The short answer
AI for lawyers in November 2026 is useful for first-pass contract review, summarising long records, drafting from firm precedents and plain-language client explanations, using work plans such as Claude Team or Enterprise, ChatGPT Business or Enterprise, or Microsoft 365 Copilot, where data is not used for training by default. In the US, ABA Formal Opinion 512 (29 July 2024) says lawyers need a client's informed consent before putting information about the representation into a self-learning tool, must check AI output, and may bill only for actual time.
| Good first uses | Contract first reads, record summaries, precedent-based drafts, translations, client letters |
|---|---|
| Keep human | Legal advice, every citation, filings, privilege calls, negotiation |
| US ethics guidance | ABA Formal Opinion 512, 29 July 2024, ABA Standing Committee on Ethics and Professional Responsibility |
| Informed consent | Required before inputting information relating to the representation into a self-learning tool; boilerplate is not enough |
| Plans with no training by default | Claude Team and Enterprise, ChatGPT Business and Enterprise, Microsoft 365 Copilot work accounts |
| Translation | DeepL paid plans: texts deleted after translation and never used for training |
How lawyers use AI in 2026
Legal work is mostly reading and writing under pressure, which is exactly what large language models are good at, and mostly confidential, which is exactly where general consumer AI tools are risky. The lawyers who get value from AI keep both facts in view. They use work plans with contractual data protection, they hand over first reads and first drafts, and they check every output as if a trainee wrote it.
The general assistants are capable enough for much of this. Claude from Anthropic is strong on long documents and careful drafting, with Claude Opus 5.5 on paid plans. Microsoft 365 Copilot works inside Word, Outlook and Teams for firms on Microsoft 365, and with a work or school account its prompts, responses and Graph data are not used to train foundation models. ChatGPT Business and Enterprise from OpenAI do not use business data for training by default. Many firms also license legal research platforms with their own AI, which are the right place for case law research. Our best AI for legal documents page compares the general tools.
The 7 legal tasks worth handing to AI
1. First-pass contract review (Claude)
Upload a contract and ask which clauses carry the most risk for your client, with the exact wording quoted for each so you can check it against the document. Claude handles long documents well. The review is a starting point for your own reading, not a substitute for it.
2. Summarising long records (Claude or Gemini Notebook on a work account)
Deposition transcripts, correspondence bundles and board minutes can be summarised by topic, with a citation to the page for every point. Gemini Notebook answers only from the sources you add; on work and school accounts Google says uploads, queries and responses are not reviewed by humans and not used to train AI models.
3. First drafts from your own precedents (Microsoft 365 Copilot, Claude or ChatGPT)
Give the assistant the firm's precedent, the deal terms and the points of difference, and ask for a first draft with every departure from the precedent flagged. Working from your own templates keeps the drafting in your style and limits invention.
4. Comparing versions (Claude)
Ask what changed between two drafts in substance, not just wording, and which changes shift risk. It is faster than reading a redline line by line, and you can then read the clauses it points to.
5. Research starting points, verified (Perplexity or your legal research platform)
Use AI to map an unfamiliar area: the main statutes, the key questions, the terms to search. Perplexity lists numbered sources in every answer, which makes it easier to check. Every case, statute and quotation must then be confirmed in an authoritative legal source before it goes anywhere near a client or a court.
6. Translation of documents (DeepL on a paid plan)
DeepL translates whole documents while keeping the layout, and glossaries fix how defined terms are translated. On paid plans DeepL says texts are kept only as long as needed for the translation, then deleted, and never used for training. The free version may use texts for training, so it is not for client documents.
7. Plain-language explanations for clients (any approved assistant)
Ask for a one-page explanation of a clause, a process or a decision for a non-lawyer client, then check it for accuracy and for anything that reads as advice you did not intend to give.
What lawyers should keep human
- Legal advice and judgement. The opinion the client relies on is yours, and so is the responsibility.
- Every citation and quotation. Courts have sanctioned lawyers for filing briefs with non-existent cases produced by AI. Nothing is cited until a person has read the source.
- Court filings and signed documents. Some courts now require disclosure of AI use in filings. Check the rules of each court you appear before.
- Privilege and confidentiality decisions. What may be shared with which tool is a judgement for the responsible lawyer, under the firm's policy.
- Negotiation and the client relationship. AI can prepare positions. People negotiate and advise.
ABA Formal Opinion 512: what it says about generative AI
In the United States, the American Bar Association's Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, Generative Artificial Intelligence Tools, on 29 July 2024. ABA opinions interpret the Model Rules of Professional Conduct and are not binding law; your own state's rules of professional conduct and its bar's guidance are what govern you. The main points:
- Competence. AI tools cannot stand in for a lawyer's own competent work, and relying on their output without checking it is risky. How much independent checking is needed depends on the tool and the task: reviewing documents needs more than generating ideas.
- Confidentiality. Before inputting information relating to the representation into a self-learning generative AI tool, a lawyer needs the client's informed consent. Boilerplate consent in an engagement letter is not enough. No consent is needed when nothing about the representation goes in, for example when using a tool for idea generation. In practice, a consumer plan that can train on your inputs is the clearest example of a tool that needs this care.
- Communication. Lawyers must reasonably consult clients about the means used to achieve their objectives, must disclose AI use if the client asks, and must tell the client in advance when AI use is relevant to the basis or reasonableness of the fee. The engagement agreement is a sensible place to record this.
- Candour and meritorious claims. Lawyers remain responsible for what they file. Unchecked AI output has led to filings with non-existent cases and misleading arguments, and courts have responded with disclosure requirements and sanctions.
- Supervision. Managerial lawyers should set clear firm policies on AI use, and supervising lawyers must make reasonable efforts to ensure others follow them.
- Fees. Lawyers billing hourly may bill only for time actually spent, and may not charge clients for time caused by their own inexperience with a tool. The opinion also explains when an AI tool's cost is overhead and when it may be charged as an expense.
Outside the US, check your own regulator's guidance; the principles of confidentiality, competence and supervision are similar in most systems, but the details differ. For the EU, see our explainer on the EU AI Act. This section summarises the opinion and is not legal advice.
A starter AI stack for lawyers
| Job | Tool | Plan | Why |
|---|---|---|---|
| Long documents and drafting | Claude | Team or Enterprise | No training by default; Enterprise adds custom data retention and audit logs |
| Inside Word, Outlook and Teams | Microsoft 365 Copilot | Work account with a licence | Enterprise data protection; sensitivity labels and retention apply |
| General assistant | ChatGPT | Business or Enterprise | No training on business data by default, SAML SSO |
| Source-grounded reading | Gemini Notebook | Google Workspace account | No human review or training on work accounts |
| Translation | DeepL | Paid plan | Texts deleted after translation, never used for training |
Read private AI for sensitive data for more on plan choice, and our Claude guide and Copilot guide for settings.
A 30-day AI rollout plan for a law firm or legal team
- Days 1 to 5: policy and plans. Write a firm AI policy: approved tools and plans, what may never go in, verification duties and who supervises. Ban personal AI accounts for client work. If anyone uses a personal Claude or ChatGPT account, they must switch off training (Claude: Settings, Privacy, Help Improve our AI models; ChatGPT: Settings, Data controls, Improve the model for everyone), though a work plan is the better answer.
- Days 6 to 10: engagement terms. Update engagement letters to describe how the firm uses AI, and design a specific informed-consent process for any use that needs it. Boilerplate alone does not meet Opinion 512.
- Days 11 to 17: low-risk pilot. Start with internal and public material: summarising published judgments, first drafts of internal memos, comparing your own templates.
- Days 18 to 24: supervised client work. Pilot contract first reads and record summaries on approved plans, with a named reviewing lawyer for every output and a log of time actually spent.
- Days 25 to 30: training and review. Train everyone on verification and confidentiality, review the pilot, and update billing practices so clients are charged only for actual time.
The risks for lawyers, and how to manage them
- Breach of confidentiality. The biggest risk. Consumer plans of ChatGPT, Claude and Perplexity can train on content unless you opt out, and DeepL's free version may use texts for training. Use work plans and get informed consent where Opinion 512 requires it.
- Fabricated authority. AI can invent cases, quotations and statutes that look real. Verify everything in an authoritative source.
- Privilege. Sharing privileged material with a third-party service may raise waiver questions in some jurisdictions. Take advice on your own position and prefer tools with clear contractual terms.
- Unreasonable fees. If AI makes a task faster, the hourly bill must reflect the time actually spent.
- Unsupervised use. Junior staff using unapproved tools is a supervision failure for the firm. Policy, training and approved tools fix most of it.
Prompts for drafting and review are in Claude prompts for writing and prompts for research. Use them only on approved plans.
Prompts for the 7 legal tasks
Use these only in tools your firm has approved for the material involved. Replace the square brackets.
Prompts to copy
1. Risk clauses with exact wording
I have attached a [type of contract]. We act for the [party]. List the ten clauses that carry the most risk for our client, ranked. For each, quote the exact wording with the clause number, explain the risk in one sentence and suggest a negotiating position. Mark anything that depends on governing law.
You get: A ranked list you can check clause by clause
2. A record summary with page references
Summarise this [transcript or bundle] by topic: [topics]. For every point, give the page and line or document reference. List any contradictions between witnesses or documents separately. Do not add anything that is not in the source.
You get: A topic summary you can verify against the record
3. A first draft from our precedent
Using our precedent below, draft a [document] for these terms: [terms]. Keep our structure and defined terms. Flag every place where you departed from the precedent and why, and list any terms I have not given you instead of inventing them. Precedent: [paste]
You get: A draft in your firm's style with every departure flagged
4. What changed between two drafts
Compare draft A and draft B attached. List every change in substance (not formatting), quote the old and new wording, and say whether each change increases or reduces risk for [party]. Put the most significant changes first.
You get: A substance-first comparison
5. Map an unfamiliar area of law
I need to understand [area of law] in [jurisdiction] for [type of matter]. List the main statutes and regulations, the key legal questions, and the search terms I should use in a legal research database. Give a source link for every statute. Do not cite any cases; I will research those myself.
You get: A research map to verify in an authoritative source
6. Translate with a glossary (DeepL)
Translate the attached [document] from [language] to [language], keeping the layout. Apply our glossary for defined terms: [term = translation]. Keep clause numbering and capitalised defined terms exactly as in the source.
You get: A translation that keeps defined terms consistent
7. Explain it to the client
Explain [clause or process] to a client who is not a lawyer, in plain English, under 300 words. Cover what it means, why it matters to them, and what happens next. Do not give advice beyond what is in my notes: [notes].
You get: A clear explanation to check before sending