Classify Systems from Facts

Use TRACE to classify systems carefully, including cases where important evidence is missing.

A simple method

You now know the main terms. The next step is to use them.

Real product pages often leave out important details. They may show a smooth chat screen but hide the search, rules, tools, and people behind it.

We need a method that works even when some facts are missing.

We will use TRACE.

TRACE is a course tool. Each letter reminds us to ask one question.

T: Task

What work does the system do?

Use clear words. For example:

• Sort an email.
• Find a document.
• Write a summary.
• Calculate a value.
• Suggest an action.
• Carry out several actions.

R: Result

What does the system produce or do?

It may produce a score, answer, image, warning, or action.

Do not assume that a future-looking result uses machine learning. A spreadsheet can copy a value into future months with a simple formula.

A: Actual method

What method is supported by facts?

Look for rules, trained models, neural networks, generation, retrieval, tools, planning, or actions.

Do not guess the missing parts.

C: Controls

What controls the system?

Look for:

• Access limits.
• Approval before action.
• Review by a person.
• Logs.
• Stop rules.
• Clear limits on what the system may do.

A control is useful only if it works in practice.

E: Evidence gap

What important fact is missing?

Write it clearly.

For example:

We know that the system extracts text. We do not know if it uses fixed templates or a trained model.

Example 1: a forecast dashboard

Description:

A spreadsheet copies the latest monthly amount into the next three months. The dashboard calls it an AI revenue forecast.

Task: show a future value.

Result: three future monthly values.

Actual method: a copying formula.

Controls: users can inspect the formula and change the starting value.

Evidence gap: there is no proof of a trained model.

Best description: fixed-formula automation. The AI claim is not supported.

Example 2: a policy-answer service

Description:

The service finds passages in approved policy documents. It gives the passages and a question to a language model. The model writes a draft. A policy officer must approve it.

Task: help answer a policy question.

Result: found passages and a draft answer.

Actual method: retrieval and text generation.

Controls: approved documents and officer approval.

Evidence gap: we do not know how well the search and model were tested.

Best description: a system that combines retrieval, generative AI, and human approval.

Your turn

A vendor says its DeepScan Agent extracts text from forms. It gives no details about training, tools, actions, or tests.

We know that it extracts text. We do not know if it uses deep learning. We also do not know if it takes actions like an agent.

The best answer is:

Text extraction is clear. The method and agent claim are not proven.

Weak answers to avoid

• It is AI because it is clever.
• It handles pictures, so it must use deep learning.
• Its name says Agent, so it is an agent.
• It uses rules, so it cannot be AI.
• A person reviews it, so it must be safe.

Each answer makes a claim without enough facts.

Remember

• Use TRACE: Task, Result, Actual method, Controls, Evidence gap.
• Keep the result separate from the method.
• A product name is not proof.
• Do not be afraid to say that facts are missing.

Next, we will use TRACE on a complete system with many parts.