Legal & professional services

An ethical wall drawn in a policy memo binds people. An agent reads whatever its credentials reach, and when one client's strategy shows up in another client's draft, the remedy is a disqualification motion, not a bug report.

The wall has to exist as policy evaluated on each retrieval: this agent, working this matter, may read these sources and no others. Enforced there, the wall holds even when a prompt tells the agent otherwise.

A human who opens the wrong document knows it and can say so. An agent that retrieves across a wall blends what it read into its output, and nobody in the workflow sees the crossing.

After-the-fact review fails here for that reason. The one observable moment is the retrieval itself, so the record and the control both have to live there.

Clients are entitled to know when and how AI assists their matter, and confidentiality duties do not relax because a model did the reading. Bar guidance on AI use is still settling, but it is settling in that direction.

Engagement-letter language works best when it points at something real: named boundaries on what an agent may access for the matter, and a record of what it actually accessed. Consent to a described control is easier to obtain, and to defend, than consent to a capability.

Every retrieval an agent makes is recorded against the matter it served, with the policy that allowed it and any masking noted alongside. Grouped by matter, conflicts counsel can read it directly.

It also answers the client who asks. Showing what an agent read for their engagement, and only their engagement, is the concrete form of the confidentiality promise the letter makes.