Chain of Thought (CoT) Datasheet

—

AI SURVEILLANCE SOLUTION

Chain of thought

Our LLM’s reasoning process

Capture, transcribe, and archive voice and digital communication data from leading voice capture platforms for seamless integration with multiple communication solutions.Contents

—
About Global Relay

Global Relay Surveillance

Regulators such as the SEC, FINRA, ESMA, and the FCA require firms to demonstrate a risk-based and explainable approach to communications monitoring. By not only detecting potential financial and non-financial misconduct—such as insider trading, market manipulation, conflicts of interest, corruption, and bribery—but also being able to clearly explain why content was flagged, organizations can act earlier, protect market integrity, and maintain an ethical working environment.

Global Relay Surveillance is a natively built, AI-enabled surveillance solution that is agentic by design. It helps organizations like yours manage risk in your communications data and meet regulatory demands. The solution uses a structured, step-by step reasoning process incorporated to the Large Language Model (LLM) prompts, called “Chain of thought” (CoT), to improve accuracy.

Before our LLM flags a potential risk in the data, the CoT process gives it time to reason through each step of the decision-making process and investigate the intent of the message for greater accuracy in risk assessment.

Chain of thought

When analyzing the message content, the LLM takes into account the full context behind it and uses this guided reasoning to decide whether the content represents a risk. It then provides an explanation for its decision. The CoT reasoning process is shown in the UI for additional transparency.

This CoT process allows a higher level of precision and recall, especially against complex, ambiguous, and context-based real-life communication.

How it works

Chain of thought is an integral part of our LLM’s decision-making process: before our LLM provides a decision alert, the CoT process takes place and ensures that the AI has gone through the entire context of the conversation, step by step.

The step-by-step process includes:

Context: It begins by summarizing the message while considering the surrounding conversation, establishing the foundation for analysis.
Potential indicators: Then it identifies any potential indicators of misconduct or regulatory concern.
Evaluation: Then it interprets meaning and aligns it with known patterns of potential risk.
Verification: The model then verifies whether there is concrete evidence or explicit language supporting a possible match.
Decision: Based on above steps, it determines the appropriate risk classification.

Alongside the decision alert, our LLM provides visibility in the UI to the exact step-by-step analytical process CoT went through in order to come to its conclusion. With this information, the alert reviewer can make an informed decision on the actions to take next.

Feature highlights

Our LLM does not require any training from the user
The CoT process leads to the LLM providing a decision
The whole conversation is analyzed, not just a sentence
Attempts at concealing illicit communication and conduct
(i.e. via slang) are flagged
The user gets a full breakdown of the reasoning process, not just a policy alert for potential illicit activity
Highly accurate results, with a significant reduction in false positives

Discover how Global Relay is leading innovation in AI surveillance

Key advantages

Improved accuracy and consistency
The CoT process helps LLMs perform better on tasks that require reasoning (i.e. common sense, symbolic manipulation, colloquial language), because the model reduces errors that can occur when jumping directly to an answer. Consequently, it reduces mistakes and consistently provides highly accurate results.

Increased transparency and traceability
CoT’s step-by-step output makes it easier to see how the model arrived at its answer, allowing users to verify logic and identify errors. This transparency is valuable for building trust in the AI’s output.

Better handling of complex tasks
CoT is particularly useful for tasks that require multi-step reasoning (such as understanding and following conversations of a complexed nature, or by uncovering hidden meanings behind words and/or phrases), where a direct answer may be difficult to calculate in a single step.

Mimics human reasoning
CoT aligns AI responses more closely with how humans approach problems, using a logical and sequential thought process that can be more intuitive for users to follow and evaluate.