The page a leader opens inside Legal AI OS.
One chart, three things to build now, and a record that grades every call.
The screen answers one question before you scroll: what should we build before the rules land?
Courts first punished fake AI citations after the fact (Mata v. Avianca was the first such sanction). Now more than 300 judges require lawyers to disclose AI use up front, and that patchwork of local orders is converging toward a single certification standard.
Verification-by-default with an auditable trail, so certification is a byproduct.
Model rules: 3.3, Rule 11
Signing a line that says 'a human reviewed it' is no longer enough. Courts and bars increasingly expect a documented verification process: a record of what was checked, how, and by whom.
Logging, trace analysis, evaluation records as the compliance artifact.
Model rules: 5.1, 5.3
The duty is shifting from 'understand the tool you use' to 'run a governance program' — training, written policies, and oversight — with mandatory AI continuing-education requirements spreading state by state.
Training and enablement at scale (iTrain bet; AI Build Lab proof).
Model rules: 1.1
Malpractice insurers are adding AI-governance questions to policy renewals and pricing coverage on the answers. They move faster than bar regulators, so insurance, not the rules, becomes the real gatekeeper for whether a firm can use AI at all.
The governance program as an insurable artifact: model inventory, approval process, oversight documentation, measurement.
Model rules: 1.1, 5.1, 5.3
The rules are a patchwork — the EU AI Act, Colorado, Texas, 35-plus state bars, 300-plus judges — each slightly different. The patchwork itself is the pressure: a firm working across jurisdictions needs one operating model built to the strictest standard.
The unifying operating model. This is the whole thesis.
Model rules: 1.1, 1.6
AI tools that learn from what you feed them, or mix data across matters, put client confidentiality at risk. Firms are being pushed to map where client data goes and get written no-training guarantees from vendors before a tool is approved.
Data-flow mapping and vendor attestation at tool onboarding.
Model rules: 1.6, 1.7, 1.9
Independent studies now publish how often each legal AI tool gets things wrong. Once error rates are public, choosing a tool becomes a measurable competence decision rather than a matter of trust.
You already built one: NERVE. A benchmark is a compliance instrument.
Model rules: 1.1
AI is shifting from drafting text to taking actions on its own across multiple steps (so-called agents). 'Just review the final output' breaks down when no single person saw every step, reopening old duties around supervision and the unauthorized practice of law.
Human-decides gates and scope-limiting by design.
Model rules: 5.1, 5.3, 5.5
AI collapses the hours a task used to take, which strains the billable-hour model and raises the question of what counts as a 'reasonable fee' when the work now takes minutes.
Measure value delivered, not hours saved.
Model rules: 1.5
Tools that predict how a specific, named judge will rule cross the line from analytics into profiling. France has already criminalized it, and US regulation is likely to move in the same direction.
Strategy stress-testing with human-reviewed, explicitly-uncertain priors; never sell 'predict the judge'. Persona models are low-confidence parameters, never determinative. (France already criminalized judicial profiling; US direction is toward rules.)
Model rules: 3.5, 8.4, —
Not scored yet — no market signal recorded.
Today the lawyer carries all the risk when an AI tool fails; there is no clear legal path to hold the vendor responsible. Pressure is building, through disclosure mandates and product-liability proposals, to shift some of that liability onto the toolmakers.
Track for a shift; document tool provenance now.
Model rules: —
Not scored yet — no market signal recorded.
Courts now expect a documented check on every AI-assisted filing, not a signature after the fact.
Attestation gives way to a recorded process: what was checked, how, and by whom.
The duty shifts from understanding one tool to running a program across all of them.
Each fault line now carries a fourth reading, Enablement (E): can the market actually supply the tooling to stand this control up. And a companion view turns the landscape into the two calls a firm actually faces, kept separate so a required control is never delayed by the economics of the AI it governs.
Required of us, and can we meet it. The billable hour never blocks it. Stand up now · build capacity first · no mandate yet.
Gated by pricing, whether the work is capturable, and market enablement E ≥ 6. Deploy now · fix pricing first · defer · watch.
The full model, with the two-firm worked example, is on how it works.
The screen carries its own calibration record: every run is checked against what actually happened, misses left in. For the full per-order breakdown and how the scoring works, see how it works.