Autonomy you can defend to a bar committee.
Every agent action is classified before it runs. The classification decides whether it ships, waits for a click, or waits for signatures — and the system, not a policy memo, enforces the answer.
From proposal to accountable action
Conceptual policy routing: an action is proposed and checked against policy, then routed to human review when required or to a permitted action. The outcome is recorded. This diagram explains the model; the controls below describe the oversight tiers.
- 1Propose an action
- 2Check the applicable policy
- 3Route for human review when required; act only when policy permits
- 4Record the outcome
Tier 1 — Low risk
- Covers
- Routine document generation, standard research, scheduling, internal summaries.
- Approval
- Auto-approved with full logging.
- Escalation
- None.
Tier 2 — Medium risk
- Covers
- Draft pleadings, settlement recommendations under the firm threshold, major client communications.
- Approval
- One-click approval from the decision queue.
- Escalation
- Escalates to the Managing Partner when agent confidence falls below 85%.
Tier 3 — High risk
- Covers
- Final settlements, court filings, new client acceptance, material strategy changes.
- Approval
- Full review plus digital signature.
- Escalation
- Managing Partner, the responsible Equity Partner, and legal ethics review.
Tamper-evident audit chain
Each audit row carries the hash of the row before it. Deleting or editing history breaks the chain and the break is detectable. Eligible plans additionally seal periodic Merkle roots so a third party can verify a snapshot without seeing the underlying records.
Attribution on every artifact
Any document, message, or analysis an agent touches carries a footer naming the agent, the model, the human approver, and the timestamp. Clients and courts see the same provenance your partners do.
Trip the Tier-3 gate yourself
The public sandbox runs a settlement-posture agent and holds the output until a partner signs.
- 57 specialist agents
- 3 oversight tiers
- Hash-chained audit trail
- EN / ES parity