How to Make AI Legal Assistants Affordable for Everyone

    This article explains how AI legal assistants can make legal information easier to reach and cheaper to use. It covers why general-purpose large language models fall short in legal work, how retrieval, rule validation, and human review fit together in a reliable system, and what cost and boundary considerations matter for individuals and small businesses.

    In legal services, the hardest problems are not only whether there are enough lawyers. They also include whether ordinary people know when to seek help, whether they can get a preliminary assessment quickly, and whether service prices match what people can actually pay.

    China’s practicing lawyer population keeps growing, but compared with its large population and number of businesses, legal services still face a clear gap between supply and demand. Many individuals and small and medium-sized enterprises face contract disputes, labor disputes, debt collection, or intellectual property issues. They often do not need less legal help. They do not know where to start, and they struggle to afford the consulting fees of traditional services.

    Making legal assistants affordable therefore does not simply mean handing lawyers’ work to AI. It means using technology to restructure the entry point, workflow, and cost structure of legal services.

    Traditional legal services mainly rely on in-person consultations, case representation, and specialized advisory work. This model is irreplaceable for complex cases, but it has several inherent limits.

    First, service supply depends heavily on professionals’ time. Lawyers must read materials, confirm facts, research laws, assess risks, and then produce opinions with legal effect or professional value. When the number of people seeking help grows quickly, relying on human labor alone makes it hard to cover the large volume of low-frequency, scattered, and small-value needs.

    Second, users often cannot describe their problems precisely. Many people only say “the other party breached the contract,” “the company wants to fire me,” or “I can’t get back money I lent out.” They do not know which facts matter or which evidence is valid. Traditional consultations often require clarifying the problem before legal analysis can begin.

    Third, there is a clear information asymmetry. Professionals know the legal rules and procedures, while ordinary users usually lack basic legal knowledge. They struggle not only to judge their own rights but also to compare the costs and risks of different solutions.

    AI can handle the standardized parts of this process, for example:

    • collecting case facts through conversation;
    • classifying natural-language questions into relevant legal scenarios;
    • retrieving legal basis from laws, judicial interpretations, and publicly available court decisions;
    • helping generate contract risk checklists;
    • helping users organize evidence and timelines;
    • providing a preliminary explanation of possible paths to resolution.

    These capabilities do not replace lawyers. They free lawyers from repetitive work and lower the barrier for users to get initial legal information.

    General-purpose LLMs are good at understanding language, summarizing text, and generating content. Legal services, however, demand higher accuracy, clear sources, and up-to-date information. A fluent answer built on incorrect legal grounds can mislead users into harmful actions and cause real losses.

    Legal AI faces four core problems.

    Laws are amended, judicial interpretations change, and local rules and industry regulations may differ. If model training data is not updated in time, the model may cite provisions that are no longer valid or ignore their scope of application.

    Take “terminating a contract” as an example. Whether it constitutes a breach depends on payment status, notice method, performance deadlines, contract terms, resulting losses, and other factors. A model cannot reach a reliable conclusion from a single keyword.

    General Q&A can accept summary explanations, but legal answers need to show where their conclusions come from. A reliable system should, where possible, provide the name of the law, the location of the relevant provision, the applicable conditions, and any necessary limitations.

    4. The cost of error is far higher than in general content generation

    In marketing copy or routine office work, a mistake may only mean redoing the work. In legal matters, an error can affect limitation periods, evidence preservation, and the exercise of rights. The system must therefore prioritize verifiability over response speed.

    3. Domain Models and Retrieval-Augmented Generation Should Work Together

    A common mistake when building a legal system is to assume that training a larger model will automatically solve domain problems. In practice, model capability, a legal knowledge base, and business workflows are all necessary.

    A more robust system typically combines model understanding, knowledge retrieval, and rule validation.

    Model understanding layer

    This layer identifies user intent, extracts key facts, determines the question type, and converts natural language into structured information. For example, in a labor dispute, the system can extract the employment start date, contract status, salary components, reason for termination, notice dates, and existing evidence.

    The system needs a curated and reviewed knowledge base. Its content may include current laws and regulations, departmental rules, judicial interpretations, procedural guidance, standard document templates, and desensitized case materials.

    A knowledge base should not be a simple batch upload of files. It needs metadata for version, jurisdiction, legal effect, and applicability conditions, so that retrieval results can be filtered accurately.

    Retrieval-augmented layer

    When a user asks a question, the system first retrieves relevant content from the knowledge base, then provides the results to the model to generate an answer. This reduces the risk of the model fabricating legal provisions from memory and makes later auditing possible.

    A simplified workflow looks like this:

    1. User question
    2. ↓
    3. Intent recognition and fact extraction
    4. ↓
    5. Retrieve statutes, procedures, and cases
    6. ↓
    7. Filter by validity, jurisdiction, and date
    8. ↓
    9. Model generates a preliminary answer
    10. ↓
    11. Citation check and risk warning
    12. ↓
    13. Escalate to human professionals when necessary

    Rule validation layer

    Tasks suited to deterministic processing, such as limitation periods, amount calculations, date determinations, and document checklists, should not be left entirely to a generative model. Rule engines or dedicated programs should handle them. The model handles understanding, the program handles calculation, and humans handle high-risk judgments. This division of labor fits real needs better.

    4. Reducing Hallucination Requires a Complete Engineering System

    Model hallucination typically appears as invented legal provisions, incorrectly cited cases, confusion between jurisdictions, or overly certain conclusions when evidence is insufficient.

    Reducing hallucination cannot rely on a single system prompt. It requires multiple layers of protection.

    Data layer: control knowledge sources

    The knowledge base should record each document’s source, update date, scope of application, and review status. Outdated or disputed materials should be labeled so the system does not mix historical rules with current ones.

    Retrieval layer: improve recall and ranking quality

    Legal questions often use varied expressions, so the system cannot rely only on keyword matching. It should combine semantic retrieval with provision hierarchy, case type, and jurisdiction when ranking results. For critical questions, the system can run multiple retrieval routes and merge the results using rules.

    Generation layer: limit answers without sufficient basis

    When the system does not find enough basis, it should state that the current information is insufficient and ask for the facts it needs, rather than forcing a complete conclusion. Answers should distinguish between legal provisions, common handling practices, and preliminary judgments based on the current information.

    Verification layer: check citations in reverse

    After generating an answer, the system can check the law names, provision numbers, and case information in it. This confirms that the citations exist and that their content matches the claim. For amounts, dates, and percentages, the system should also verify the results by calculation.

    Interaction layer: expose uncertainty

    A reliable legal assistant should not present every answer as a definitive conclusion. When facts are incomplete, local rules may differ, or professional judgment is needed, the system should clearly