Executive Summary:
The HR platform had two recurring problems: teams waited for manual performance reports, and sales staff answered the same product questions repeatedly. Rapyder deployed two Amazon Bedrock assistants: a Text-to-SQL bot in Microsoft Teams and a documentation-grounded RAG assistant for pre-sales. Internal analytics responses became 65–75% faster, pre-sales workload fell 40–50%, and each channel supports more than 1,000 queries per day.
About Customer:
The customer is a fast-growing HR technology platform that helps organizations manage OKRs, automate performance reviews, and build a fair, transparent, engagement-driven workplace culture. The platform integrates behavioral science, data analytics, and machine learning to give enterprises tools for transparent performance tracking, and enables HR leaders and managers to monitor employee productivity, track objectives, and reward high-performing employees through gamified engagement models.
Problem Statement:
HR teams depended on manually generated reports to answer routine OKR, KPI and review-completion questions. Without self-service analytics, managers needed technical support to access performance insights. At the same time, sales staff repeatedly answered product and implementation questions, creating a pre-sales workload that was difficult to scale as prospect interest grew.
Industry:
Industries, Others
Offering:
Offerings, Data Analytics
Business Challenges:
- Slow internal answers: Internal teams could not get quick answers to performance and analytics questions without manual report generation.
- Manual manager reports: Managers waited on manually generated reports to answer basic OKR, KPI, and review-completion questions.
- No self-serve analytics: There was no self-serve way for HR staff to query performance data without technical or analytics support.
- Repetitive sales questions: The sales team spent significant time manually answering repetitive product and implementation questions from prospects.
- Scaling pre-sales work: Manual pre-sales query handling did not scale as inbound interest and prospect volume grew.
Solution Implemented:
- Text-to-SQL assistant: Deployed a Text-to-SQL conversational analytics assistant embedded in Microsoft Teams, powered by Amazon Bedrock with Claude 4 Sonnet.
- Change Data Capture: Replicated MongoDB collections into Amazon S3 as Parquet files via AWS Database Migration Service with Change Data Capture, keeping analytics data continuously current.
- Schema catalogue: Registered up-to-date schema metadata in the AWS Glue Data Catalog, using an AWS Glue Crawler to scan the S3 datasets.
- Bedrock query flow: Used AWS Lambda to invoke Amazon Bedrock with schema context, converting natural language questions into schema-compliant SQL, executed via Amazon Athena.
- RAG pre-sales assistant: Deployed a Retrieval-Augmented Generation (RAG) pre-sales virtual assistant on the customer’s website, grounding responses in product manuals and documentation embedded as vectors in MongoDB.
- Access controls: Enforced Azure Active Directory authentication and role-based access control so Teams users only retrieve performance data relevant to their role.
- Audit logging: Logged every query, retrieved context, and response in Amazon DynamoDB for both chatbots, supporting auditability and future omnichannel expansion.
Technology Stack:
- Amazon Bedrock (Claude 4 Sonnet)
- AWS Lambda
- Amazon Athena
- AWS Glue Data Catalog & Crawler
- AWS Database Migration Service (DMS)
- Amazon S3
- Amazon API Gateway
- Amazon EC2
- Amazon DynamoDB
- Amazon CloudWatch
- AWS CloudTrail
Business Benefits:
- 65–75% faster analytics response – Natural language Text-to-SQL replaced manual report generation for internal performance and analytics questions.
- 88–92% SQL generation accuracy – Schema-aware prompting with Claude 4 Sonnet and Glue Catalog context kept generated queries reliably schema-compliant.
- 1,000+ queries handled daily per channel – Serverless Lambda, Athena, and DynamoDB scaled query capacity without added infrastructure.
- 40–50% lower pre-sales workload – RAG-based automation handled repetitive product and pricing questions, freeing the sales team to focus on qualified leads.
- 85–90% RAG response accuracy – Retrieval-grounded generation from verified product documentation kept pre-sales answers accurate and low-hallucination.
- 25–35% lower analytics infrastructure cost – A serverless Lambda/Athena model replaced the cost profile of traditional BI infrastructure.
Still answering HR and product questions manually?
Build analytics and pre-sales assistants with Amazon Bedrock.
Key Takeaways:
- Schema-aware analytics – Schema-aware prompting and Glue Catalog context matter as much as the model when Text-to-SQL runs at scale.
- Documentation grounding – Grounding answers in verified product documentation gives the pre-sales assistant a dependable source of truth.
- Role-based access – Role-based access must be enforced before a query reaches sensitive HR data.
Frequently Asked Questions
For HR managers, routine questions about OKRs, KPIs and review completion previously meant waiting for a manual report. Text-to-SQL converts their plain-language questions in Microsoft Teams into SQL, using Amazon Bedrock with schema context and Amazon Athena to query the data.
Pre-sales questions need product-specific answers rather than generic AI responses. The RAG assistant retrieves relevant passages from verified product documentation stored as vectors in MongoDB and uses that context to generate responses. This grounding helps reduce unsupported answers; it does not guarantee error-free responses.
Because the assistant queries employee performance data, access controls are part of the query workflow. Azure Active Directory authenticates Teams users, while role-based access controls restrict retrieval to performance data relevant to their role.
Both the internal Teams assistant and the public pre-sales assistant needed capacity for changing query volumes. AWS Lambda, Amazon Athena and Amazon DynamoDB provide serverless query processing and storage, supporting more than 1,000 queries per day per channel.
