What AI can and cannot do in legal work

This lesson is educational and is not legal advice. Laws, professional duties, court rules, contracts, and organizational policies vary by jurisdiction and situation. Use qualified legal counsel for decisions about actual legal rights, obligations, or strategy.

Legal work contains many language-heavy activities: finding terms, comparing versions, summarizing documents, organizing facts, drafting alternatives, and checking whether a document follows a template. Modern AI can often help with these activities because it can transform and classify text quickly.

That ability is easy to overread. A fluent response does not prove that the model identified the controlling law, received all material facts, interpreted authority correctly, or understood an organization’s risk tolerance. The model does not represent a client, hold a professional license, or accept responsibility for the result. It may not know that a source has been reversed, a regulation has changed, an exception applies, or a contractual definition changes the ordinary meaning of a word.

The useful distinction is between assistance and determination. Assistance makes human work easier to inspect. Determination commits the organization to a legal conclusion, strategy, representation, waiver, filing, or negotiated position. AI may prepare inputs for a determination, but an authorized and qualified person must make it.

With approved tools, permitted data, and human review, AI can support:

• creating a first-pass document inventory;
• extracting parties, dates, defined terms, obligations, notice periods, and renewal mechanics;
• comparing a draft against an approved playbook or clause library;
• generating issue lists and questions for a reviewer;
• summarizing a source supplied in the prompt, with references to its sections;
• proposing search terms, research paths, and alternative interpretations;
• converting a lawyer-approved position into a plain-language explanation;
• producing a draft that clearly marks assumptions and missing inputs;
• checking internal consistency, such as mismatched dates or undefined terms.

These are bounded tasks. They have identifiable inputs, observable outputs, and review criteria. “List every notice deadline in these five contracts and quote the supporting language” is more controllable than “Tell me if these contracts are safe.”

AI should not be the final decision-maker for whether to sign, sue, settle, disclose, waive, file, terminate, report, or accept risk. It should not independently:

• decide that conduct is lawful or a document is enforceable;
• establish an attorney-client relationship or give personalized legal advice;
• choose litigation or negotiation strategy;
• certify that research is complete or current;
• approve a filing, representation, legal hold, or regulatory response;
• determine whether privilege applies or has been preserved;
• promise an outcome to a client, counterparty, regulator, or court;
• make a high-impact decision about a person based only on generated analysis.

Some tasks are unsuitable even as assistance when the tool is not approved for the data, the source set is incomplete, or the output cannot be meaningfully reviewed. Speed is not a benefit if nobody can reconstruct what the model saw and why its output changed the work.

Before using AI, write a five-part boundary:

1. Job: the narrow transformation or analysis requested.
2. Inputs: the authorized documents and facts the model may use.
3. Non-goals: decisions and conclusions it must not make.
4. Evidence: quotations, section references, or source links required.
5. Gate: the person who reviews the output and decides the next action.

For example: “Using the approved template and this draft, identify changed limitation-of-liability language. Quote both versions and label additions, deletions, and ambiguity. Do not assess enforceability or recommend acceptance. Counsel reviews every item before the draft is circulated.”

That boundary converts a vague request into a reviewable work product. It also makes stopping conditions visible. If the model cannot access an attachment, if definitions are missing, or if jurisdictions conflict, it should return an open item rather than complete the story.

Not every use needs the same process. Reformatting a public policy may need ordinary proofreading. Summarizing confidential investigation notes, analyzing a termination right, or preparing language for a court filing requires tighter access controls, source verification, specialist review, and documented approval.

Ask three questions: What could happen if the output is wrong? How easy is the error to detect? Who has authority to rely on the result? High consequence, low detectability, or unclear authority means stronger controls—or no AI use at all.

The goal is not to remove humans from legal work. It is to use automation where it improves coverage and consistency while preserving the judgment, duties, and accountability that the system cannot supply.