The Impact
18+
20+
4
24/7
About the Customer.
REGION
INDUSTRY
JUMP TO SECTION
Share this Article
Executive Summary
A global workforce management company serving multinational organizations engaged Matellio, an Amazon Web Services (AWS) Advanced Partner, to eliminate its dependence on manual support operations for routine HR queries. The company’s Time and Expense Management platform had no self-service query capability — employees and managers depended on support teams for information already available in the system. Matellio built a production AI chatbot on Amazon Bedrock, embedded directly into the platform, that answers natural language HR queries in 18+ languages, enforces role-based access at the query level, and routes every question to the appropriate response pipeline automatically.
Customer Challenge
The company’s support operations were heavily dependent on manual intervention, human availability, and repeated knowledge sharing. A significant proportion of support requests involved routine queries – timesheet lookups, leave balance checks, policy questions, and process guidance – for information already available in the system but not easily accessible to users.
As the platform expanded across regions and profiles, this created compounding risk:
Routine queries required human intervention even when answers were already in the system.
Individual team members gave different answers to the same question, eroding user trust.
Users in one region waited hours for responses from support teams in another.
New users had no self-service path to understand platform workflows.
Every new user added to the platform added to the support burden — with no structural improvement in sight.
Without a scalable solution, the company faced rising costs, slower resolution times, declining user satisfaction, and an inability to grow its platform operations without proportionally growing its support headcount.
Why AWS
The company’s platform was already hosted on Amazon Web Services (AWS), with its web server on Amazon Elastic Compute Cloud (Amazon EC2) and production data in Amazon Relational Database Service (Amazon RDS) MySQL. Building the chatbot on AWS meant direct connectivity to the live production database inside the same Amazon Virtual Private Cloud (Amazon VPC), with no data leaving the existing security boundary.
Amazon Bedrock provided managed access to Claude — the large language model (LLM) with the strongest multilingual and natural language to SQL (NL-to-SQL) performance in evaluation — alongside Amazon Bedrock Knowledge Base for profile-isolated vector search and AWS Lambda for serverless, auto-scaling execution.
Matellio had built the company’s foundational platform phase, giving the team direct knowledge of the production database schema, the role-based access control (RBAC) model, the JWT token structure, the SAP integration schedules, and the application business rules.
Combined with proven production experience on Amazon Bedrock model invocation, Bedrock Knowledge Base metadata filtering for multi-tenant document isolation, and AWS Lambda optimization for latency-sensitive workloads, Matellio could move from requirements to working production code without a knowledge transfer phase.
Partner Solution
Matellio built a production AI chatbot embedded directly inside the company’s Time and Expense Management platform – a system that previously had no self-service query capability. The chatbot answers natural language questions drawn from the live production database, enforces role-based access automatically, supports 18+ languages without manual selection, and maintains persistent conversation memory across sessions.
AWS Lambda serves as the core serverless backend, running the complete orchestration pipeline on every request: JWT token decoding for role and profile identification, intelligent query routing, NL-to-SQL generation, retrieval-augmented generation (RAG) retrieval, and natural language response generation. Amazon API Gateway exposes the chatbot as a secure REST API endpoint, routing authenticated HTTP requests from the web frontend to Lambda.
Amazon Bedrock hosts Claude Sonnet and Claude Haiku as the reasoning models — used for query classification, NL-to-SQL generation, language detection, and response generation depending on the task.
Amazon Bedrock Knowledge Base with Amazon OpenSearch Service stores and retrieves policy and procedural documents using vector search, with 20+ profile-isolated knowledge bases and metadata filtering to prevent cross-profile data leakage.
Amazon RDS MySQL serves as the production database for timesheet, quota, attendance, and project data, and stores full conversation history per session for multi-turn context. Amazon S3 stores source policy documents indexed into the Bedrock Knowledge Bases.
How the System Works — Step by Step
Every query is classified on arrival – timesheet data, quota and leave, policy questions, or general conversation – and handled by the appropriate pipeline. Users never configure anything.
The system decodes the user’s JWT token on every request. A regular employee sees only their own records. A project manager sees their team. An HR associate’s scope depends on their vendor assignment. This logic covers 20+ account types and 20+ company profiles – enforced at both the routing and knowledge base level.
For data queries, AWS Lambda invokes Amazon Bedrock (Claude Sonnet) to convert the user’s plain language question into a precise SQL query against the live Amazon RDS MySQL database. The result comes back as a conversational answer.
For policy and procedural queries, Amazon Bedrock Knowledge Base retrieves the relevant documents from profile-isolated knowledge bases using vector search with metadata filtering — ensuring users only ever see documents for their own company profile.
Language is detected from the user’s question. The entire response pipeline — SQL results, knowledge base answers, error messages, conversational replies — is returned in the user’s detected language. No language selection required.
Every conversation turn is stored in MySQL and retrieved at the start of each new request, enabling multi-turn conversations where users can ask follow-up questions without repeating context.
Solution Architecture
The architecture below shows the end-to-end AWS deployment: the core orchestration layer, API layer, NL-to-SQL pipelines for timesheet and quota queries, RAG knowledge base retrieval for policy questions, and the data layer.
Results and Benefits
The AI chatbot is live in production and delivered the following qualitative and operational improvements across the company’s support operations. Quantitative metrics will be confirmed following client review and approval.
Employees and managers can now resolve routine HR queries instantly without raising a support ticket or waiting for a team member to respond.
Support is available in 18+ languages, auto-detected from the user’s question — removing language as a barrier for global users across all regions.
Every response is scoped to the user’s role and profile automatically — a regular employee sees only their own data, while HR associates and project managers see data appropriate to their access level.
Responses are generated from the live production database and indexed policy documents, eliminating the inconsistency that came from individual team member knowledge.
A structured feedback loop captures thumbs-up and thumbs-down ratings on every response, with negative ratings storing correction comments — creating a dataset for ongoing SQL prompt refinement and knowledge base updates.
Support capacity no longer scales linearly with user growth — the chatbot absorbs routine query volume regardless of timezone, user count, or time of day.
Let’s turn your next big idea into a scalable product, engineered for speed, reliability, and real-world impact.
