Key findings

• Treat every AI tool as a third-party service requiring matter and vendor due diligence.

• Enforce matter permissions and ethical walls at retrieval, logging, and evaluation.

• Verify every legal citation and factual proposition against authoritative sources.

• Make reviewer identity and corrections part of the matter record.

• Do not automate legal judgment, filings, advice, or client commitments.

Use cases that assist rather than replace lawyers

Small firms can apply AI to intake extraction, document classification, deposition or interview summaries, first-pass chronologies, issue lists, matter-status drafts, billing narratives, and retrieval from approved internal knowledge. A tool can compare a draft with a checklist or identify passages that may require attention. These uses can reduce organization time, but the lawyer must decide relevance, strategy, privilege, accuracy, and what is communicated or filed.

Begin with an internal task whose source documents are available to the reviewer. A chronology draft is safer than an argument; a list of potentially relevant authorities is safer than a legal conclusion; a client-update draft is safer than unsupervised delivery. Generated research can omit controlling authority or fabricate citations. Therefore, use authoritative legal research systems for verification and require a lawyer to read the cited source, confirm that it exists, and assess whether it supports the proposition.

• Good pilots: intake fields, matter document routing, summaries, chronology drafts, checklists, and internal knowledge retrieval.

• Review-first uses: research leads, contract comparisons, discovery support, correspondence drafts, and billing narratives.

• Keep under lawyer control: advice, strategy, privilege decisions, filings, undertakings, settlements, and fee judgments.

A matter-centric reference architecture

Require user, client, matter, task, and purpose on every request. The orchestrator checks the firm's identity and access controls, retrieves only material the user can access, and filters by document status. The model returns a structured draft with document IDs, page or paragraph references, uncertainty, and missing-information flags. Policy rules block cross-matter retrieval, unapproved export, unsupported citations, and external delivery. The result enters the matter workspace for named lawyer review.

The practice management, document management, time, billing, and records systems remain authoritative. Store prompts and outputs under the matter's retention and access rules because they may contain client information or work product. Log source identifiers, model and prompt version, reviewer edits, approval, and destination. Separate development from live matters and use synthetic or properly authorized data for testing. Maintain manual procedures and a kill switch so model or retrieval failure never forces staff to bypass matter controls.

Confidentiality and vendor diligence

Before using a tool with client information, understand what data it receives, where it is processed, whether it is retained or used for training, who its subprocessors are, how access is controlled, and how data is deleted or returned. Review engagement terms, client instructions, protective orders, court rules, professional obligations, and applicable privacy law. Consumer-grade accounts and informal browser tools may not provide the controls a matter requires.

Apply least privilege, ethical walls, encryption, secure authentication, and matter-level logging. Do not paste an entire production, medical record, transaction file, or client database when a redacted excerpt is sufficient. Treat system prompts, evaluations, and support logs as potential client data. Test indirect disclosure and prompt injection in opposing or third-party documents. A vendor contract is only one control; the firm's configuration, supervision, staff behavior, and incident process determine how the service operates in practice.

Ethical oversight and independent verification

ABA Formal Opinion 512 addresses generative AI through existing duties including competence, confidentiality, communication, candor, supervision, and fees. It states that lawyers should understand relevant capabilities and limitations and review output for accuracy. For court submissions, lawyers must verify analysis and citations and correct errors. Firm policy should translate those duties into allowed uses, prohibited data, required review, client communication criteria, vendor approval, and training.

A supervising lawyer should be accountable for work produced with AI by lawyers and nonlawyers. Review means comparing factual assertions with the record, reading authorities, checking jurisdiction and currency, assessing omissions, and applying professional judgment. Do not bill model processing as lawyer time or pass through charges without considering applicable fee rules and disclosures. Preserve how the final work was verified. Courts and jurisdictions may impose additional requirements, so the responsible lawyer must check the rules that apply to the matter.

KPIs for quality, service, and risk

Measure intake turnaround, document classification precision and recall, chronology correction rate, time to a reviewed first draft, source-link coverage, citation verification failure, material edit rate, and reviewer minutes. For knowledge retrieval, have lawyers rate relevance and completeness against a defined question set. Track how often the tool abstains appropriately when evidence is missing.

Operational metrics can include matter cycle time, write-offs caused by avoidable rework, realization, and cost per completed administrative task, but do not let speed or margin override duty. Add confidentiality incidents, cross-matter access test failures, unsupported legal propositions, missed deadlines, filing corrections, client complaints, and time to disable a faulty workflow. Segment by matter type and task. A tool that performs well on standard commercial documents may be unsuitable for litigation, family, immigration, or criminal work.

A 90-day law-firm rollout

Days 1-30: choose one internal workflow, appoint a partner owner and security owner, map matter data and obligations, review vendor terms, baseline time and corrections, and create an evaluation set with representative documents, ambiguous facts, nonexistent citations, conflicting authorities, privileged content, sealed information, and prompt injection. Build read-only retrieval and disable external sending and filing.

Days 31-60: run shadow mode with a small trained team. Lawyers verify every source and record corrections. Test ethical walls, logs, deletion, outages, and whether the system exposes uncertainty. Days 61-90: allow a narrow production use inside selected matters with mandatory lawyer approval and matter-record retention. Audit an initial sample, review fee and client-communication implications, and train supervisors and staff. Expand only after documented quality, confidentiality, and competence review; each new task category receives its own evaluation.

• Day 30: approved vendor, matter boundary, owner, baseline, policy, and adversarial evaluation set.

• Day 60: shadow mode passes citation, factual, confidentiality, and supervision gates.

• Day 90: limited matter use has lawyer approval, audit evidence, fallback, and a documented scope decision.

When not to automate

Do not use AI when the firm cannot protect client information, enforce matter walls, verify authoritative sources, or assign a competent lawyer to review. Avoid it where a protective order, client instruction, court rule, or vendor term is incompatible with the proposed processing. Do not use generated text to fill evidentiary gaps or create confidence that the record does not support.

Keep legal advice, strategy, privilege, conflicts, client consent, negotiations, settlement authority, final contract interpretation, filings, and representations to tribunals under lawyer control. Do not permit a public chatbot to form an unintended relationship or deliver individualized legal advice without designed intake and supervision. Rules-based checklists may be better for deadlines and required fields. Manual work is appropriate for novel, high-stakes, or fact-sensitive matters where evaluation coverage is weak and an error cannot be readily reversed.

Sources and methodology

This article synthesizes the primary sources below as of the publication date. Forecasts and recommendations are directional scenarios, not guarantees; they should be tested against your workflow, data, risk tolerance, and current vendor documentation.

American Bar Association: ABA Formal Opinion 512: Generative Artificial Intelligence Tools (accessed 2026-07-30)

American Bar Association: ABA announcement on ethics guidance for lawyers' use of AI (accessed 2026-07-30)

National Institute of Standards and Technology: AI Risk Management Framework (accessed 2026-07-30)

National Institute of Standards and Technology: Privacy Framework (accessed 2026-07-30)