Lightweight DTM for Building “Decision-Capable AI” — Can Customer Support Truly Be Automated?

■ Introduction

When I talk about the Decision Trace Model (DTM), I often hear:

“That kind of rich design is not feasible in real-world operations.”

In such cases, I respond:

“DTM can actually be started in a much lighter way.”

DTM does not require a full setup from the beginning (such as Ontology, Multi-Agent, Behavior Trees, etc.).

In practice, it has the following characteristic:

👉 It can start from a minimal configuration (Light DTM) and be expanded step by step.

---

In this article, I will explain how Light DTM can improve AI-driven customer support.

Today, many companies have introduced AI into customer support in the following ways:

• Automatic classification using LLMs
• Automated FAQ responses
• Chatbots for first-line support

At first glance, it appears that AI has already “automated customer support.”

However, in real-world operations, the following problems still occur:

• Misclassification happens
• Complaints and critical cases are overlooked
• Responses vary depending on the person in charge
• It is unclear why a particular decision was made

These may seem like separate issues, but they share a common root cause:

👉 AI is producing outputs, but it is not making decisions.

---

■ Signal ≠ Decision

This is a critical point.

What machine learning models and LLMs produce are not decisions.

For example:

• “This is a complaint”
• “There is an intention to cancel”
• “Customer dissatisfaction is high”

These are all:

👉 Signals

However, in reality, we need decisions such as:

• Complaint → escalate to a human or handle automatically?
• Cancellation intent → prevent or accept?
• Fraud risk → escalate or proceed?

This is:

👉 Decision

In other words:

👉 There is a fundamental gap between Signal and Decision.

---

■ Light Decision Trace Model (Minimal Structure)

Light DTM fills this gap with a minimal structure.

The configuration is very simple:

That’s all it takes.

What we are essentially doing is:

👉 separating Signal and Decision

---

In Light DTM:

• The model remains unchanged
• No need to build a new AI
• Only simple decision rules are added

In other words:

👉 It can be introduced without major system changes

That is why it is “lightweight.”

---

■ Why does it work despite being lightweight?

So why does this simple approach work?

Because:

👉 It extracts “decision” as an explicit structure

---

Previously:

Signal (classification / score / generation) = directly used

This resulted in:

• No explanation for decisions
• Inconsistent behavior across operators
• No systematic improvement

---

With Light DTM:

This enables:

• Clear decision criteria
• Consistent outcomes under the same conditions
• Continuous improvement through rule updates

---

In short:

👉 From “AI that outputs”
👉 To “systems that can decide”

---

■ Horizontal Expansion of Light DTM

This structure is not limited to customer support.

The key point is:

👉 The minimal structure
Signal → Decision → Human → Log
can be applied across many domains.

No need for new AI systems.

👉 Just add decision structures to existing data and models.

---

① Fraud Detection

Signal: risk score / anomaly detection
Decision: block / allow / escalate

Problems in traditional systems:

• Too many false positives due to threshold-based rules
• Difficult trade-offs between detection and misses

With Light DTM:

• Explicit decision rules and escalation logic
• Adjustable balance between false positives and misses
• Explainable decisions

👉 Improves stability in finance and e-commerce

---

② Approval Workflow

Signal: LLM-based summarization / evaluation
Decision: approve / reject / escalate

Traditional problems:

• Decisions vary by person
• Past decisions cannot be reused

With Light DTM:

• Decision criteria are externalized
• Decisions become consistent
• Decision history becomes reusable

👉 From subjective judgment to reproducible decision-making

---

③ Store-Level Decisions

Signal: customer data / purchase history / LLM suggestions
Decision: discount / recommend / hold

Traditional problems:

• Reliance on individual experience
• Inconsistent decisions

With Light DTM:

• Decisions are structured and logged
• Variability is reduced
• Successful patterns can be reused

👉 Turning operational decisions into organizational assets

---

■ What is common across all cases

The key takeaway is:

👉 This is not about improving AI accuracy
👉 It is about structuring decisions

And importantly:

👉 This can be achieved even with a minimal Light DTM setup

---

■ From Light DTM to Multi-Agent

This structure can naturally evolve into a more advanced system.

Even though:

is sufficient at the beginning,

real-world decisions involve multiple perspectives.

For example:

• Risk (Risk Agent)
  → fraud, failure, complaint risks
• User Experience (UX Agent)
  → customer satisfaction
• Business Impact (Business Agent)
  → cost and revenue

---

Instead of merging everything into a single model:

👉 Separate them by role

This creates:

---

This enables:

👉 Decision-making based on multiple explicit perspectives

---

Key benefits:

• Decision rationale is decomposed
• Influence of each perspective is visible
• Rules and weights can be adjusted later

---

In short:

👉 From black-box decisions
👉 To controllable and tunable decisions

---

This is:

👉 Distributed decision-making through Multi-Agent systems

The goal is not to increase the number of agents,

but:

👉 To decompose decision-making into perspectives

---

This entire structure can be implemented using:

👉 Multi-Agent Orchestrator Core

as described in:

“Decision System Execution Layer — What Multi-Agent Orchestrator Core Enables”

---

■ Conclusion

AI can already produce outputs.

However, what is required in real-world operations is:

👉 deciding what to do

---

Light Decision Trace Model provides:

👉 a minimal way to introduce decision-making

---

Even with the same system, the outcome changes significantly depending on how much structure is introduced.

• The goal is not to build a full system from the start
• Start light, then expand as needed

---

And from there:

👉 naturally evolve into advanced decision-making through Multi-Agent systems

---