What Generative AI Can and Cannot Do in Legal Work

Hello, I’m Noriaki Asato, representative attorney at LegalAgent.

Ask generative AI to review a services agreement and it may produce polished comments on liability and intellectual property. Sending those comments to the counterparty requires further review. Does the response quote the actual clause? Has it confused the customer with the service provider? Is it trying to change a condition the business team has already accepted? Fluent writing can make it easier to overlook these questions.

In legal work, I believe it is important to distinguish the tasks assigned to AI from the decisions the company must own. AI can assist with drafts and document comparisons, while people verify the evidence, commercial background and approval process. Contract review provides a useful example of where to draw that boundary.

The Role of Generative AI in Legal Work

Generative AI can help summarize agreements, identify issues, and draft amendments and responses. Given documents and internal standards, it can also compare them and suggest approaches. That does not guarantee that it has understood every document correctly or reflected all relevant information in its conclusions.

For example, an agreement may contain a liability cap in its main text and exceptions in a schedule. Providing only the main text and receiving an answer that liability is limited does not amount to reviewing the entire agreement. Identify the relevant agreement, schedules, purchase orders and amendments, and keep a record of the materials supplied.

Services equipped with search, document retrieval or calculation tools may be able to research and check facts beyond their training data. OpenAI’s explanation describes how tools such as search can improve accuracy while errors and fabricated citations remain possible. There is no need to treat every AI system as incapable of researching current information. It remains essential to distinguish the functions used from the materials actually checked.

When allocating work, consider both whether AI can produce a proposal and who will verify and approve it.

Tasks AI Can Assist With and Decisions People Must Make

Tasks that can readily be checked against source text or a defined standard are good starting points. Examples include extracting contract terms and termination notice periods, listing exceptions to confidentiality duties, and explaining differences between versions. Separating a lengthy business email into confirmed facts, requests and missing information can also help start a review.

Ask for clause numbers and source quotations alongside summaries. A statement that a contract lasts one year may be inadequate if it omits renewal conditions or early termination rights. Give particular attention to numbers, deadlines, exceptions and conditions when checking the result.

Whether to accept a liability cap, by contrast, cannot be determined from the clause alone. The contract value, foreseeable incidents, insurance coverage, commercial importance and availability of alternatives all matter. AI can compare options and reasons using that information, but the decision about the risk the company will accept remains with authorized staff and decision-makers.

The same applies to disputed legal interpretations and negotiation strategy. When using AI analysis as input, verify that cited statutes and cases exist, apply at the relevant time, and concern comparable circumstances. The existence of a statutory provision and its application to the transaction are separate questions.

Example Instructions for Preparing Decision Materials

A broad initial question such as “Can we sign this agreement?” can encourage assumptions to fill gaps in the materials. First define the task and output, and require the system to leave unknown matters unresolved. The following examples assume an approved environment and documents that may be entered into it.

For the following non-disclosure agreement, please:

1. Extract the definition of confidential information, restrictions on use,
   and obligations that survive termination.
2. Show the clause number, source quotation and summary separately.
3. Include exceptions and cross-references to other clauses.
4. Distinguish matters not established by the text from matters requiring
   additional documents.

If you cannot find a provision, say "not confirmed" rather than filling
the gap with an assumption. Do not determine whether the parties have
agreed to a condition or whether the agreement should be accepted.

[NDA and schedules approved for use in this environment]

The purpose is to obtain an output that directs a reviewer back to the relevant clauses. It is not to get the AI to declare the agreement satisfactory. Check the quoted text against the actual documents as well.

For a liability clause, a service provider could request the following decision materials.

Review this liability clause from the service provider's perspective.

- Separate the conditions for liability, recoverable losses, liability cap,
  and exceptions to the cap.
- Distinguish facts stated in the agreement from your assumptions.
- Explain the implications for data leaks, infringement of third-party
  rights, and delayed delivery.
- List missing transaction information and possible sources for checking it.
- For any proposed amendment, explain its reason and effect on the other party.

A person will decide which proposal to adopt. Give sources for statutes
and cases you cite, and state when you have not verified their text.

[Clause and related contractual conditions approved for use]

Specific instructions do not guarantee accuracy. Check consistency with the allocation of responsibility throughout the agreement, including exclusions and indemnities elsewhere. A well-structured response and a legally sound response require separate checks.

Common Failures in Business Use

One concern is sending an AI-generated amendment without checking the negotiation context. For example, AI may seek to delete a condition as a general risk even though the business team accepted it to secure a delivery date. Legal staff need to establish who accepted what and, where necessary, refer the issue for a further internal decision.

Another failure is relying on a nonexistent clause or an outdated version. Similar filenames, reference templates and signed agreements can make the source of an answer unclear. Fix the version being reviewed, distinguish it from reference materials, and verify where each proposed change belongs in the target document.

At the other extreme, returning all extraction and comparison work to manual handling limits the opportunities to use AI. I believe a practical approach is to begin with tasks that are easy to check, such as extracting contract terms, then adjust the scope according to recurring errors and the time required for verification.

Checks for Day-to-Day Operation

Before starting, check the service and plan, use of data for training, retention periods, access permissions, connections to external services and internal rules. A “business” label does not establish permission to enter every kind of confidential or personal information. Determine the permitted scope in light of service terms and client agreements, and remove unnecessary information.

Removing names may still leave someone identifiable through a combination of company names, case history and amounts. Take care when assuming information has been anonymized. If the environment cannot accept the information needed for a decision, use hypothetical facts to explore general issues and separately compare them with the actual matter.

Build the following checks into the workflow:

• Before input: confirm the version, scope of materials and information permitted.
• After output: check source text, figures, deadlines, exceptions and cited laws and cases.
• Before adoption: confirm commercial assumptions and internal approval authority.
• Before external communication: check amendments, comment wording, recipients and attachments.

Even when a search result includes a source link, read the source to see whether it supports the conclusion. If the original material cannot be obtained, leave that point unverified and choose another means of checking it. When AI can operate external services, distinguish authority to draft from authority to send or register information.

The appropriate scope of AI use depends on more than model capability. It also depends on the company materials that can be supplied, the availability of qualified reviewers, and whether approval records can be retained. I believe defining these conditions for specific workflows supports sustained use in practice.