What Should Never Be Fully Delegated to a Legal AI Agent?

The useful question is not whether an AI agent can technically perform an action. It is whether the firm can assign the action, information, consequence, verification, and remedy to an accountable person. Legal work combines factual uncertainty, professional judgment, client objectives, authority, deadlines, and duties that a product cannot assume on the lawyer’s behalf.

An AI system may help collect, compare, classify, draft, calculate, monitor, or propose. It should not become the final decision maker for legal advice, strategy, conflicts, privilege, client disclosure, citation validity, filing, signature, production, sending, deadline satisfaction, settlement authority, or other actions whose error can materially affect a client or tribunal.

Human review is meaningful only when the reviewer has the competence, authority, time, sources, and context to detect a bad result. A generic approval button does not establish supervision. Define what the system may prepare, what evidence it must provide, which conditions require escalation, and what the reviewer must independently confirm before authorizing downstream use.

The American Bar Association’s Formal Opinion 512 identifies duties involving competence, confidentiality, client communication, supervision, meritorious claims, candor, and reasonable fees. It also says lawyers cannot rely solely on generative AI for functions requiring personal judgment and remain responsible for client work. Applicable state rules, court orders, client commitments, insurer conditions, and matter circumstances may add stricter requirements.

Require the system to stop when sources conflict, authority is missing, jurisdiction is uncertain, a deadline cannot be verified, a recipient is ambiguous, restricted information appears, or a requested action exceeds the approved tool scope. A well-designed workflow reports uncertainty and preserves evidence instead of hiding it behind confident prose.

Reviewers must be able to reject, correct, narrow, or rerun the work without pressure to approve automation. Record important overrides and recurring failure patterns. If people routinely click through a gate, the organization should redesign the workload, staffing, interface, or authority rather than treating nominal human involvement as a control.

Start in observation, draft, or recommendation mode. Separate retrieving information from changing it, and separate preparing external communication from sending it. Use allowlisted tools, limited destinations, least privilege, transaction previews, explicit confirmation, logging, and rollback where available. Test prohibited and ambiguous requests, not only the happy path.

Some activities may remain permanently human even after strong pilot results because their consequence, professional character, or remedy makes automation inappropriate. This is not a judgment that legal AI lacks value. It is an operating design that places automation where it can reduce routine effort while preserving human responsibility where judgment and accountability matter most.