Agents in contract review, and the line they should not cross

A contract arrives. Before anyone can form a view on it, somebody has to read the whole thing, find the clauses that matter, compare them against the position the organisation normally takes, and note where this one differs.

That is hours of careful attention, and almost all of it is location rather than decision.

What the agent does well

Finding the clauses. Liability, termination, indemnity, data protection, governing law, renewal terms. An agent that extracts each one with its exact text and position turns a full read into a review of extracts.

Comparing against your standard position. If the organisation has a playbook, the agent can mark every place this contract departs from it and by how much.

Producing a first-pass summary of what changed from the previous version, which is otherwise a diff somebody does by eye.

What it must not do

Decide whether a departure is acceptable. Approve. Sign. Advise.

The distinction is not about capability. A contract position is a judgement about risk the organisation is willing to carry, and that judgement needs a person who can be asked to account for it. An agent producing an approval is producing something nobody is accountable for, which is the one output legal work cannot use.

Why provenance matters more here than anywhere

A summary of a clause is not a clause. If an agent says the liability cap is a particular figure and a reviewer relies on that summary, the summary has become the contract in practice.

Every extraction should carry the exact source text and its location, and the reviewer should be reading the clause rather than the description of it. The agent's job is to put the right paragraph in front of a person quickly, not to replace reading it.

The failure mode

Missing a clause rather than misreading one. A contract with an unusual structure, an obligation buried in a schedule, a definition that changes the meaning of a term used forty pages later.

An agent that reports nothing unusual is indistinguishable from an agent that found nothing because it did not look properly. This is why evaluation sets for contract work should include documents with deliberately awkward structure, and why absence should be reported as low confidence rather than as a clean result.

Practical constraints

Contracts are confidential and frequently cannot leave the organisation, which makes self-hosted deployment the starting assumption. Access boundaries enforced at the point of action. A complete audit trail, because the question of who saw what and when is a real one in a dispute.

Starting shape

One contract type, extraction and comparison only, with a lawyer reviewing every output. Named owner in legal. Evaluation built from real past contracts including the difficult ones.

Next

The Agent Orchestration System pages cover policy, audit and evaluation. The deployment pages cover self-hosted and air-gapped options.