Claude for legal work can summarise long contract sets, compare clauses, extract information and prepare first drafts. The label "Claude for Legal" does not change the basic requirement: the result depends on the workflow, the supplied sources and human review.
This guide focuses on concrete legal tasks and the questions that should be resolved before confidential documents enter the system.
What Claude can usefully do in contract work
Anthropic presents contract review, redlining, extraction and drafting as legal workflows. These tasks work best when there is a document, a defined instruction and an output that a lawyer can verify.
Useful requests include: "Compare the termination periods and renewal mechanisms in a table", "Find every liability cap and cite its clause", or "Draft an alternative using this approved standard". An open request to decide whether an entire agreement is "safe" lacks the party perspective, risk tolerance and governing legal sources needed for a reliable answer.
Five workflows with a reviewable result
1. Explain a clause
Ask separately for the wording, commercial effect, dependencies and undefined terms. Require clause references and an explicit statement where information is uncertain.
2. Extract information from several documents
Define the columns before the review, such as party, term, termination, liability and governing law. Check each row against the source. A neat table does not prove that extraction was complete.
3. Compare against an approved standard
Provide the standard and the reviewed agreement together. Ask Claude to separate the deviation, source clause and possible consequence. Without a reference, a model may invent an apparent market standard.
4. Prepare drafting
State the party interest, negotiation goal and limits. Ask for the smallest workable change, then inspect definitions, cross-references and consequences elsewhere in the document.
5. Verify the result
Legal claims need current primary sources. Check citations, jurisdiction and effective date independently. Claude can organise the work but is not a substitute for an authoritative legal database or legal approval.
Concrete example: reviewing a supplier agreement
A useful first test starts with a contract type for which an approved playbook exists, such as a supplier MSA. In the first pass, Claude extracts term, service changes, acceptance, warranties, liability, data protection, IP and termination with clause references. In the second pass, the model receives the approved standard and maps each deviation to a specific playbook position.
The output should separate three layers: what the agreement says, how it differs from the standard, and which question the legal team must answer. This makes clear whether Claude is extracting text or making an assessment. A statement that a liability clause is "market standard" is insufficient without the clause, comparison text and effect on the selected party.
Drafting comes only after that review. Each accepted issue receives the smallest workable redline proposal. A person then checks definitions, cross-references, precedence and commercial consequences in the full document. This creates an auditable sequence instead of one large prompt that produces an apparently final answer.
Claude is not one single product context
Personal Claude accounts, Claude for Work or Enterprise, and the Anthropic API have different contracts, administration and retention behaviour.
Anthropic says that commercial product chats and API content are not used for model training by default unless a customer explicitly opts in. Retention is a separate question. Under the published standard, API inputs and outputs are deleted within 30 days unless an exception or different agreement applies. Stateful products retain chats so users can continue them, while Enterprise provides custom retention controls. Zero data retention is a separate arrangement and does not automatically cover every product surface.
Location and data residency also require separate review. Anthropic states that commercial data may be processed across several regions while storage is US-only by default, unless otherwise agreed. Access from Europe therefore does not establish European storage.
Checks before using confidential documents
Confirm the contracting entity, product surface, DPA, current subprocessors and transfer mechanism. Determine how long chats, projects, files, logs and backups remain and which administrators can export data or enable features.
Review connectors and plugins separately. They can improve the workflow while extending the data sources and actions Claude can reach. Permissions, approvals and logging should match the task. Legal privilege and professional secrecy require an additional analysis beyond data protection compliance.
Compare Claude and CASUS by task
Claude is a general AI work system with chat, documents, integrations and an API. CASUS Risk Review focuses on structured contract review, while CASUS Chat Actions apply specific changes in the contract document. The CASUS security page documents the technical basis. A fair evaluation uses the same agreement and instruction for each system.
Measure source accuracy, missed clauses, correction effort, transfer into the document, permissions and data terms. A general assistant may fit research and drafting; an embedded legal workflow may reduce handoffs for repeatable contract reviews.
How to measure quality fairly
Define ten to twenty questions whose correct answers are already known from the agreement. Include easy fields and difficult cases such as exceptions, distributed definitions and conflicting schedules. For each question, record whether the source and answer are correct, incomplete or wrong.
Measure false positive findings as well. A system that produces many alleged risks may create more review work than one with a shorter, precise list. For drafting, count proposals that can be accepted without material correction and watch for unintended changes to definitions or neighbouring clauses.
The comparison is fair only where the product, documents, instruction and available sources remain the same. Record changes to the model, connectors or retention separately. The test should establish fit for the team's workflow rather than reward one impressive demo response.
A prompt for a controlled first test
"Use only the supplied documents. Create a table with topic, source clause, current wording, deviation from our standard and one open question. If information is absent, write 'not found in the document'. Do not give final legal approval."
Start with an approved or anonymised document. Record the product, settings, sources and expected output so the test can be repeated. To make a direct comparison, try CASUS.
FAQ
Is Claude a specialised legal AI model?
Claude is a general AI model and product ecosystem. Anthropic offers workflows and integrations for legal teams, but legal specialisation comes from sources, instructions, connected systems and the controlled process.
Does Anthropic train on contracts from commercial products?
Anthropic says commercial product content is not used for training by default unless the customer explicitly opts in. This does not answer the separate question of storage and retention.
Are Claude data stored in Europe?
Anthropic says commercial customer data are stored in the United States by default unless otherwise agreed, while processing may occur in several regions. Check the actual contract and configuration.
Is zero data retention sufficient for confidential legal work?
No. It can be one control, but it does not replace the DPA, subprocessor and transfer review, permissions, professional duties or verification of the output.
Which contract tasks suit Claude?
Bounded tasks with reviewable outputs work best: extraction, comparison, first drafting and summaries tied to source clauses. Final legal and commercial decisions remain with the legal team.
