AI for legal teams
An eighty-page contract arrives for review. Before a lawyer reads a line, its structure has to be mapped, the clauses that depart from the template found, the cross-references, the definitions. An hour of preparation for an hour of useful reading. The task that costs is that hour — not the opinion, which stays with the lawyer, and must.
The tasks we remove first in this industry
What this industry pays when nothing moves.
Ranges, not promises. They come from tasks we meet in this industry and orders of magnitude we measure in the survey — yours may fall outside, in either direction.
What we remove, and what we leave.
Each task is described with what the agent takes on and what stays with a person. If the second column is empty, we have not understood the industry.
Prepare the reading of a contract
What the agent does
It maps the structure, the definitions, the cross-references, and sets each clause against your template, flagging deviations — without saying whether they are acceptable.
What stays with a person
The reading itself, and the opinion. ‘This clause differs from the template’ is an observation ; ‘this clause is risky’ is an opinion, and it belongs to the lawyer.
Build a file
What the agent does
Documents, correspondence, chronology : it assembles the file in your format, every fact pointing to its document, and flags what is missing.
What stays with a person
The characterisation of the facts and the strategy. The file is ready ; what is done with it is not.
Track the monitoring
What the agent does
It reads the sources you designate and flags what affects your matters or your templates, with the exact extract and the reference.
What stays with a person
The interpretation and the update of templates. It alerts ; the lawyer decides.
Answer first-line internal requests
What the agent does
‘Where is the matter’, ‘which template for’, ‘have we dealt with’ : it answers from the file and the template base, citing.
What stays with a person
Any question whose answer is an opinion. It says ‘template X was used three times for this case’, not ‘use template X’.
Write up notes and attendance notes
What the agent does
From hearing, meeting or client notes, it produces a note in your format, with the points to action.
What stays with a person
The review, and what was said. The note is the lawyer’s, not the machine’s.
No legal opinion is produced by an agent. Ever.
It is one of the three founding limits of the agency, written on every quote. An agent spots, assembles, compares to a template and cites. It does not characterise, recommend or advise. The line between ‘this clause differs’ and ‘this clause is risky’ is the one in the diagram below.
In both branches, the agent produces an observation. The opinion appears nowhere in this diagram — deliberately.
Ten systems, one survey, one system opened at a time — the one with the shortest payback. Read the full method, or go back to the AI agency page.
What we open it with, and why here.
The choice of tool comes last, and it is reversible. But in this industry, three constraints keep coming back and steer the choice before we even start.
Files are covered by legal professional privilege. Before any agent, we decide where the documents live — on your premises, in most cases — and who has access. Nothing is used to train a model.
The agent is only worth what your templates are worth. The first job is often to gather them, date the versions and remove duplicates.
It puts two texts side by side and cites. It does not say which is better. As soon as a question calls for an opinion, it hands over.
What we hear in this industry, before we start.
Is legal privilege compatible with an agent ?
With an agent installed inside your perimeter, yes : the documents do not leave it, access is named, and nothing is used to train anything. It is the first decision of the workshop. If it does not suit you, we stop there.
An agent that flags deviations — isn’t that already an opinion ?
No : a deviation is a fact. ‘Clause 12 received differs from clause 12 of the template’ can be checked in one click. Saying whether the deviation is acceptable is the opinion — and the agent never formulates it, technically as much as by rule.
Our templates are out of date. Is that a problem ?
It is the most frequent case, and it is the first job : gather, date, de-duplicate. An agent comparing against an outdated template flags deviations that are not — we know it, we start there.
And for an in-house legal team ?
The tasks are the same, plus internal requests : ‘can we sign this’, ‘which template for’. The agent answers ‘this template served three times for this case’, never ‘sign’.
The areas where we are called most for this.
The same tasks, in neighbouring industries : AI for banking, AI for real estate, AI for insurance.
We will tell you which system to open first.
Describe the task that costs you the most. We will tell you what is feasible, how long it takes and what it costs. If the answer is no, you will leave with the two reasons why.