AI Agents in Legal Practice
Chain of custody for digital decisions. When an AI agent drafts a motion or reviews a contract, the legal profession demands an unbroken audit trail. Most AI platforms cannot provide one.
The prudent pace
The legal profession is adopting AI tools with a caution that other industries might describe as deliberate and practitioners would describe as prudent. There are good reasons for that pace. Law is a profession built on evidence, provenance, and the sanctity of the record.
When a lawyer cites a case, the court expects that case to exist. When a document enters the evidentiary record, its chain of custody must be unbroken. When attorney-client communications are asserted as privileged, the basis for that assertion must be demonstrable.
AI agents threaten each of these foundations – not because they are unreliable in principle, but because most AI tools today produce outputs without producing proof. And in legal practice, output without proof is not merely unhelpful. It is dangerous.
The regulatory landscape
Since the Mata v. Avianca decision, courts and bar associations across multiple jurisdictions have moved to address AI in legal practice. The trajectory is clear: disclosure, competence, and accountability requirements are expanding.
Mata v. Avianca, Inc.
Attorneys Steven Schwartz and Peter LoDuca sanctioned for submitting six fabricated case citations generated by ChatGPT. No audit trail of the AI interaction existed.
▸ Triggered nationwide awareness of AI hallucination risk in legal filings
California's Practical Guidance
Issued guidance on generative AI use in the practice of law, addressing disclosure and competence obligations directly.
▸ Established state-level AI competence expectations
Florida Advisory Opinion 24-1
Addressed AI disclosure requirements and competence obligations for Florida practitioners.
▸ Reinforced supervision duties for AI-assisted work product
Standing Orders (Multiple)
Individual judges issued standing orders requiring disclosure of AI use in court filings.
▸ Created patchwork of disclosure requirements across jurisdictions
EU AI Act – High-Risk Obligations
AI systems in "administration of justice" classified as high-risk under Article 6(2) and Annex III, point 8. Conformity assessments, risk management, and human oversight required.
▸ Firms with European clients face dual compliance burden
The ethical architecture: ABA Model Rules and AI
The American Bar Association's Model Rules of Professional Conduct were written for human practitioners, but their obligations extend to any tool a lawyer employs. Three rules are directly implicated by AI agent adoption.
Lawyers must provide competent representation, including staying abreast of the benefits and risks associated with relevant technology.
AI implication: Understanding how an AI tool identifies clauses, what sources it draws from, and its known failure modes. Competence means understanding the AI methodology, not just its outputs.
Prohibits revealing client information without informed consent. Lawyers must make reasonable efforts to prevent inadvertent or unauthorised disclosure.
AI implication: When an AI agent processes merger agreements or litigation strategy memos, data flows through its pipeline. If that pipeline transmits data externally or retains it for training, confidentiality is breached.
Lawyers must ensure that non-lawyer assistants act compatibly with the lawyer's professional obligations.
AI implication: An AI agent performing legal research or document drafting is an assistant whose work product the supervising lawyer must be able to review, verify, and correct. You cannot supervise what you cannot see.
Keeping clients reasonably informed about how their matters are handled
Ensuring subordinate lawyers properly supervise AI tools
AI competence framework
Competence in 2026 means understanding the AI in your workflow – not just its outputs, but its methodology. The following framework maps the progression from baseline awareness to the auditability that trust infrastructure enables.
Risk scenarios: what goes wrong without audit trails
The legal risks of deploying AI without trust infrastructure are concrete and immediate. Each scenario below represents an exposure that firms face today.
AI agent processing privileged documents transmits data to a third-party service or stores it in a shared environment. A court finds privilege has been waived.
▸ In active litigation, privilege waiver can be catastrophic. Every AI interaction with client documents that cannot be fully documented is a potential breach.
A litigation hold does not encompass AI agent activity logs. Processing history is itself electronically stored information (ESI) under the Federal Rules of Civil Procedure.
▸ Failure to preserve AI processing records creates spoliation risk and potential sanctions.
AI performs contract due diligence but no record exists of what it reviewed, flagged, or missed. The acquirer later claims it was not informed of a material risk.
▸ Malpractice exposure and failure of supervision under Rule 5.3.
AI generates plausible-sounding case citations that do not exist. Without an audit trail, lawyers cannot demonstrate what prompt was given or what the system retrieved.
▸ Sanctions, reputational damage, and potential malpractice claims.
Chain of custody, privilege, and eDiscovery
In litigation, chain of custody is foundational. Physical evidence must be tracked from collection through presentation at trial. Each transfer is documented. Each custodian is identified. Any gap creates grounds for challenging admissibility.
AI agents introduce a new category of digital chain-of-custody requirements. When an AI agent performs contract due diligence, it reads documents, identifies clauses, flags risks, and produces a summary. Each step is an event in a chain. If the analysis is later challenged, the question becomes: what did the AI review, what did it flag, and what did it miss?
The privilege implications are equally severe. Attorney-client privilege can be waived inadvertently. If an AI agent processing privileged documents transmits data to a third-party service, a court could find that privilege has been waived. Demonstrating reasonable steps to prevent disclosure requires showing exactly how the AI handled the privileged material.
The Federal Rules of Civil Procedure require parties to preserve and produce electronically stored information (ESI) relevant to litigation. If an AI agent has processed documents as part of legal work, its processing history is itself ESI.
A litigation hold that does not encompass AI agent activity logs is incomplete. eDiscovery practitioners are already grappling with how to collect and produce AI processing records when those records do not exist in a structured, exportable form.
The EU dimension
The exposure extends beyond ABA rules. Under the EU AI Act, AI systems used in the "administration of justice and democratic processes" are explicitly classified as high-risk under Article 6(2) and Annex III, point 8 – triggering conformity assessments, continuous risk management, technical documentation, and human oversight requirements.
Deadline: 2 August 2026
With obligations for high-risk AI systems taking effect by this date, firms with European clients or matters face a dual compliance burden: ABA ethics obligations domestically and EU AI Act requirements internationally.
What trust infrastructure looks like for legal AI
The requirements are clear. Lawyers need AI tools that produce a verifiable record, maintain confidentiality through auditable data handling, enable meaningful supervision, and generate exportable evidence for eDiscovery. The question is what technical architecture satisfies these requirements.
Every event in an AI agent's processing pipeline – every document accessed, every clause identified, every risk flagged – recorded as a hash-chained entry in a tamper-evident ledger.
Legal benefit: Digital chain of custody with comparable integrity guarantees to a properly maintained physical evidence chain.
Every instruction given to an AI agent and every response it produces wrapped in a standardised envelope recording sender, recipient, scope, and context.
Legal benefit: Documentation that privilege assertions require: proof that the AI operated within the bounds of the attorney-client relationship.
Each stage of processing – ingestion, analysis, output generation – emits its own auditable trace across 11 defined layers (L0 through L10).
Legal benefit: Transforms the AI from a black box into a transparent pipeline. This is what competence under Rule 1.1 looks like in practice.
The ability to re-execute an AI agent's analysis against the same inputs with updated parameters – planned for VOLT v0.3. When available, this would enable verifiable quality assurance: re-run a review with the latest precedent on force majeure and verify the results. The difference from existing discovery review tools, which have long offered audit logs and re-execution, is cryptographic verification that the replay faithfully reproduces the original execution environment.
The path forward
The legal profession's caution about AI is not an obstacle to be overcome. It is a signal about what trust infrastructure must provide. AI agents can deliver extraordinary value – faster contract review, more comprehensive research, more consistent compliance monitoring. But that value is contingent on the ability to prove what the AI did.
In a profession where the record is everything, the firms that move first will not be those that adopt AI fastest. They will be those that adopt AI with the evidentiary infrastructure legal practice has always required – now extended to digital decisions.
The critical question
Every AI interaction with client documents that cannot be fully documented is a potential privilege breach waiting to surface in litigation. The privilege waiver question alone should concentrate the mind.
Quox (quox.ai) builds trust infrastructure for AI agent operations. Its open protocols – AEE for standardised agent messaging, AOCL for orchestration control and observability, and VOLT for cryptographic evidence chains – provide the accountability architecture that autonomous systems require.