On May 28, I will give a talk at the 178th Square Free Seminar on the theme:
Designing AI Systems with Embedded Decision-Making
— Structuring Human Thought and Judgment into AI Systems —
178th Square Free Seminar
“The Frontline of Trust and Systems Built with AI”
Theme 2
Designing AI Systems with Embedded Decision-Making
— Structuring Human Thought and Judgment into AI Systems —
Date: Thursday, May 28, 2026
Time: 19:00–20:30 JST
Participation fee: Free
Format: Zoom / YouTube Live
Registration:
Square Free Seminar registration page
Details:
Square Free Seminar information page
With the rapid evolution of generative AI and AI agents,
AI is no longer just a chat tool.
Today, AI is beginning to:
- write code
- conduct research
- reason
- operate as an agent
- interact with external systems
In other words, AI is shifting from being:
a system that generates information
to becoming:
a system that executes actions.
But this shift raises a very important question:
How far can we trust AI?
This is not simply a question of AI accuracy.
In the real world, we must deal with:
- exceptions
- responsibility
- safety
- organizational judgment
- past-success bias
- fragmented information across departments
- weak signals
- human approval
Real-world decision-making cannot be built on inference alone.
AI Is Moving from Prediction to Decision-Making
Many AI systems are still mainly designed around:
- searching
- generating
- answering
But in real operational environments, what matters is:
- how decisions are made
- where the system should stop
- when it should return control to humans
- who takes responsibility
- why a particular decision was made
This is where the following concepts become important:
- Runtime
- Boundary
- Human Gate
- Decision Trace
I organize these ideas as:
Decision Trace Model (DTM)
DTM Designs Trust
The key point of DTM is not to make AI fully autonomous.
Rather, it is the opposite.
The important question is:
How far should we trust AI?
DTM is a way to design that trust.
In DTM, decision-making is structured and recorded through the following flow:
Event
↓
Signal
↓
Runtime
↓
Flow
↓
Boundary
↓
Human Gate
↓
Execution
↓
Ledger
This makes it possible to trace:
- AI judgments
- human intervention
- boundary conditions
- reasons for execution
- decision history
In other words, DTM is not simply a mechanism for recording AI outputs.
It is a mechanism for structuring the decision-making process itself in order to support trust.
Weak Signals Matter in Manufacturing
In this talk, I will use a manufacturing example to discuss how to handle:
- small anomalies
- fragmented information across departments
- past-success bias
- lack of re-evaluation
- boundary violations
- Human Gate
In real operational environments, teams such as:
- QA
- Field Team
- Fast Charging Team
may each hold different pieces of information.
A signal that does not appear serious on its own may become a major risk when viewed as part of the whole.
However, in conventional systems:
- information is often fragmented
- reasons for decisions are not recorded
- issues may stop at “not a major incident”
- problems may only become visible later
DTM addresses this through:
- weak signal integration
- boundary management
- human escalation
- decision trace
This makes it possible to re-evaluate:
why a particular decision was made.
What the AI Era Needs Is Not the “Strongest AI”
What truly matters in the AI era is not only the performance of AI itself.
What matters is:
- how AI is connected to business operations
- how human judgment is embedded
- how boundaries are managed
- how decisions are traced
- how decision processes become reusable
AI is not prediction.
It is decision.
And decisions can become corporate assets.
Blog:
deus-ex-machina-ism.com
GitHub:
masao-watanabe-ai GitHub
AIシステム設計・意思決定構造の設計を専門としています。
Ontology・DSL・Behavior Treeによる判断の外部化、マルチエージェント構築に取り組んでいます。
Specialized in AI system design and decision-making architecture.
Focused on externalizing decision logic using Ontology, DSL, and Behavior Trees, and building multi-agent systems.
