AI agents

Agents that prepare the matter, and lawyers who decide it

Seven agents, one per stage. They triage, read, compare, assemble redlines and watch deadlines. They don’t send, sign, accept deviations or approve.

The seven agents

One stage each, and a line none of them crosses

Every agent writes what it found and why. When a person needs to act, the agent prepares the decision and waits. Accepting a deviation, sending to a counterparty, approving a matter and recording a signature are refused to an agent in the code, whatever its level.

AgentStageWhat it preparesWhat it never does
Intake agentIntake and readThe contract type, whose paper, personal data, and what is missing; a request to the counterparty for itSend the request, or invent a document nobody sent
Clause extractorIntake and readWhich clauses the reader found, each with page and line, and the required ones it couldn’t findInvent a clause: each is quoted from its page and line
Playbook reviewerReview and redlineThe position of every clause and the review summary for the counsel to useDecide a position: the playbook decides
Redline drafterReview and redlineThe playbook’s fallback language for each open deviation, and the redline package, heldWrite clause language of its own, or send the package
SupervisorRoute for approvalWhat the approval controls say, and the submission, to the General Counsel where referredApprove: people approve
Obligations agentAfter signingWhat is overdue or due soon, requests to counterparties for what they owe, reminders for ownersMark an obligation done, or invent one
Renewals agentAfter signingAuto-renewals whose notice deadline is close, and a renewal review before the window shutsDecide to renew or terminate, or invent a deadline
Autonomy

Autonomy is earned, one level at a time

Each agent starts at L1 for each action. An administrator can raise it, up to the tenant ceiling of L2 (Ask first), set by CONTRACTS_AGENTS_MAX_LEVEL and capped again in code. A promotion is recommended only once people agreed with at least 90% of 50 or more answers over thirty days, and a manager accepts it; an agent drops back at once if agreement falls below 70% over 20 or more.

LevelNameWhat the agent may do
L1DraftPrepares a draft; a person does the work
L2Ask firstPrepares the decision and asks a person before it acts. The ceiling for every InsightContracts agent
L3Act and tellActs, tells the person and lets them pull it back. Locked
L4ActActs on its own. Locked
Grounding

The model explains. It never supplies the clause.

Reading, the playbook comparison, the cross-document checks, the controls and the deadlines are deterministic code. Where a language model is used, it phrases what an agent found over numbered facts; a sentence carrying a name or number that isn’t in the facts is thrown away and the engine’s own wording stands.

Deterministic

The engines decide positions

Clauses, positions, severities, controls and notice deadlines are calculated, not generated.

Verbatim

Redlines are the playbook’s words

No model writes clause language. The suggested redline is your fallback text, copied exactly.

Fallback

Model off is a supported mode

With the model switched off, every agent runs on its rules and says so on each step.

Prepared decisions

Accept, reject or ask

What an agent prepares lands on the matter and on the counsel’s home. Accepting carries out only what is described; rejecting records that a person disagreed; asking sends a question back. Each answer is kept and counts toward the agent’s agreement rate.

Watch the agents prepare a matter

We run them on the demo company’s matters and walk through every prepared decision.

Request a demo