The supervisory layer for regulated AI.

Matter operations today is a stack of non-integrated systems held together by paralegal time. Intake, conflicts, coverage, reserves, billing — each lives in its own tool, or in nobody's tool at all. In the plaintiff bar, the economic consequence is severe: contingency cash flow gates on caseload throughput, and senior attorneys spend senior-attorney hours on work that should run on rails.

The first wave of legal AI did not fix this. It bolted unbounded chatbots onto the existing chaos and asked attorneys to supervise probabilistic output one query at a time. That moves the compliance burden from the system to the human, then asks the human to document it after the fact. Regulated work demands the inverse: bounded workflows that enforce procedural compliance at the architectural layer, with an immutable audit trail produced as a byproduct of the work itself.

We picked legal first because it is the highest-stakes information-processing industry on earth and the regulatory perimeter is fully written. We picked plaintiff-side small firms because the economics finally invert. Under the billable hour, AI efficiency was a revenue penalty — fewer hours, fewer dollars. Under contingency, AI compresses cost basis without touching settlement revenue. The incumbent advantages of AmLaw — capital, headcount, brand — do not translate into this architecture. Small plaintiff firms are the segment where the leverage curve bends fastest.

ABA Formal Opinion 512 was issued in July 2024 and is widely read as a procurement checklist. We read it as an architectural specification. Rules 1.1, 1.6, 5.1/5.3, and 1.5 have been transformed from policy commitments into testable system properties. A platform that satisfies them at the workflow layer — bounded agent actions, role-gated review, complete event log, billable defensibility — is not a feature set. It is the substrate on which defensible legal AI will run for the next decade. We are building for the 36-month window in which the firms that adopt this architecture restructure the industry's cost base.

LiveCards. Multi-agent matter intake for plaintiff firms. Parallel specialist agents handle conflicts, coverage analysis, reserve estimation, and UTBMS budget forecasting. Attorney sign-off is the explicit human-in-the-loop gate; the full event-level audit trail is the deliverable. Five canonical demo matters across PI, GL, CA, MM, and OP are live today. Shipping next: high-throughput PDF intake and a Rust-based document ingestion pipeline benchmarking an order of magnitude faster than current cloud incumbents.

• Supervision is the architecture, not a feature.Policy can be circumvented by a busy attorney. Architecture cannot.
• The audit trail is the deliverable, not a side effect.In a system that runs on probabilistic models, the immutable, timestamped log isn't documentation of the work. It is the work.
• AI proposes; humans dispose.The agent surfaces options inside a bounded action space. Substantive legal judgment stays with the attorney, alone.
• Efficiency without evidence is a liability.A system that makes a firm faster but cannot prove compliance to a malpractice carrier or a bar disciplinary panel is an existential risk, not a productivity tool.

Legal is the wedge, not the ceiling. The thesis generalizes to any industry where supervision-by-architecture is a regulatory, fiduciary, or insurance requirement — sectors in which a governing body or counterparty needs more than a model output and demands a defensible record of how that output was produced and reviewed. Legal is where we prove the architecture.

Argentis Labs was founded by John Marin. A decade of AI engineering in regulated industries, working with private-equity- and venture-backed platforms — production machine learning at the scale of hundreds of millions of applications, fraud and credit-risk systems built under direct federal and state regulatory oversight. JD from UC Law San Francisco, MSc in Financial Engineering from HEC Lausanne, graduate work in machine learning at Stanford. The throughline across all of it: building infrastructure for industries where the wrong output is a compliance event, not a UX bug.