Understanding the Payemoji AI Chatbot

Understanding the Payemoji AI Chatbot

Understanding Our AI Chatbot: Compliance with the EU AI Act

Introduction

Our AI-powered chatbot delivers efficient and tailored customer interaction by leveraging predefined workflows and WhatsApp-native features such as button messages, interactive lists, external link buttons, and dynamic forms. Designed to handle common customer queries, the chatbot ensures a seamless user experience while aligning with your brand’s voice and objectives.

This document explains the chatbot’s technical foundations, operational flow, benefits, and limitations, with a focus on compliance with the EU AI Act (Regulation (EU) 2024/1689).  Payemoji’s AI service require transparency measures, such as disclosing AI interaction to users and offering human escalation options. We achieve compliance through audited Azure cloud services, GDPR-aligned data handling, and regular system performance reviews.

How the AI Works

Our chatbot employs a Retrieval-Augmented Generation (RAG) architecture, powered by Azure OpenAI for natural language generation and Azure AI Search for intelligent data retrieval. RAG combines the information retrieval and generative AI to produce accurate, context-specific responses grounded in customer-provided data. The process is illustrated and detailed below.

 

Data Ingestion and Preparation            

  1. Website Crawl: We conduct a weekly crawl of your specified website domains to collect up-to-date content. Our crawler adheres to robots.txt instructions, ensuring only permitted sections are accessed. Customers must explicitly approve crawls, and you can prune specific sections (e.g., internal portals, outdated pages, product catalogues where catalogues are uploaded seperately) to maintain relevance.

  1. Customer-Provided Data: You supply additional data, such as FAQ documents or individual “facts” (e.g., product specifications or policy statements) as well as entire product catalogues to ensure accuracy. These are securely ingested and indexed alongside website content for efficient retrieval.

User Interaction via WhatsApp

Users interact with the chatbot through WhatsApp’s messaging API, which supports structured workflows:

·            Button Messages: Quick options like “Yes/No” or topic selectors (e.g., “Check Order Status” or “Product Info”).

·            Interactive Lists: Multi-option menus for navigating complex queries (e.g., selecting a product category).

·            Forms: Structured inputs for tasks like submitting feedback or updating account details.

·            External Links: Call buttons, URL links

·            eCommerce Product Catalogues and Checkout: native WhatsApp eCommerce product selectors and cart functionality

Notes
For predefined workflows, the chatbot responds instantly. For open-ended or complex queries, it proceeds to the RAG process.

The user’s WhatsApp message is sent to Azure AI Search, which employs semantic ranking to identify the most relevant content from your indexed data.    Semantic ranking uses advanced natural language processing to understand query intent, returning the top 5 results based on contextual relevance rather than simple keyword matching.  Top 5 is default however this can be customized depending on the individual customer particularly those with large product catalogues.

Notes
Example: A query like “What’s the return policy for ABC Limited?” retrieves FAQ entries or website sections about returns, even if phrased differently.

Response Generation

The top 5 search results are passed to Azure OpenAI, along with a customer-specific prompt tailored to your brand’s tone and policies.     The default prompt includes strict instructions:
- Respond only using the provided search results and customer data, ignoring the AI’s general knowledge base.

- If no relevant results are found or the query is off-topic, use a fallback response, e.g., “I’m sorry, I don’t have that information available, but I can help with anything else related to ABC Limited!”

The AI summarizes the results into a concise, user-friendly response, delivered via WhatsApp with optional interactive elements (e.g., buttons for follow-up actions).

Notes
Example: For a query about store hours, the chatbot might respond, “ABC Limited stores are open from 9 AM to 6 PM, Monday to Saturday. Would you like to check hours for a specific location? [Yes/No].”

Monitoring and Updates

·            Interactions are logged securely (with anonymization) to monitor performance and compliance.

·            Weekly website recrawls keep the data fresh, and you can request manual updates for FAQs or facts as needed.

·            Regular review of logs combined with different methods of customer and user feedback.

·            Updates are made based on recommendations and mandates from Azure to maintain compliance using their service.  Security and content policies from Azure are also used to provide the service

Benefits of Using RAG

The RAG approach offers several advantages for your business and end-users:

·          Reduced Hallucinations: By grounding responses in your verified data, RAG minimizes the risk of generating incorrect or fabricated answers, enhancing trust and reliability.

·         Customized Messaging: Responses reflect your brand’s voice and content, ensuring consistency across customer interactions (e.g., matching your website’s tone or policy wording).

·       Efficiency and Scalability: Semantic search enables quick and accurate responses.

·       Compliance Support: RAG’s reliance on customer data supports EU AI Act transparency requirements by making responses traceable to specific sources, reducing biases from broad AI training data.

·        Flexibility: Weekly crawls and manual data updates allow the chatbot to adapt to changing information, such as new products or policies.

Potential Downsides of RAG

While RAG is a great solution to ensure focus on your brand voice and limiting the expansiveness of responses, it has limitations:

·         Dependency on Data Quality: Outdated, incomplete, or inconsistent data (e.g., old FAQs or old product catalogues) can lead to inaccurate responses. Regular updates and review are key.

·         Limited Scope for Novel Queries: RAG excels at factual responses but may struggle with queries outside your data, resulting in frequent fallback messages for off-topic questions.

·         Performance Overhead: Semantic search and summarization may introduce minor latency (typically milliseconds) although this is minimized through the use of Azure’s infrastructure.

We address these through regular data audits, customizable prompts, and performance monitoring, but your input is essential for optimal results.

What This Means for You as a Customer

To maximize the chatbot’s effectiveness and ensure compliance with the EU AI Act, customers have the following responsibilities and best practices:

·            Provide Up-to-Date Data: Regularly update FAQs, facts, or time-sensitive information (e.g., promotions or policy changes) not covered by website crawls. Outdated data may lead to incorrect responses, impacting user trust.

·            Test AI Responses: Before and after deployment, simulate common queries to verify accuracy, tone, and workflow integration. For example, test queries like “How do I return an item?” or “What are your prices?” to ensure proper handling.

·            Review Website Content: Approve crawls and exclude irrelevant or sensitive sections (e.g., employee portals) via
robots.txt or manual pruning. Regularly check your website for outdated content.

·            Monitor for Errors: Review chatbot logs and error logs periodically to identify unintended patterns, such as biased responses from source data (e.g., FAQs favoring one product). Report issues for prompt correction.

·            Ensure User Transparency: Inform end-users they’re interacting with an AI, as required by the EU AI Act. For example, include a welcome message like, “Hi! I’m ABC Limited’s AI assistant. How can I help you today?" 

·            Offer Human Escalation: Provide clear options for users to escalate to human support, such as a button or command (e.g., “Talk to a person”). This aligns with the Act’s emphasis on user control.

·            Comply with GDPR: Ensure any customer data shared with us (e.g., in forms) complies with GDPR. We handle data securely, but you’re responsible for obtaining user consents where applicable.

Conclusion

Our AI-powered WhatsApp chatbot leverages RAG to deliver accurate, brand-aligned support while meeting EU AI Act requirements for transparency and accountability. By understanding its functionality and maintaining high-quality data, you can ensure a reliable, compliant, and user-friendly experience.

InfoFor further assistance or customization, contact our support team at support@payemoji.com.



    • Related Articles

    • Payemoji Dynamic Messaging

      Payemoji Conversational & Dynamic Messaging Payemoji Dynamic Messaging allows the use of a single API to send many different Notification message (see how to send a Notification Message here) and other message types. WhatsApp and other channels allow ...

    • What is Payemoji? How does it work

      Picture this: 600 million people message businesses on WhatsApp every day, and it’s growing 50% year over year. Why? Because WhatsApp delivers: 98% open rates, 60% click-throughs, and trusted, verified senders. Compare that to spammed emails, ignored ...

    • FAQ: Integration of Payemoji WhatsApp Business Platform and Firefish Software Recruitment CRM

      FAQ: Integration of Payemoji WhatsApp Business Platform and Firefish Software Recruitment CRM Overview What is the Payemoji and Firefish integration? The integration between Payemoji’s WhatsApp Business Platform and Firefish Software’s recruitment ...

    • Run a Campaign With Payemoji

      Maximize Your Marketing Campaigns with WhatsApp Payemoji have a Campaign management system built right inside our unified inbox! You can send a message to one person or thousands with a few clicks. This article shows you how. 1. Log Into Your ...

    • Provisioning a WhatsApp Number with Payemoji

      1. Log Into the Payemoji Portal Log in at https://portal.payemoji.com/login using your email and the password you have set (if you have not set a password please use Forgot Password). 2. Log Into Facebook You will be greeted with a sign up page. ...