LEGALISS's LLM Strategy 2026:
Towards a "reliable AI" that understands Japanese knowledge and business practices and drives business operations.

1. Japan Needs More Than AI That Speaks Japanese

Generating Japanese text is not the same as understanding Japanese society.

In Japanese workplaces, language varies across organizations and industries, from contracts, internal regulations and approval requests to reports, product specifications, customer service records, and medical and nursing care documents. To earn trust in real work, AI must understand omitted subjects, honorific language, implicit assumptions, long contexts, and company-specific abbreviations.

Japanese companies also place particular importance on the following requirements when adopting AI:

• Preventing the external disclosure of personal and confidential information

• Making the evidence and sources behind answers verifiable

• Detecting incorrect answers and inappropriate actions

• Enabling human approval and intervention to stop operations

• Continuously evaluating quality, cost, and safety after deployment

• Avoiding excessive dependence on a single model or cloud provider

The Japanese government is advancing both the adoption and governance of generative AI. In fiscal 2026, the Digital Agency is conducting trials of its government AI, Gennai, for approximately 180,000 staff across ministries and agencies, while promoting the evaluation and procurement of domestic LLMs. AI has moved beyond experimentation into a phase where its reliability as social infrastructure is being tested.

​

2. LEGALISS Builds AI Systems, Not Just Models

Introducing a single capable LLM does not automatically solve a company’s challenges.

Real business operations require integration of enterprise data, search infrastructure, access permissions, external tools, workflows, evaluation systems, and human approval alongside the LLM.

Spanning LLMs and SLMs, SaaS AI platforms, AI applications, IoT, and robotics, LEGALISS designs complete systems that enable AI to operate safely within business processes.

​

A Portfolio Approach to Models

The largest model is not necessary for every task.

We use high-performance LLMs for advanced reasoning and lightweight SLMs for routine processing and real-time responses. For highly confidential work, we select models that run on premises or in dedicated enterprise environments. Model selection balances accuracy, speed, cost, and security.

Our goal is a sustainable AI architecture with interchangeable models, so model updates do not require rebuilding the entire business system.

​

3. Design Japanese Data for Value, Not Volume

The quality of an LLM is not determined by data volume alone.

Accuracy, domain expertise, usage rights, freshness, and intended purpose all matter. To use Japanese-language and company-specific data safely, LEGALISS emphasizes:

• Managing data provenance, rights, and terms of use

• Detecting duplicate, low-quality, and personal information

• Structuring industry terminology, specialist documents, and company-specific expressions

• Preventing contamination between evaluation and training data

• Building a knowledge infrastructure that keeps pace with information updates

Not all proprietary knowledge needs to be incorporated through model retraining. We combine retrieval-augmented generation (RAG), which retrieves relevant information and provides answers with source references, with additional training tailored to specific use cases.

Research into open models with strong Japanese capabilities is also advancing. LLM-jp, led by the National Institute of Informatics, has released models trained on Japanese corpora and mixture-of-experts (MoE) models. LLM-jp

LEGALISS combines research from Japan and abroad with company-specific knowledge to design AI suited to Japanese workplaces.

​

4. From AI That Reasons to AI That Executes Work

Since 2025, reasoning has become one of the defining advances in LLM development.

Reinforcement learning and inference-time computation have enabled models to analyze problems step by step and evaluate multiple candidate solutions. Research on DeepSeek-R1 demonstrated that large-scale reinforcement learning can elicit reasoning capabilities, while also highlighting challenges such as readability and language mixing that reasoning alone does not resolve.

The next step is the AI agent, connecting these reasoning capabilities to business operations.

AI agents do more than answer questions. They plan around objectives, search internal information, operate APIs and software, and carry out work while checking the results.

Rather than granting agents unlimited autonomy, LEGALISS designs permissions according to operational risk:

• Allowing read-only access

• Requiring human approval before actions

• Restricting the tools and data an agent can use

• Recording execution history and decision outcomes

• Stopping operations and handing control to a person when anomalies occur

We prioritize dependable operation within authorized boundaries over the ability to do everything.

​

5. Expanding into Multimodal and Physical AI

Businesses work with far more than text.

Multimodal AI integrates and understands images, audio, video, tables, drawings, and sensor data. It is becoming increasingly important across manufacturing, retail, healthcare, tourism, entertainment, and many other industries.

Alongside language model development, LEGALISS works in AI applications, virtual humans, IoT, and robotics.

Drawing on this broad technical foundation, we aim to develop AI that not only processes digital information but also perceives the physical world, interacts with people, and translates understanding into appropriate actions.

Applications connecting LLMs to society include AI that understands store information and supports staff, explains warning signs in equipment data, or guides users through an understanding of speech and video.

​

6. Evaluate Business Outcomes, Not Just Benchmark Scores

High scores on general benchmarks do not necessarily mean a model will perform correctly in real business operations.

LEGALISS designs evaluation criteria for each organization and use case.

​

Quality Evaluation

• Natural Japanese and accurate understanding of instructions

• Understanding technical terms, proper names, and internal expressions

• Accuracy and consistency of answers

• Faithfulness to reference information

• The ability to withhold an answer when uncertain rather than guess

​

System Evaluation

• Response latency

• Inference cost per request

• Stability under concurrent use

• Changes in quality following model updates

• Recovery from failures

​

Safety Evaluation

• Resistance to prompt injection

• Prevention of confidential and personal information leaks

• Prevention of unauthorized actions

• Measures against harmful and discriminatory outputs

• Action logs and auditability

It is essential to evaluate the entire system, including retrieval, tool use, access control, and human oversight, rather than the model alone.

​

7. Build AI Governance into Development from the Start

AI governance is not about adding rules after a system has been completed.

Intended uses and prohibited activities must be defined during planning, risks assessed during development, and logs, incidents, and changes in quality continuously monitored after deployment.

LEGALISS incorporates governance into system design, taking Japanese institutions and guidelines into account.

The AI Guidelines for Business issued by the Ministry of Economy, Trade and Industry and the Ministry of Internal Affairs and Communications describe the responsibilities of AI developers, providers, and users, as well as risk management throughout the lifecycle.

In July 2026, the AI Safety Institute also published version 1.20 of its Guide to Evaluation Perspectives on AI Safety.

LEGALISS aims to create an environment where appropriate management allows companies to confidently unlock AI’s capabilities, rather than letting rules restrict its potential.

​

8. LEGALISS’s Principles for LLM Development

Six principles guide our LLM development:

1. Japanese Intelligence
   AI that deeply understands Japanese language, culture, and business knowledge.

2. Model Independence
   An infrastructure that optimally combines LLMs and SLMs without dependence on a specific model.

3. Enterprise Knowledge
   Mechanisms that securely connect enterprise data and produce evidence-based answers.

4. Agentic Execution
   AI that integrates search, decisions, and tool use to support work from start to finish.

5. Safety by Design
   Design that incorporates permission management, auditing, and human approval from the development stage.

6. Continuous Evaluation
   A framework for continuously measuring accuracy, cost, and safety and improving systems in operation.

​

Trusted AI to Drive Japan’s Future

Japan needs more than simply importing AI from overseas or competing to build the largest models.

It needs AI that understands the Japanese language, Japanese industries, and the quality standards of Japanese companies, protects critical data, and moves real work forward.

LEGALISS treats LLMs, SLMs, SaaS, AI applications, IoT, and robotics as one technical foundation, connecting model development with business implementation, evaluation, and operation.

We are building more than AI that appears intelligent in a demonstration.

Our AI is designed to operate responsibly within companies and society, support human judgment, unlock human creativity, and carry Japan’s industrial competitiveness into the next era.

Understanding Japan. From Japan to the world. LEGALISS is creating a new standard for trusted AI.