• 

Salesforce Hackathon Project

My Name is Thomas DERASSE, France based, Salesforce Architect since a year, I am participating in a Salesforce Hackathon with a project based on Salesforce Agentforce, aiming to demonstrate a scalable, modular chatbot architecture tailored to real-world business needs, particularly in banking, marketing, and legal compliance sectors.

Architecture and Agent Design

The solution is structured around three specialized agents, with transfert capabilities, each operating within a well-defined perimeter:

1. General Banking Agent

Handles general banking inquiries related to accounts, transactions, and standard support services. Capability to send detailed transactions via PDFs by customer default email.

1. Marketing Agent (Pricing and Offers)

Manages marketing-related queries, specifically pricing, promotional offers, and brand campaigns. Based on catogorized Knowledge source topics

1. Legal Agent (Terms and Conditions)

Covers legal aspects such as terms of service, user rights, and data protection policies. Based on catogorized Knowledge source topics

Each agent operates independently within its domain and is restricted to knowledge articles strictly relevant to its responsibilities.

= Data Access and Knowledge Management

The agents' capabilities are built using the Salesforce Data Cloud Knowledge - Agentforce Library:

Knowledge articles are categorized by data type and business domain. Each agent has access only to the knowledge articles relevant to its scope. This strict segregation ensures:

Data compliance and security: no cross-domain data exposure. Operational clarity: each agent provides only domain-specific responses. Minimization of legal risks and support for auditability and GDPR compliance. Thus, agents cannot access information outside their assigned knowledge scope, maintaining a strong data governance model.

= Request Routing and Agent Transfer Logic

The conversation routing between agents respects a structured validation process to ensure a compliant and respectful user journey:

Initial Conversation Analysis:

The system analyzes the user’s initial message (intent detection via keywords or categorization). If the query falls within the current agent’s scope, it is handled directly.

= When Transfer is Needed

If the user’s request does not match the current agent’s domain, the agent politely proposes a transfer. No automatic conversation or data transfer happens between agents. Instead, the system asks for explicit user consent before proceeding:

"Your question concerns legal matters. Would you like me to connect you with our Legal Support team?"

= After User Validation:

Upon user agreement, the conversation is restarted with the newly assigned specialized agent. Only minimal session information (such as the user ID and opt-in confirmation) is passed to ensure continuity. Conversation context (previous messages) is not forcefully pushed to the next agent to respect privacy and limit unnecessary data exposure.

= This transfer logic maintains:

User control over their interaction journey. Data minimization principles. Compliance with consent-driven interactions.

= Hosting and Integration Architecture

The chatbot is deployed externally on a Google Site using a custom HTML/JavaScript integration while relying on the Salesforce Agent API backend.

= Compliance with Consent and Data Protection Rules:

Explicit consent is requested from users before agent transfer. Only the minimal necessary information (user validation for transfer) is preserved across agent hops. Sensitive data remains confidential and protected through Salesforce security policies. Data are categorized

= Knowledge Base Governance:

Knowledge articles are domain-segregated to avoid unnecessary exposure of unrelated data. Knowledge content updates are traceable and auditable.

= User-Centric Design:

Clear messaging throughout the experience. Smooth transitions between agents without user frustration. Contextually relevant answers provided rapidly through specialized agents.

= Monitoring and Auditability:

Logs capture conversation flow, agent transfer validations, and user decisions. Monitoring dashboards can be established using Salesforce reporting.

==========================

= = MVP Scope for Hackathon = =

The initial MVP delivers:

3 specialized agents fully operational (Banking, Marketing, Legal). User-driven dynamic transfer between agents. Deployment of the chatbot on a Google Site, fully integrated with Salesforce backend. Respect of GDPR and security compliance through controlled knowledge access and consent-driven transfers.

= Planned post-MVP improvements include:

Expansion of available knowledge bases per agent. Enhancement of intent detection models for better routing. Analytics integration for user behavior and chatbot performance tracking.

Built With

• agentforce
• apex
• datacould
• email
• experiencecloud
• flow
• gmail
• javascript
• knowledge
• prompt
• visualforce