nature. They do not make the castles more efficient. They make it possible to claim territory (the $1 trillion of legal services spend) without building castles or moats at all.
If agents can do the work of junior lawyers, then the pyramid starves. Law firms have to change their whole business model from hourly billing to something else in order to sustain the partner salaries.
If agents become the ones who do the grunt work, then in-house teams can deploy their own agents, and NewMod law firms can deploy their own agents. No one needs to come to the castle at all, except for something beyond simply “getting the work done”.
Can’t the Castles Deploy AI Agents?
The obvious counter-move: law firms can deploy AI agents themselves. i.e. Buy the missiles, and fire them from the battlements.
Some firms are trying. The captive ALSP model — ClearyX, A&O Shearman Advance, Nextlaw — is the structural expression of this counter-move. Build a separate entity, different economics, different staffing, use the parent firm's relationships as the distribution channel.
The logic is sound. But the captive ALSP are constrained by reality:
they cannot go too cheap — that undermines the parent's rates.
they cannot move too fast — that embarrasses partners who took six weeks to do the same work.
they cannot replace lawyers too aggressively — the parent's entire workforce model depends on them.
Captive ALSPs exists to retain the castle's interests, not to win territory.
The one real advantage a captive ALSP has over a pure startup is warm introductions — direct access to the GCs of the world's largest companies. That pipeline has value, but it has a half-life. As clients become more sophisticated about agent-delivered work, they will evaluate providers on outcomes and economics, not on who made the introduction.
Who Will Capture the Economic Value?
The castle analogy only goes so far — but let’s play out the battle over the next few years… who will win the territory?
We anticipate three things will happen simultaneously, in proportions we cannot yet confidently predict.
First, clients simply spend less. The GC who deploys AI agents does not redirect her outside counsel budget to a new provider. He or she returns it to the CFO. The value does not transfer to a new winner. A portion of the legal spend simply evaporates from the legal market entirely.
Second, a new class of entity captures the spread between token cost and outcome value. The AI agent costs a few dollars. The legal outcome is worth hundreds or thousands of dollars. Whoever sits between those two numbers owns the highest margin business model in the market. That entity does not fully exist at scale yet. It is being assembled right now — by Crosby charging flat fees per contract, by Norm billing on outcomes rather than hours, by a cohort of YC-backed AI-native firms that look like law firms to clients but run like software companies inside.
Third — and this is where Jevons' paradox might become relevant — agents may unlock latent demand the castles never served. Most individuals and small businesses are legally underserved not because lawyers do not exist, but because the fees are too high. At sufficient quality, agents create new legal activity in territory the billable hour model made inaccessible. The pie does not just get redistributed. Some portion of it may grows.
But, there are assumption. None of this plays out unless agents can deliver work that clients will actually rely on — not demo quality, not a first draft requiring complete revision, but decent output. Right now, as at April 2026, that bar has not been cleared. The verification problem is real. The hallucination problem is real. The inconsistency problem is real.
Law firms fall if:
AI agent quality crosses the reliability threshold for a meaningful category of legal work;
clients become willing to accept agent-delivered outcomes without insisting on human sign-off;
regulation lags long enough for new providers to establish themselves before bar associations move on unauthorized practice; and
law firms fail to transform their own economics fast enough — which their own incentive structures make very difficult.
The Death of the Castles
Under a specific, narrow, but entirely plausible set of assumptions, there is no clear survival path for the majority of law firms, simply because the castle ran out of reason to exist.
The brand is still there. The partner relationships are intact. The walls are still standing. But the math stops working. And in business, when the math stops working, the castles fall.
Everything is just a question of how long the battle takes.
Post Script: the Prisoners’ Dilemma
There is a bigger problem underneath all of this, and it deserves its own piece… later.
Every player in this landscape — existing law firms, in-house teams, and agent-native NewMod firms — faces a version of the same dilemma. Each individually rational decision leads to a collectively irrational outcome.
Firms deploy AI agents to survive, which accelerates the erosion of their own business model.
In-house teams absorb the quality risk because the cost savings are too compelling to ignore.
Agent-native providers race to capture territory before quality is reliable, because first-mover advantages are too valuable to cede.
The legal system gets hollowed out. Quality of work degrades. The junior pipeline disappears. The clients who thought they were getting cheaper legal services may discover, too late, that they were getting lower quality protection at a price that disguised the real cost.
Nobody chooses this outcome. It is emerging from everyone making individually sensible decisions under competitive pressure.
The individual client may genuinely benefit, but the system may pay the price.
