AI is changing the unit of work inside the firm

For decades, many legal tasks were organized around hours devoted to locating, sorting, comparing and drafting information. AI can reduce part of that preparatory work, but it does not remove the need to understand the problem or validate the answer.

Work may shift from initial production toward question formulation, critical review, negotiation, strategy and client communication.

AI can reduce the time required to produce an answer. It cannot determine whether that answer is legally correct, strategically useful or appropriate for the client.

Use cases carry different levels of risk

Controls should increase whenever an output may affect rights, deadlines, strategy, confidentiality, assets or substantive decisions.

AI-assisted research still requires returning to the original source

Generative systems can help formulate questions, identify concepts and organize a research path. They can also invent authorities, combine jurisdictions or present amended law as current.

• Define the jurisdiction, date and legal question.
• Request identifiable sources rather than conclusions alone.
• Open and review every original document.
• Confirm currency, authority and scope.
• Check citations, provisions, docket numbers and names.
• Identify exceptions and conflicting authority.
• Separate client-provided facts from system assumptions.
• Record who verified the research.

Contract review may become faster, but should never become automatic approval

AI can extract dates, parties, obligations, liability limitations and deviations from a template. Its usefulness depends on the firm having defined what it is looking for and under which standard.

The process should preserve the original document, automated output, corrections and final approval. Without traceability, the firm cannot explain how a conclusion was reached.

The deepest opportunity may exist within internal knowledge

Law firms accumulate templates, opinions, precedents, research and experience that frequently remain dispersed across email, folders and individual systems.

An AI layer may improve access, but it cannot independently correct duplicated, outdated or poorly classified information.

• Define which documents may be included.
• Remove duplicates and obsolete versions.
• Assign owners responsible for updates.
• Separate general knowledge from client information.
• Apply permissions by practice, team and matter.
• Display the source and date of every answer.
• Allow users to open the underlying document.
• Record questions, answers and corrections.

AI also changes the client experience

Clients may expect faster responses, greater matter visibility and clearer formats. This does not mean that every communication should be automated.

A chatbot should not present itself as a lawyer or provide personalized legal advice outside the appropriate professional process. It should explain its role, limitations and route to human assistance.

Risk extends beyond inaccurate answers

The firm must understand what happens to every item of data entered

• Does the provider use information to train models?
• Where is information stored and processed?
• Which subprocessors are involved?
• How long is information retained?
• Can it be deleted verifiably?
• How is data encrypted?
• Which administrative logs are available?
• Can access be restricted by user and matter?
• What happens at contract termination?
• Which incident-notification commitments apply?

Human supervision should be designed rather than assumed

Stating that an output will receive “human review” is insufficient when nobody knows what to review, which source to use or who has authority to approve.

A useful policy distinguishes tools, data and levels of risk

• Inventory of approved and prohibited tools.
• Classification of public, internal, personal and confidential information.
• Authorized use cases by level of risk.
• Review and approval requirements.
• Rules concerning research, citations and sources.
• Protocols for external content and communications.
• Access, security and retention controls.
• Procedure for incidents and unreliable results.
• Initial training and periodic updates.
• Continuous assessment of providers and models.

The policy should not be so restrictive that lawyers turn to hidden tools, nor so broad that every experiment becomes an authorized practice.

How to evaluate an AI tool for legal use

AI will transform roles and working models

Junior lawyers may spend less time on mechanical tasks, but they will need to learn how to research, review and challenge outputs at an earlier stage.

Partners will need to decide how efficiency affects fees, budgets and value propositions. The use of a tool does not automatically justify charging more or less; the firm should explain the value, risk and professional work involved.

• Train people to frame problems—not merely write prompts.
• Teach source verification and error detection.
• Preserve opportunities to learn legal fundamentals.
• Develop knowledge, data and legal-operations roles.
• Update models of supervision and delegation.
• Review budgets, efficiency and client expectations.
• Avoid measuring adoption only through user counts.

Value should be measured alongside quality and risk

A 90-day implementation plan

A strong pilot does not seek to prove that the tool works. It seeks to discover the conditions under which it works, the errors it produces and the controls it requires.

Common AI adoption mistakes

• Purchasing before defining the problem. The tool has no clear use case.
• Entering confidential information. The team does not understand the configuration or terms.
• Trusting plausible answers. Original sources are not reviewed.
• Automating a defective process. Errors are reproduced more quickly.
• Assigning no accountability. Everybody uses the tool, but nobody controls the risk.
• Measuring only time saved. Corrections, quality and exposure are ignored.
• Publishing without review. The firm loses voice, accuracy and credibility.
• Removing junior development. Opportunities to build professional judgment disappear.
• Depending on one vendor. There is no portability or continuity plan.
• Confusing adoption with innovation. Using AI does not prove that service improved.

Frequently asked questions