Matter-scoped agents on one Czar in the practice
A practice's material is privileged, confidential, and walled by matter. Agents on one Czar in the practice draft, review, and research inside each matter's scope, and nothing client-facing leaves without the responsible partner.
· 4 min read
IN SHORT
A legal, audit, or advisory practice runs its agents on one Czar inside the practice. Agents in Agency Studio draft from the practice's own precedents, review the file, and research within the matter's scope, and each agent sees only the matter it was authored for. The responsible partner approves anything that goes to a client, and every retrieval and draft is on a log the practice holds.
A practice's material is privileged, confidential, and walled by matter. Within the perimeter there are further perimeters: a team on one matter must not see the file of another, and the practice is answerable for keeping it that way. The drafting, review, and research that agents are good at would, in a cloud assistant, mean sending that material out of the practice and across its own walls in the same step. One Czar in the practice runs the agents inside each matter's scope, and nothing goes to a client without the partner who is responsible for it.
The wall inside the practice
Most organizations have one boundary to defend. A practice has many. A lawyer on a transaction cannot read a litigation file in the next corridor. An audit team on one engagement cannot see the working papers of another. An advisory team is walled from the one advising the other side. These walls are not a convenience; they are the condition under which the practice is allowed to hold its clients' material at all.
The work inside each wall is familiar. A first draft assembled from precedents the practice has used before. A review of a file to find what differs from the last version, the standard form, or the client's instructions. Research that stays inside the question the matter is asking. All of it is reading the practice's own material and writing from it, and all of it ends with a partner deciding what the client receives.
Why a cloud assistant fails here
Client material cannot leave the practice to be drafted from. That rules out the obvious tool before any other question is asked. The subtler problem is the wall. A cloud assistant with access to the practice's documents sees them as one corpus, and a matter wall inside the practice is as hard a boundary as the perimeter; an assistant that can retrieve across matters has breached it even if nobody noticed. The third problem is authority: nothing client-facing goes out without the responsible partner, and an assistant that drafts and sends in one motion has removed the partner from the step that matters. The last is the log. Who read what, and when, must be on a record the practice holds, by matter, and a vendor's usage report is not that record.
Drafting from the practice's own precedents
On one Czar in the practice, agents in Agency Studio assemble first drafts from the practice's own material, inside the matter. Retrieval serves the agent only from the matter's scope and from the precedents the practice has approved for use; Inference writes the draft; each passage carries a reference to where it came from, so the associate reviewing it can read the source instead of trusting the output. Memory is kept per matter, so what the agent learned on one file is not carried to the next. Adaptation can tune the model on the practice's own drafting, on the appliance, so that a first draft reads the way the practice writes.
Review and research, matter by matter
Review follows the same discipline. An agent reads the file and the incoming version and flags what differs: a clause that moved, a figure that changed, a representation that disappeared. It researches within the scope the matter sets, from the practice's own library and the sources the practice has declared, and it reports what it found and where. The partner decides what any of this means for the client and what goes out.
That is the boundary this solution is built around. Agents prepare and propose; a named person approves. Nothing customer-facing, client-facing, or production-changing leaves the appliance without that approval, and every step is on the record the team holds. In a practice the named person is the responsible partner, and the client never receives an agent's draft.
An access log the practice holds
Every retrieval and every draft is written to the record by matter: who asked, what the agent read, what it produced, who approved it, and when. The log is on the appliance, and it is the practice's. When the practice's own supervision asks what was accessed on a matter and by whom, or when a client asks the same question, the answer comes from one log rather than from a reconstruction. Evaluation scores drafts and reviews against work the partners have already signed off, so the practice can see where the agents are reliable. Watchman verifies what is running, and any run can be replayed to show exactly what the agent saw.
One Czar, placed and bought
The Czar is a sealed node in the practice's own rack, on the practice's own network, in a room the practice already secures. Its routes are declared one by one at commissioning: the document management system, the matter system, the precedent library, the research sources the practice already licenses. There is no default path out, and the appliance can run with none at all. The practice holds one subscription for the configured appliance, with the hardware, the platform license, and sustainment in one line, fixed for the term, so a new matter or a new agent does not move it. A solution brief for legal, audit, and advisory practices describes the intended architecture; it is substantiated at commissioning and makes no customer claim.
A practice has always been able to do this work; the question was whether it could have help with it without compromising the thing it is trusted with. One Czar inside the practice keeps the material inside the perimeter and inside the wall, puts the partner where the partner has always been, and leaves the practice holding the record of all of it.
Questions this raises
- How is a matter wall enforced on the appliance?
- Each agent in Agency Studio is authored for one matter and bounded to that matter's material. Retrieval serves it only from that scope, Memory is kept per matter, and the log records every retrieval by matter. An agent on one matter has no way to read another, and a run can be replayed to show exactly what it read.
- Can an agent send a draft to a client?
- No. Drafts are held in a queue for the responsible partner, who approves, amends, or rejects them, and the approval is written beside the draft. The route to the client is for the partner's action, not the agent's. The practice's own supervision is the boundary, and the appliance is built to keep it there.
- What does the practice hold afterward?
- The whole record, on its own appliance: who asked, what was retrieved, what was drafted, who approved it, and when, by matter. If a client or the practice's own supervision asks what was read on a matter and by whom, the answer comes from one log the practice holds rather than from a vendor's workspace.
KEEP READING