Employee Self-Service AI

Organizations process 211,000+ HR tickets annually, with 70% being routine questions.

Employees frequently need answers to:

• HR policies and benefits
• IT troubleshooting
• Legal compliance questions
• Operations procedures

This creates:

• 3-5 day delays for simple questions
• $75/ticket support costs
• Inconsistent responses across teams
• Employee frustration
• Overwhelmed support teams

Designed an AI agent with intelligent scope definition, guardrails, and escalation workflows:

Technical Architecture:

• RAG (Retrieval-Augmented Generation) for accurate, explainable answers
• Role-based permissions for sensitive data handling
• Accuracy validation mechanisms to maintain quality
• Seamless ServiceNow integration for ticket creation and tracking

Phased Rollout Strategy:

1. Phase 1: High-volume, low-risk HR queries (benefits, PTO)
2. Phase 2: IT troubleshooting and common issues
3. Phase 3: Legal and compliance questions with escalation paths

Focus on augmentation, not replacement - complex cases still route to human experts.

Cross-Functional Team Leadership:

• Led ML engineers, backend developers, UX designers, and security team
• Partnered with HR, IT, Legal, and Compliance stakeholders
• Defined knowledge domains and escalation paths across departments
• Conducted workshops to understand unique needs of each domain

Key Technical Decisions:

RAG vs. Fine-Tuning:

• Chose RAG architecture for better explainability and easier knowledge base updates
• Trade-off: Slightly higher latency, but more transparent and maintainable

Accuracy vs. Speed:

• Prioritized accuracy (95%+ threshold) over instant response
• 2-3 second latency acceptable vs. instant but less accurate responses

Phased Feature Rollout:

• Started with text-only responses before adding document retrieval
• Ensured quality baseline before expanding capabilities

Challenges Navigated:

• Strict data privacy regulations (GDPR, CCPA)
• ServiceNow API rate limits affecting scalability
• Existing knowledge base quality varied widely across departments
• SOC 2 compliance requirements throughout implementation
• Change management for employees used to ticket-based support

Market Differentiation:

vs. Generic Chatbots:

• Built deep ServiceNow integration (not just API wrapper)
• Role-based access controls for sensitive data
• Domain-specific knowledge (HR, IT, Legal, Compliance)

vs. Microsoft Viva & Workday Assistant:

• Multi-domain coverage (not just HR)
• Superior escalation workflows with context preservation
• Better accuracy through RAG architecture vs. generic LLMs

Key Advantage: Purpose-built for enterprise support workflows, not adapted from consumer chatbots.

Key Learnings:

The phased rollout was crucial for building trust. Starting with low-risk HR queries (benefits, PTO) allowed us to refine the system before handling sensitive legal/compliance topics.

Most Important Lesson: "Garbage in, garbage out" applies strongly to RAG systems. Would invest more in knowledge base cleanup upfront - quality documentation directly determines AI answer quality.

What Worked Well:

• Building confidence with HR queries before legal/compliance
• Human escalation paths for edge cases
• Role-based access controls for sensitive data

What I'd Do Differently:

• Spend 2-3 weeks on knowledge base audit before AI integration
• Create content quality standards and templates earlier
• Build internal content owner champions in each department