# AI for Legal, Compliance & Risk URL: https://justuju.org/role/legal/ For your role AI reads and drafts faster than any associate: research memos, contract redlines, document review. It also invents authorities with total confidence, and courts have sanctioned lawyers for filing them. The duty to check has not moved. ## How AI is changing this work - Research tools plan multi-step work and return memos with linked citations. - Contracts are reviewed against a playbook and redlined in minutes; intake and renewals are handled by agents. - Compliance teams must now govern the company's own AI use as well as use AI themselves. ## Key layers, most important first ### 1. Layer 9 · Applications & Interfaces URL: https://justuju.org/layer/applications-interfaces/ Why it matters: Legal AI arrives as assistants built for the profession, tied to case law, statutes and your documents. The choice of tool decides what sources stand behind an answer. What to do here: - Research with linked authority: Ask a legal question and get a memo whose every citation opens the source and its treatment. (see https://justuju.org/concept/chat-assistants/) - Playbook redlining: Review a contract against your standard positions and propose edits for a lawyer to accept. (see https://justuju.org/concept/copilots/) - Assistant in the drafting tool: Draft, compare and summarise inside the word processor, with firm precedents as the base. (see https://justuju.org/concept/embedded-ai/) Examples as of 2026-11-11: Harvey (https://www.harvey.ai/platform/workflow-agents); CoCounsel Legal (Thomson Reuters) (https://legal.thomsonreuters.com/en/products/cocounsel-legal/features); Lexis+ with Protégé (https://www.lexisnexis.com/en-us/products/lexis-plus-protege.page); Ironclad (https://ironcladapp.com/product/ironclad-ai) ### 2. Layer 7 · Orchestration & Tooling URL: https://justuju.org/layer/orchestration-tooling/ Why it matters: The difference between a usable answer and a sanction is grounding: answers built from retrieved, real sources, with the citation checked. That is built at this layer. What to do here: - Retrieve, then answer: The tool searches authoritative sources first and answers only from what it found. (see https://justuju.org/concept/rag/) - Review at document scale: Classify thousands of documents for relevance or privilege, each decision citing the passage. (see https://justuju.org/concept/rag/) - Verify every citation: Open each authority and confirm it exists and says what is claimed. Tools reduce errors; they do not remove them. (see https://justuju.org/concept/guardrails-evals/) Examples as of 2026-11-11: Relativity aiR for Review (https://www.relativity.com/blog/a-year-of-air-reflecting-on-2024-and-what-lies-ahead/); Deep Research in CoCounsel Legal (https://www.thomsonreuters.com/en/press-releases/2026/august/thomson-reuters-launches-next-generation-of-cocounsel-legal-the-ai-ecosystem-built-for-legal-professionals) ### 3. Layer 6 · Inference & Optimization URL: https://justuju.org/layer/inference-optimization/ Why it matters: Confidentiality turns on where the model runs and what the provider keeps. Those are questions about how the model is served. What to do here: - Know the data path: For each tool: where prompts are processed, how long they are kept, and whether they train anything. (see https://justuju.org/concept/model-apis/) - Contract for it: Use business terms that bar training on your inputs; consumer chat tools usually lack them. (see https://justuju.org/concept/model-apis/) - Run it yourself for the most sensitive work: An open model on your own hardware keeps privileged material in house. (see https://justuju.org/concept/local-inference/) Examples as of 2026-11-11: Amazon Bedrock (https://aws.amazon.com/bedrock/); Ollama (https://docs.ollama.com/api/openai-compatibility) ### 4. Layer 4 · Data URL: https://justuju.org/layer/data/ Why it matters: Training data is the centre of AI's biggest legal questions: copyright, licensing, personal data and what model makers must disclose. Your clients will ask. What to do here: - Training-data diligence: In supplier reviews, ask what the model was trained on and what licences or opt-outs were respected. (see https://justuju.org/concept/data-provenance/) - Opt-outs and licences: Know the text-and-data-mining rules where you operate and how rights holders reserve their rights. (see https://justuju.org/concept/data-provenance/) - Your documents as training data: Decide whether client or company material may ever be used to improve a supplier's model. (see https://justuju.org/concept/pretraining-data/) Examples as of 2026-11-11: EU Copyright Directive 2019/790 (text and data mining) (https://eur-lex.europa.eu/eli/dir/2019/790/oj); EU AI Act duties for general-purpose models (https://eur-lex.europa.eu/eli/reg/2024/1689/oj) ## High-leverage use cases - First-pass research memos: A cited memo on a defined question, with every authority opened and checked by a lawyer. - Contract review against a playbook: Flag departures from standard positions and propose redlines for approval. - Document review: First-pass relevance and privilege calls across large sets, with the reasoning shown. - Clause and obligation extraction: Pull parties, dates, renewal and termination terms from a contract portfolio into one table. - Policy and regulation tracking: Summaries of new rules mapped to your policies, with links to the official text. ## Start here: the next 30 to 60 days - Week 1 to 2: Write an AI use policy for the team: approved tools, what client data may go in, and the duty to verify. - Week 3 to 4: Pilot playbook review on one contract type, such as confidentiality agreements, and compare with lawyer review. - Day 30 to 60: Test a research tool on ten questions you already know the answer to. Count wrong and unsupported citations. - Day 30 to 60: Add AI questions to supplier due diligence: data path, retention, training use, model sources. ## Risks and pitfalls - Invented authorities: A US court fined lawyers $5,000 for filing a brief that cited cases a chat tool had made up. Source: https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2022cv01461/575368/54 - Specialist tools still err: A Stanford and Yale study found leading legal research tools gave wrong or unsupported answers in roughly 17% to 33% of queries. Source: https://arxiv.org/abs/2405.20362 - Professional duties apply in full: The American Bar Association's first opinion on generative AI applies competence, confidentiality, communication, supervision and fee rules to its use. Source: https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/ - Privilege and confidentiality: Client material in a tool that keeps or trains on inputs may be a disclosure. Check terms before use. - Regulated uses inside the company: Compliance must know where the business uses AI for hiring, credit or similar decisions, which the EU treats as high-risk. Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj ## Sources - Harvey (harvey.ai): https://www.harvey.ai/platform/workflow-agents - CoCounsel Legal (Thomson Reuters) (legal.thomsonreuters.com): https://legal.thomsonreuters.com/en/products/cocounsel-legal/features - Lexis+ with Protégé (lexisnexis.com): https://www.lexisnexis.com/en-us/products/lexis-plus-protege.page - Ironclad (ironcladapp.com): https://ironcladapp.com/product/ironclad-ai - Relativity aiR for Review (relativity.com): https://www.relativity.com/blog/a-year-of-air-reflecting-on-2024-and-what-lies-ahead/ - Deep Research in CoCounsel Legal (thomsonreuters.com): https://www.thomsonreuters.com/en/press-releases/2026/august/thomson-reuters-launches-next-generation-of-cocounsel-legal-the-ai-ecosystem-built-for-legal-professionals - Amazon Bedrock (aws.amazon.com): https://aws.amazon.com/bedrock/ - Ollama (docs.ollama.com): https://docs.ollama.com/api/openai-compatibility - EU Copyright Directive 2019/790 (text and data mining) (eur-lex.europa.eu): https://eur-lex.europa.eu/eli/dir/2019/790/oj - EU AI Act duties for general-purpose models (eur-lex.europa.eu): https://eur-lex.europa.eu/eli/reg/2024/1689/oj - Mata v. Avianca, Inc., No. 1:2022cv01461, Document 54 (S.D.N.Y. 2023): sanctions order: https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2022cv01461/575368/54 - Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (Magesh et al.): https://arxiv.org/abs/2405.20362 - ABA issues first ethics guidance on a lawyer's use of AI tools (Formal Opinion 512): https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/ --- Justuju · https://justuju.org/ · Content CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote with a link to the page.