Client and Business Context
The project was carried out for a large technology company in the adtech industry. The company needed to streamline and automate its customer service process, especially in responding to frequently recurring questions.
Challenge: Scalable 24/7 Customer Support
The main challenge was to provide 24/7 customer support directly within the company's product, using the existing HubSpot Chatflows. The key was to intelligently route inquiries - automatically answering standard questions about pricing, documentation, integrations, and churn, while efficiently redirecting more complex or specific cases to the live support team. The system had to be scalable and capable of processing information from the company's extensive knowledge base.
- Handling customer inquiries 24/7.
- Integration with the existing HubSpot Chatflows system.
- Automatic responses to questions about pricing, documentation, integrations, and churn.
- Intelligent routing of conversations to live support.
- Need to utilize an extensive internal knowledge base (documentation, FAQ).
- Ensuring data security and verification of personal data in inquiries.
- Monitoring performance and conversation topics.
General architecture of the solution using Vertex AI Agent Builder, Vertex AI Search, and external integrations.
Solution: AI Agent in Vertex AI Agent Builder with RAG
In response to the client's challenges, I designed and implemented an advanced conversational AI agent that serves as the first line of support. The solution is based on:
- Vertex AI Agent Builder: A Google Cloud platform for creating and managing AI agents. A Playbooks-based architecture was used, which allowed for precise definition of conversation flows, business logic, and tool integrations for various types of queries (e.g., separate Playbooks for pricing, documentation).
- Vertex AI Search: Used as a scalable RAG (Retrieval-Augmented Generation) system. This enabled the agent to use the client's extensive, unstructured knowledge base (documentation, FAQ) to generate precise and up-to-date answers.
- Make.com (Middleware): An automation platform used as a middleware layer to integrate the AI agent with HubSpot Chatflows and other client systems (e.g., CRM, notification systems). Make.com also handles the logic for retrieving current prices and logging conversations.
- Custom Functionalities:
- Personal Data Verification: A mechanism that identifies and blocks the processing of sensitive data in queries.
- Question Categorization and Routing: The main Playbook ("Customer Help Center") analyzes and categorizes the query, then directs it to the appropriate task-specific Playbook (e.g., "Pricing Center", "Documentation Center").
- User Context from HubSpot: Retrieving user data (pricing plan, assigned Account Manager) to personalize responses.
- Multi-language Support: Automatic translation of queries into English (for RAG) and responses back into the user's language.
- Team Notifications: Automatic alerts for Account Managers or the Customer Success team for specific inquiries (e.g., regarding churn).
- Real-time Pricing: A dedicated Tool (Make.com webhook) that provides the agent with current pricing data fetched from the website.
- Extended Logging: An additional Tool (Make.com webhook) that saves detailed information about the conversation flow (Playbook used, language) for analytical and debugging purposes.
Playbook configuration in Vertex AI Agent Builder, defining conversation steps, tools used, and parameters.
Make.com scenario acting as middleware, integrating HubSpot Chatflow with the AI Agent's API and handling additional functions (logging, price retrieval).
Key Benefits for the Client
- 24/7 Support Availability: Immediate responses to customer inquiries at any time.
- Reduced Response Time: Significant acceleration of responses to common questions.
- Reduced Support Team Workload: Automation of handling repetitive inquiries allows the team to focus on more complex issues.
- High Quality of Answers: Using RAG (Vertex AI Search) ensures answers are based on current company documentation.
- Scalability: The solution easily scales with the growth in the number of inquiries and the knowledge base.
- Personalization: Taking user context (plan, AM) into account in responses.
- Proactive Notifications: Alerts for relevant teams (AM, CS) at key moments of customer interaction.
- Multilingualism: Serving customers in their preferred language.
- Data Insights: Detailed conversation logging enables analysis and continuous improvement of the agent.
Frequently Asked Questions (FAQ)
The implementation time depends on the complexity of the required conversation flows (Playbooks), the number and type of integrations (Tools), the volume and structure of the knowledge base (for Vertex AI Search), and specific client requirements. A basic implementation of an agent handling a few main topics can take from 2 to 4 weeks, while more advanced projects with numerous integrations may take longer. The process includes needs analysis, architecture design, configuration in Vertex AI, creating scenarios in Make.com, training, and testing.
Maintenance costs mainly consist of fees for Google Cloud services (Vertex AI Agent Builder, Vertex AI Search) and the Make.com platform. Vertex AI fees depend on the number of processed queries and resource usage (e.g., RAG). Make.com has free and paid plans, depending on the number of operations and data transfer. Costs are generally flexible and scale with usage. Compared to the cost of maintaining a 24/7 support team, an AI solution is often more cost-effective, especially with a high volume of repetitive inquiries.
Vertex AI Agent Builder offers a graphical interface that simplifies managing Playbooks and configuring the agent without needing deep programming or AI knowledge. Basic management of the knowledge base content (for RAG) also does not require specialized skills. However, designing complex conversation flows, configuring advanced tools (Tools), or optimizing prompts may require some technical knowledge or expert support.
Thanks to the implemented architecture with automatic translation, the agent can handle inquiries in multiple languages. The query is translated into English to search the knowledge base, and the answer is then translated back into the user's original language. This allows for effective support for global customers while maintaining a single, primary knowledge base.
Standard HubSpot Chatflows rely on predefined decision trees or simple keyword-based bots. The agent based on Vertex AI Agent Builder uses advanced language models (LLMs) and a Playbook architecture, allowing for a more natural understanding of user intent, flexible conversation management, and the execution of complex tasks. Integration with Vertex AI Search (RAG) enables dynamic use of an extensive knowledge base, which is not offered by standard Chatflows. Additionally, custom integrations (Tools) and logic implemented in Make.com provide much greater automation and personalization capabilities.
The agent uses the RAG (Retrieval-Augmented Generation) mechanism based on Vertex AI Search. This means its knowledge comes mainly from the provided data (documentation, FAQ, websites). Updating the agent's knowledge involves updating these data sources. The language models (LLMs) themselves are managed by Google Cloud. Additionally, analyzing conversation logs helps identify areas for improvement in the configuration of Playbooks, prompts, or conversation examples, enabling continuous refinement of the agent's performance.
Yes. By using Vertex AI Agent Builder, which supports defining custom API-based tools (Tools), and the flexibility of the Make.com platform as middleware, the agent can be integrated with virtually any system that has an API. This can include CRM systems, databases, e-commerce platforms, project management systems, internal communication tools (e.g., Slack), and many others.