Mohamed Imran

🛡️ Agentforce AI Trust & Compliance Audit System

Role: Senior Salesforce Administrator & Consultant Tools: Agentforce (Einstein Copilot), Einstein Trust Layer, Salesforce Flow, Custom Objects

Cover Image

📌 Project Objective

Architect an automated governance framework to capture, audit, and adjudicate AI-generated interactions in real-time. This system ensures Enterprise AI is safe, auditable, and compliant with strict regulatory requirements.

🎥 Project Demos

🚫 1. Business Problem & Solution

The Problem: Enterprises deploying Generative AI face strict regulatory requirements to monitor LLM outputs for toxicity, bias, and data leakage. Standard logging often lacks the granularity required for legal auditing or immediate remediation.

The Solution: I built a custom “Human-in-the-Loop” architecture. This system captures the raw prompt and model response from Agentforce, commits them to a permanent audit log, and executes a real-time safety evaluation using Salesforce Flow.

📂 2. Data Model Configuration

I created a custom object AI_Interaction_Audit__c to serve as the immutable ledger for all AI conversations.

Field Label API Name Data Type Description
Prompt Context Prompt_Context__c Long Text Stores the raw user input sent to the LLM.
Model Response Model_Response__c Long Text Stores the generated response from the LLM.
Toxicity Score Toxicity_Score__c Number (3,2) Ingests the safety score from the Einstein Trust Layer.
Action Taken Action_Taken__c Picklist Values: New, Automated Approval, Flagged for Review.
Timestamp Interaction_Timestamp__c DateTime Captures exact execution time.

🌉 3. Integration Layer: Autolaunched Flow (The Bridge)

Flow Label: Subflow: Log AI Interaction Audit Type: Autolaunched Flow (No Trigger)

Configuration Rationale: I selected an Autolaunched Flow to act as the synchronous API between Agentforce and the Salesforce database. This ensures the Agent waits for a successful write operation before confirming to the user.

Logic Design:

  1. Start: Invoked by Agentforce Action.
  2. Inputs: Accepts varPromptContext (User utterance) and varModelResponse (LLM generation).
  3. Create Records: Creates one AI_Interaction_Audit__c record.
  4. End: Returns success status to the Agent.

Autolaunched Flow Diagram

🧠 4. Agentforce Configuration (The Brain)

Topic Name: Compliance Logging Scope: Defined a strict scope to prevent hallucination. “My job is only to log every interaction for compliance…”

Instructions:

  1. Log Interaction: Explicit command to execute the Log AI Compliance Audit action.
  2. Input Mapping: Dynamic mapping of conversational context to the Flow variables defined in Section 3.

Agentforce Debugger

⚙️ 5. Automation Layer: Record-Triggered Flow (The Engine)

Flow Label: AI Interaction Audit - Safety Check Trigger: After-Save (Actions & Related Records) on AI_Interaction_Audit__c

Logic Design:

  1. Decision (Check Safety Scores): Evaluates Toxicity_Score__c and Masking flags.
  2. Path 1 (High Risk): If Score > 80 OR Masking = True:
    • Update Action_Taken__c to “Flagged for Review”.
    • Action: Send Custom Notification to the Compliance Team queue.
  3. Path 2 (Compliant): Default outcome.
    • Update Action_Taken__c to “Automated Approval”.

Record-Triggered Flow Diagram

✅ 6. Validation & Project Achievements

Testing Scenario: User requests “Check the compliance logs for today.” Result:

Audit Record Result

🏆 Key Outcomes

Return to Home