Legal AI is often sold as faster research. Legal work is broader: intake, conflict checks, chronology, evidence, contracts, diligence, hearings, advice, filing, billing, and compliance. Each stage has different data, authority, error cost, and review.
A defensible enterprise strategy uses specialized models around the matter lifecycle. Speech models capture proceedings. Document models extract clauses and facts. RAG retrieves authorized sources. Small language models classify, structure, and draft. Human lawyers remain responsible for professional judgment and final work product.
Adoption has accelerated
Thomson Reuters reported that generative-AI use in legal organizations rose from 14% to 26% in its 2025 survey of nearly 1,800 professionals. Among users, 33% in law firms and 46% in corporate legal departments accessed GenAI multiple times a week.
Demand is also coming from clients. Thomson Reuters found that 57% of corporate clients wanted outside firms to use generative AI, while 71% did not know whether their firms were doing so. Only about 20% of organizations said they measured return on investment. Adoption is outpacing evaluation.
The reliability warning
A 2025 peer-reviewed Stanford evaluation tested more than 200 legal queries on Lexis+ AI, Westlaw AI-Assisted Research, Ask Practical Law AI, and GPT-4. The specialist legal tools hallucinated between 17% and 33% of the time under the study's definition. Lexis+ AI answered 65% of queries accurately, Westlaw AI-Assisted Research 42%, and Ask Practical Law returned incomplete answers for more than 60% of queries.
The paper found that RAG reduced hallucination relative to a general model, but did not eliminate it. Its conclusion is compact: "RAG systems are no panacea."
The products may have changed since testing, and the query set cannot represent every practice. The enterprise lesson remains: citations and legal branding do not replace local evaluation and lawyer verification.
The model stack across the legal workflow
The model stack across the legal workflow
Intake and matter opening
Speech or text models capture the client's account, entities, dates, jurisdictions, adverse parties, and requested outcome. A deterministic service runs conflict checks.
Legal speech intelligence
Domain ASR transcribes depositions, hearings, interviews, consultations, calls, and recorded evidence.
Matter-specific RAG
A matter RAG system retrieves pleadings, evidence, correspondence, contracts, transcripts, research, and internal work product subject to ethical walls and permissions.
Chronology and fact extraction
Document and language models extract people, organizations, dates, events, amounts, obligations, and source spans. They can build a draft chronology or fact table.
Contract intelligence
Models can identify clause types, extract terms, compare against a playbook, route deviations, and draft approved language.
Research and drafting
RAG can retrieve authority and produce a research memo skeleton. The lawyer verifies propositions, treatment, jurisdiction, date, and citation.
Intake and matter opening
Speech or text models capture the client's account, entities, dates, jurisdictions, adverse parties, and requested outcome. A deterministic service runs conflict checks. An SLM classifies the matter and identifies missing intake fields.
Measure required-field completeness, entity accuracy, conflict false negatives, time to open, and abandonment.
Legal speech intelligence
Domain ASR transcribes depositions, hearings, interviews, consultations, calls, and recorded evidence. It must recognize case names, citations, legal phrases, speaker roles, and exhibits.
A 2025 UK Supreme Court-focused ASR paper reported a 9% accuracy improvement for a custom legal speech system over its generic baseline. Remote Legal reported changing its transcription architecture after a provider model update introduced phantom text, then achieving 85% lower test-environment cost with a new stack. The cases show why legal speech needs both domain adaptation and regression tests after model changes.
Matter-specific RAG
A matter RAG system retrieves pleadings, evidence, correspondence, contracts, transcripts, research, and internal work product subject to ethical walls and permissions. It should distinguish record fact, party allegation, legal authority, and attorney analysis.
Every answer needs passage-level citation, document status, date, jurisdiction, and privilege state. The system should abstain when the corpus does not support the claim.
Chronology and fact extraction
Document and language models extract people, organizations, dates, events, amounts, obligations, and source spans. They can build a draft chronology or fact table. The lawyer verifies material entries.
Measure entity and event precision, source-span accuracy, contradiction detection, and review time. Missing an event is different from misstating one, so precision and recall both matter.
Contract intelligence
Models can identify clause types, extract terms, compare against a playbook, route deviations, and draft approved language. A deterministic rules layer should check dates, amounts, defined terms, and cross-references.
Measure clause recall, deviation precision, accepted suggestions, review time, and post-sign leakage of nonstandard terms.
Research and drafting
RAG can retrieve authority and produce a research memo skeleton. The lawyer verifies propositions, treatment, jurisdiction, date, and citation. Generated text should never enter a filing without source-level review.
The 2023 Mata v. Avianca sanctions order imposed a $5,000 penalty and required notice to judges falsely identified as authors of fabricated opinions. It remains a concrete demonstration that plausible citations create professional and financial consequences.
Compliance monitoring
A regulatory RAG system tracks effective rules, obligations, controls, policies, and evidence. Specialized classifiers map new requirements to control owners. Human counsel decides legal interpretation and materiality.
Deployed solution examples
Harvey reports that Pierson Ferdinand users save about 10 hours per attorney per week, with power users reporting 15 to 20. Shumaker estimates seven to ten hours weekly. Harvey's 2024 review reported an average of three hours a week across users and deployment to more than 150 Repsol legal professionals.
These are vendor-published customer estimates, not audited time studies. They demonstrate adoption and perceived value. A buyer should instrument exact tasks, accepted outputs, and realized capacity.
Reference architecture
- Matter identity: client, matter, team, jurisdiction, ethical wall, and retention.
- Content layer: native files, OCR, email, transcript, metadata, and versions.
- Authority layer: primary law, citator status, effective date, and jurisdiction.
- Model router: extraction, classifier, ASR, SLM, larger model, and deterministic checks.
- Permission-aware RAG: retrieval enforced before model generation.
- Work product layer: draft, review, approval, filing, and knowledge capture.
- Audit layer: source, prompt, model, output, edits, approver, and final use.
Evaluation scorecard
| Workload | Model metric | Legal or business metric |
|---|---|---|
| Research | proposition accuracy, citation support, treatment | verified research time |
| Matter RAG | recall, groundedness, permission leakage | time to evidence, missed document |
| Contract | clause recall, deviation precision | review time, fallback acceptance |
| Chronology | event precision and recall | preparation time, correction count |
| Legal ASR | legal-entity and speaker accuracy | correction time, transcript turnaround |
| Drafting | factual consistency, source coverage | accepted text, review time |
| Compliance | obligation mapping accuracy | control coverage, response time |
A sensible first deployment
Start with a bounded internal workflow such as clause extraction, transcript summarization, or knowledge retrieval. Build a test set from closed matters with lawyer-approved answers. Include outdated authority, ambiguous language, missing sources, and permission traps.
Require a measurable gate: for example, at least 95% clause recall for defined high-risk clauses, zero unauthorized retrieval in the test set, and 30% lower review time without increased material corrections.
The strategic conclusion
The durable legal-AI asset is not access to a foundation model. It is the firm's matter structure, permissions, authority metadata, expert corrections, and evaluation set.
Specialized models make that asset operational. They narrow the task, expose evidence, control cost, and keep lawyer judgment where consequence demands it.
Research note
Research is current through September 5, 2026. Product performance may change after published evaluations. Vendor customer metrics are labeled. This article is not legal advice.
Continue the research
- How to build trustworthy legal RAG
- Legal speech recognition for hearings and matters
- Why enterprises should use small language models
- How enterprise model routing balances quality and cost
- Sovereign AI for government and public services
Building a Production-Ready System
Conscious Engines builds legal AI solutions around controlled sources, matter permissions, legal vocabulary, and reviewable outputs. Our stack can combine legal speech recognition, task-specific document models, enterprise RAG, and workflow agents while keeping professional judgment and legal authority with accountable people.