AI for legal research: where it helps, where it gets lawyers sanctioned

Where legal AI adoption stands

Legal adoption of generative AI is accelerating fast: per the Thomson Reuters 2025 Generative AI in Professional Services Report, the share of legal organizations actively using generative AI rose from 14% in 2024 to 26% in 2025, and 78% of legal professionals expect it to become central to their work within five years.¹ The demand is coming from clients too, with 59% of corporate legal clients wanting their outside firms to use the technology.

Usage is becoming routine, well past the experimental stage. Nearly 70% of law-firm respondents who use generative AI report using it at least weekly, with the top reported tasks being document review at 77%, legal research at 74%, and document summarization at 74%.¹ Sentiment has flipped with it: in 2024 hesitancy led at 35%, while in 2025 the leading feelings were hopefulness at 28% and excitement at 27%.

The gap that matters for any build is readiness, and enthusiasm has run well ahead of it. Only 41% of organizations have established a governance policy, only 40% provide staff training, and only 20% measure ROI, while 71% of corporate clients do not even know whether their outside counsel uses the technology.¹ That readiness gap, more than raw adoption, is what a careful project has to close. This page supports the broader AI agents guide; here the focus is the legal angle and its specific duties of care.

Use cases that actually fit legal work

The use cases that fit are agentic workflows where the agent plans, retrieves, drafts, and cites, then routes the result to a lawyer: contract review and redlining, first-draft drafting, legal research, due diligence, e-discovery, matter intake and triage, and compliance monitoring. The pattern that matters is plan, retrieve, draft, cite, and route for human review, which is a step beyond a one-shot chatbot prompt.

The recurring patterns map cleanly to where firms already spend time. The table below lays them out; each is delivered through our AI agent development work, grounded for the legal sector in our legal engineering practice. Read every row with the same caveat, that the agent proposes and a lawyer disposes.

How to build a defensible legal AI agent

You build a defensible legal AI agent by connecting it to the firm’s document systems and trusted legal databases, grounding it with retrieval-augmented generation over primary law and the firm’s own precedent, requiring an auditable citation on every assertion, and wrapping it in privilege filters, citation checks, audit logs, and a mandatory attorney sign-off. It is an orchestrated, retrieval-grounded, human-in-the-loop system, a world away from a chatbot pointed at the open web.

Work through it as a layered architecture.

1. System integrations. Connect to the document management system, contract repositories or contract lifecycle management, matter-management systems, and trusted legal databases, so the agent works from vetted institutional sources instead of the open internet.
2. RAG over authoritative corpora. Ground generation in primary law, meaning case law, statutes, and regulations, plus the firm’s own precedent, templates, and playbooks. Retrieval narrows the model to vetted sources instead of its parametric memory, which is the main lever for reducing hallucination.
3. Citations and grounding. Every material assertion carries an auditable citation back to the source document or authority. Grounding the output to retrieved text rather than the model’s recall is what makes a draft checkable, and checkability is the whole point in a filing.
4. Confidentiality and privilege filters. Apply privilege and confidentiality controls at the data boundary so client material is not exposed to an uncontrolled model. Under ABA Model Rule 1.6, putting client information into a self-learning tool can be improper disclosure, so data handling is a design constraint from the outset.⁵
5. Guardrails and human-in-the-loop. Add confidence thresholds, citation-existence and quote-verification checks, scope limits on autonomous actions, and full audit logs, ending in a mandatory attorney review-and-sign-off gate before anything is filed or sent. The agent proposes; a lawyer accepts, rejects, or refines.

The framing that keeps this defensible: the agent is built for verification, confidentiality, and supervision from day one, so a licensed attorney can stand behind the output. That is the work our AI agent development team does, grounded in the sector context our legal engineering practice brings.

How accurate is AI for legal research, and the sanctions risk

The defining risk of AI for legal research is hallucinated citations, and it is documented and sanctionable even in purpose-built legal tools. A peer-reviewed Stanford RegLab study found leading commercial legal research tools hallucinated roughly 17% to 34% of the time, and courts have sanctioned attorneys for filing AI-generated cases that do not exist, starting with Mata v. Avianca in 2023.³⁴ A tool being built for law does not make its output safe to file unverified.

Lead with the evidence before any reassurance. In Mata v. Avianca, Inc. (S.D.N.Y., 2023), plaintiff’s attorneys submitted a brief citing six cases that did not exist, generated by ChatGPT, and the court imposed a $5,000 sanction, stressing that the deeper failure was not catching the error and then standing behind the fake cases.⁴ The problem then scaled instead of fading: a research database tracking AI-hallucinated content in court filings logged hundreds of cases worldwide through late 2025, the majority decided that year, with sanctions commonly running from about $1,500 to $6,000 and escalating to mandatory training, bar referrals, and disqualification.²⁷

The Stanford finding is the part that should change how teams scope a build. The study, published in the Journal of Empirical Legal Studies, tested leading commercial tools and found Lexis+ AI answered about 65% of queries accurately while hallucinating more than 17%, and Westlaw AI-Assisted Research was accurate about 42% of the time while hallucinating more than 34%.³ The table below is the cautionary centerpiece of this page.

Why attorney review is non-delegable

Attorney review is non-delegable because ABA Formal Opinion 512 (2024) places the duties of competence, candor, confidentiality, and supervision on the lawyer, and lawyers cannot blindly rely on AI output. An agent can draft and cite, but only a licensed attorney can take professional responsibility for a filing, so every substantive AI-generated element must be reviewed and verified by a responsible lawyer before it is used.

The duties are specific. Under ABA Model Rule 1.1 as applied by Formal Opinion 512, a lawyer must understand a tool’s capabilities and failure modes, including hallucination and knowledge-cutoff limits.⁶ Rule 3.3 on candor to the tribunal means unverified AI citations cannot be filed; Rules 5.1 and 5.3 treat the tool as a non-lawyer assistant the firm must supervise; Rule 1.6 governs confidentiality, and Rule 1.5 means a client should not be billed for time the AI saved as if it had been done by hand.⁵

Put the three threads together and the conclusion is plain. Purpose-built tools still hallucinate, courts are actively sanctioning attorneys for unverified output, and the ABA rules make verification a duty the lawyer cannot hand to software. The honest position, and the one a defensible build is engineered around, is that an AI agent augments lawyers and never replaces attorney review and judgment. The agent does the repetitive drafting, research, and review; the lawyer supervises, verifies, and decides.

Kanika Mathur

Head of Service Delivery, Resourcifi

Kanika Mathur is Head of Service Delivery at Resourcifi, where her engineering pods build retrieval-grounded agents wired to a firm’s document systems and primary law, with citation verification and privilege-aware data handling. She has scoped the review gates and audit logging that keep a licensed attorney in the loop on every output, which is the lens this guide is written from. It never claims software can practice law.

Resourcifi on LinkedIn →

Keep reading

Related guides worth your time

← →

Agents & RAG Agentic RAG: When to Use It and How to Build It Agentic RAG explained: how it differs from naive and advanced RAG, the key patterns like corrective RAG and self-RAG, the... Read guide → Agents & RAG AI Agent for Fintech: Risk, Compliance, Ops, Customer AI agents in finance: fraud, AML, KYC and servicing use cases, how to build with money-movement guardrails and human appr... Read guide → Agents & RAG AI Agent for Healthcare: Use Cases, Governance & Implementation AI agents in healthcare: the use cases that pay off first, how to build one HIPAA-safe on FHIR with clinician review, and... Read guide → Agents & RAG AI Agent for HR: Recruiting, Onboarding, People Ops AI agents for HR: screening, employee Q and A and onboarding use cases, how to build them, and the bias, EEOC and Local L... Read guide → Agents & RAG AI Agent for SaaS: How to Embed Autonomous Agents in Your Product AI agents' disruptive impact on the SaaS industry in 2025: Gartner sees agentic AI at 30% of app-software revenue by 2035... Read guide → Agents & RAG AI Agent for Sales: Pipeline & Outreach Automation AI agents for sales: lead qualification, outreach and CRM use cases, how to build with guardrails on autonomous outreach... Read guide → Strategy, architecture & ops AI Architecture Patterns Agentic design patterns explained: reflection, tool use, planning, and multi-agent collaboration, with a framework to pic... Read guide → Strategy, architecture & ops AI Architecture Patterns for SaaS: A Technical Guide Generative AI architecture for SaaS: layered design, multi-tenant isolation, LLM gateway, RAG, and security. Built by Res... Read guide → Building AI AI Copilots for SaaS: Build vs Buy Guide AI copilot vs AI agent for SaaS: a copilot assists, an agent acts. How an in-app copilot works, the RAG and multi-tenant... Read guide →