The Difference Between “Madoless” and “General-Purpose Generative AI”: How Does Madoless, Specialized in Procedural Guidance, Control Hallucinations?

When you hear “AI agent specialized for procedures,” you might think, “Can’t general-purpose generative AI (general AI) also answer procedural inquiries?”.
However, the nature of general-purpose AI can become a weakness in the field of procedural guidance.
Today, I will explain the weaknesses of general-purpose AI in procedural guidance and the approach of Madoless to overcome them.

Challenges General-Purpose AI Faces in Procedural Guidance

General-purpose AI constructs sentences from a wide range of learned knowledge. If you ask a question, it will usually return some kind of answer. It is not good at saying “I don’t know” when it doesn’t know something, and plausible errors (hallucinations) can get mixed in.

For example, if you ask a general-purpose AI, “I’ve moved, so please tell me the necessary procedures,” it will return an answer that organizes general procedures such as moving-in notifications and national health insurance.
However, the AI does not know the document formats, reception counters, or unique systems used by that specific local government. Similarly, for financial institutions, the documents required for insurance claims may differ depending on the product or policy terms.

General-purpose AI may sometimes guess parts it doesn’t know from general information and write them in a plausible way. Users may end up heading to the counter without being able to distinguish for themselves whether the answer is correct.

Procedural guidance is a field that is incompatible with this nature. Required documents, submission destinations, and deadlines differ by local government, financial institution, and product, and systems are subject to change. It is entirely possible that an answer supplemented by general knowledge will conflict with the actual rules of that organization.

Another point is that not all procedures can be determined by a system. Depending on individual circumstances, the judgment of a staff member may be required.
Telling the user when something is unknown and connecting them to a person. This action determines the quality of the guidance.

Madoless’s Approach: Enabling AI Answers to be Provided as Organizational Guidance

Madoless is built on the premise that AI answers will reach users directly as organizational guidance.
Therefore, we have designed it to include not only what the AI answers, but also what it does not answer, and who is responsible for managing it.

Answering Only Within the Scope Determined by the Organization

Madoless answers based only on materials registered and approved by the organization. For questions where no basis is found, it does not supplement with the AI’s general knowledge, but instead replies “I cannot answer” and guides the user to a counter or phone. This is because we believe that in procedural guidance, the decision not to provide an answer is just as important as providing one. Personal information handled at the consultation stage is also limited to the scope necessary to identify the procedure, and identity verification is performed on an authenticated application screen.

Providing Evidence and Connecting Unknowns to People

Madoless’s answers are accompanied by the name of the material that served as the basis. Users can verify the source of the guidance, and consultations that cannot be answered are handed over to staff.
Rather than having the AI complete everything, we pass consultations that should be handled by people to people. We consider this line to be indispensable in procedural guidance and enforce it thoroughly.

Ability to Continuously Manage Materials and Answers

Systems and formats are subject to change. With Madoless, the organization approves materials by version, and the AI refers only to the latest approved version. Before publication, it is tested with hypothetical questions, and changes to answers and settings are kept as a history. It is a mechanism that allows the organization to responsibly manage and update AI guidance.

Summary

For an AI that provides procedural guidance, as much as the ability to answer correctly, a design that limits the scope of answers, provides evidence, and connects unknowns to people is required.
Madoless enables the organization to manage its operation through a mechanism that bases answers only on registered and approved materials, pre-publication testing, and the recording of change history.

If you are a local government, government agency, or financial institution interested in Madoless, please feel free to contact us.

Contact us here: https://service.opro.net/madoless/contact