How Does AI Make Decisions in a World Without Data? — The Potential of Decision Systems in Medical Triage —

1. The Question

Can AI do anything without data?

Most AI systems today assume one thing:

👉 Data comes first

But in many real-world situations—especially in healthcare—this assumption does not hold.

• Patients arrive with incomplete information
• Symptoms are noisy and ambiguous
• Ground truth is unknown at the moment of decision

And yet, decisions must be made.

So the real question becomes:

👉 How can AI make decisions when data does not yet exist?

---

2. The Limitation of Conventional AI

Traditional AI follows this structure:

Data → Training → Inference

This leads to several constraints:

• AI cannot function without sufficient data
• It depends on labeled datasets
• It fails in cold start situations

In other words:

👉 No data = No decision

But in domains like medical triage, this is unacceptable.

You cannot wait for data.
You must act under uncertainty.

---

3. A Shift in Perspective

The core assumption must change.

AI is not fundamentally a prediction system.

👉 AI is a Decision System

Instead of asking:

• “Do we have enough data to predict?”

We ask:

• “How can we structure decisions under uncertainty?”

This shift changes everything.

---

4. The Approach

To address this problem, we combine three components:

• Simulation
• Decision Trace Model
• Multi-Agent

These are not independent techniques.
They form a connected process that enables decision-making without prior data.

---

4.1 Why These Three Must Be Combined

The challenge is not just making decisions.

It is:

👉 How to start, observe, and improve decisions without data

---

Step 1: Simulation — Creating the Initial Form

When no data exists, we cannot begin with learning.

Instead, we create a hypothetical environment:

• Virtual patients
• Defined symptom patterns
• Assumed progression models

The purpose of simulation is not accuracy.

👉 It is to make decisions run for the first time

Simulation creates:

👉 The initial structure of decision-making

---

Step 2: Real-World Operation — Introducing Reality

Once deployed:

• Real patients arrive
• AI makes decisions
• Doctors intervene
• Outcomes are observed

Now we obtain something new:

👉 Real logs

---

Step 3: Decision Trace Model — Identifying the Gap

Simulation-based decisions will always differ from reality.

The key is not avoiding the gap, but understanding it.

Decision Trace Model decomposes decisions into:

• Event
• Signal
• Decision
• Boundary
• Human
• Log

This allows us to see:

👉 Where did the decision go wrong?
👉 Which assumption failed?

---

Step 4: Multi-Agent — Enabling Targeted Improvement

Real-world decisions are not singular.

They involve multiple perspectives:

• Symptom interpretation
• Risk assessment
• Prioritization

By distributing these into agents:

👉 We can isolate which part caused the deviation

Instead of retraining everything, we can:

👉 Fix specific components of the decision structure

---

■ What Emerges from This Combination

When these three are integrated, a new loop appears:

1. Simulation creates the initial form of decision-making
2. Real-world execution generates logs
3. Decision Trace identifies gaps
4. Multi-Agent enables targeted correction

And this leads to:

👉

Shape decisions through simulation,
detect distortions through real data,
and refine the decision structure based on those distortions

---

■ Why This Matters

The key is this:

We do not assume data exists at the beginning.

What we need is not:

• Perfect models
• Large datasets

But:

👉 A structure that can start decisions, record them, and improve them

Only then can meaningful data emerge.

---

5. Use Case: Medical Triage

In emergency departments, decisions must be made immediately:

• Which patient should be prioritized?
• How urgent is the condition?
• What test should be performed first?

The critical point is:

👉 There is no correct answer at the moment of decision

---

■ Decision Trace Flow

---

■ Role of Simulation

• Test triage decisions with virtual patients
• Record decision outcomes
• Compare with real-world cases

---

■ What Happens in Improvement

Example:

• AI predicts: Medium risk
• Actual outcome: Critical condition

👉 There is a gap

---

Instead of retraining blindly, we analyze:

• Was the signal interpretation wrong?
• Was the risk threshold too low?
• Should escalation have happened earlier?

Then we adjust:

• Rules
• Boundaries
• Agent responsibilities

---

👉 Improvement is not about model retraining
👉 It is about refining the decision structure

---

6. The Essence

The key insight is:

👉 Data is not the starting point

Traditional approach:

Data → Learning → Inference

This approach:

Decision → Execution → Log → Data

---

Data is not something we begin with.

👉 It is something that emerges from decisions

---

More precisely:

• Decisions generate actions
• Actions produce outcomes
• Outcomes are recorded as logs
• Logs accumulate into data

---

👉 Data is the byproduct of structured decision-making

---

7. Conclusion

In uncertain domains like healthcare:

• Inputs are unstable
• Ground truth is delayed
• Data does not exist upfront

And yet, decisions must be made.

---

That is why we need:

• Simulation
• Decision Trace Model
• Multi-Agent

---

These are not tools for prediction.

They are components of a decision-making system.

---

■ Final Insight

AI does not run on data.

👉 AI runs on decisions—and data emerges from them.