H2O.ai Agents : From Theory to Practice - Support Presentation
1 like•147 views
This is the support slide deck for H2O Agents AI: From Theory to Practice course. These slides cover AI agent architecture, h2oGPTe capabilities, industry applications across finance, healthcare, telecom, and energy sectors, plus implementation best practices. They're designed as a helpful reference while following the video course or for quick review of key concepts in agentic AI. To access the full course and more AI learning resources, visit https://h2o.ai/university/
![H2O.ai Confidential
Key Ideas : Agentic AI for Contact Centers
Automate customer conversation
Automate during-call action items
- Conversations [Data]
- Meta Data [Documents]
- Intent Identification
Agentic Actions
- Conversation Intent Validation + Metadata Matching
- Action Identification
- API / Function call identification
- Executing the Function call (ie. automated actions)
- Database Update
- CRM update
- Notifications, Logging
Automated Quality Assurance](https://image.slidesharecdn.com/h2o-250411064807-38d7d0e6/85/H2O-ai-Agents-From-Theory-to-Practice-Support-Presentation-30-320.jpg)
![H2O.ai Confidential
h2oGPTe Retrieval Tool Legend:
Experiment Autodoc
H2O DAI
Documentation
Stage 1
Stage 2
[optionally]
Researcher’s
suggestions passed
to engineer
Launch experiment
and report failures
to agents
Stage 3
Chat
Chat
Manager
User-proxy agent
Researcher
Assistant agent
Manager
User-proxy agent
Engineer
Assistant agent
User proxy
LLM Access](https://image.slidesharecdn.com/h2o-250411064807-38d7d0e6/85/H2O-ai-Agents-From-Theory-to-Practice-Support-Presentation-34-320.jpg)
H2O.ai Agents : From Theory to Practice - Support Presentation
- 1. H2O.ai Confidential H2O.ai Agents AI From Theory to Practice H2O.ai
- 2. H2O.ai Confidential Introduction, Spectra of AI Agents, Tool Use 1. Intro to Agentic AI Features, GAIA Benchmark 2. h2oGPTe Agents Advice and guidance for implementing AI Agents in practice 4. AI Agents in Practice Goal Gain an understanding of AI agents and how to effectively implement real-world use cases in h2oGPTe. Audience AI enthusiasts interested in AI agents as well as enterprise technology leaders and business stakeholders who need to understand and implement h2oGPTe agents in their organisations. Agenda 3. Demos & Use Cases Industry specific demos
- 3. H2O.ai Confidential Agentic AI Agentic AI actively interacts with systems, using tools or taking actions to achieve specific goals. VS Generative AI Generative AI focuses on producing content like text, images, or code based on learned patterns. RAG enhances this by retrieving external information for context. Introduction to AI Agents AI Agent Environment Agent interacts with the environment. E.g. Customer Support System Tools Memory LLM Task A task is given to the agent along with a variety of tools such as code execution or RAG.
- 4. H2O.ai Confidential Tools or Functions Functionality that can accomplish a specific task given a set of inputs and are interfaces that an agent is "aware of" and can access and use to interact with. Tools Available with h2oGPTe: ● Python and Bash code execution ● Web browsing ● Web search via Google or Bing ● Multi-modal understanding (text, images, audio) via OCR, caption, vision, and transcription models ● File handling across various formats ● Data science modeling and forecasting (including with DriverlessAI) ● Image generation ● Wolfram Alpha search ● Wikipedia search ● Multi-file code editing via Aider ● Advanced reasoning via OpenAI o1 or Qwen QwQ
- 5. H2O.ai Confidential Large Language Model (LLM) An LLM serves as the “brain” of the AI agent, coordinating between understanding user needs, planning solutions, using tools and communicating effectively. Frontier models are often preferred for building agents: ● Claude 3.7 Sonnet ● GPT-4o ● Deepseek R1 ● GPT-o1 ● Qwen 2.5 72B ● Meta LLama 3.3 70b ● Mistral Large Instruct
- 6. H2O.ai Confidential Memory Memory allows the agent to ‘learn’ and adapt over time by utilising past interactions to inform current actions. Memory is often implemented in the form of a Vector Database: ● h2oGPTe can be used to store and retrieve memory ● Chroma ● Pinecone ● Qdrant ● FAISS ● mem0
- 7. H2O.ai Confidential Agentic AI Agentic AI actively interacts with systems, using tools or taking actions to achieve specific goals. VS Generative AI Generative AI focuses on producing content like text, images, or code based on learned patterns. RAG enhances this by retrieving external information for context. Introduction to AI Agents AI Agent Environment Agent interacts with the environment. E.g. Customer Support System Tools Memory LLM Task A task is given to the agent along with a variety of tools such as code execution or RAG.
- 8. H2O.ai Confidential Spectra of Agentic AI No Agency Autonomous Execution Self-Determined Completion Limited Agency Supervised Agency Full Agency Simple LLM Calls Pure Text Response Tool Use Predefined Actions Action Planning Human Oversight
- 9. H2O.ai Confidential Spectra of Agentic AI No Agency Autonomous Execution Self-Determined Completion Limited Agency Supervised Agency Full Agency Simple LLM Calls Pure Text Response Tool Use Predefined Actions Action Planning Human Oversight
- 10. H2O.ai Confidential ENTERPRISE GENERATIVE AI h2oGPTe Agentic AI converges generative AI and predictive with purpose-built SLMs Document AI with multimodal guided JSON generation Multimodal audio & vision analysis Coding assistant. Rapid prototyping & development Autonomous agentic AI: execute multi-step workflows autonomously Citation-based verification for transparent Retrieval Augmented Generation (RAG) Customizable guardrails for AI safety Intelligent model routing: optimal model selection for every task Model risk management for enhanced compliance and interpretability The industry’s first multi-agent Generative AI platform combines code-first agents with planner, memory, and execution modules, enabling tool calling and customizable tools (coming soon), while integrating the strengths of Generative and Predictive AI with air-gapped, on-premise deployment options.
- 11. H2O.ai Confidential The best of Generative & Predictive AI: Run AutoML using Agents
- 12. H2O.ai Confidential H2O.ai h2oGPTe: The World's Best Agentic AI Sets GAIA Benchmark Record H2O.ai Takes the Top Spot in GAIA Benchmark: Surpassing Google by a Stunning 15.8% Margin Source: https://huggingface.co/spaces/gaia-benchmark/leaderboard
- 14. H2O.ai Confidential Phase 1: Market Intelligence Phase 2: Compliance Check Phase 3: Risk Assessment Analyse market trends and financial reports Review regulations & compliance requirements Calculate exposure & stress test scenarios Phase 4: Investment Decision Report Tools: Python coding AI Financial Analyst Workflow Investment Analysis for H2O.ai TOOLS Google Search Internet Access TOOLS “Ask questions about documents” TOOLS Code Execution Advanced Reasoning
- 15. H2O.ai Confidential Phase 1: Policy Research Phase 2: Public Impact Phase 3: Impact Assessment Review existing policies & legal frameworks Access service delivery & community accessibility Calculate economic & social outcomes Phase 4: Policy Decision Report Tools: Python coding AI Government AI Policy Workflow Develop a National AI Skills Development Framework TOOLS Google Search Internet Access TOOLS “Ask questions about documents” TOOLS Code Execution Advanced Reasoning
- 16. H2O.ai Confidential Phase 1: Literature Search Phase 2: Evidence Analysis Phase 3: Clinical Synthesis Search latest research & systematic reviews Evaluate quality & relevance of research findings Synthesise research into clinical recommendations Phase 4: Diagnosis Report Generation Tools: Python coding AI Medical Diagnosis Researcher Workflow Evidence-based Research Assistant TOOLS Google Search Scholar Papers Search TOOLS “Ask questions about documents” TOOLS Code Execution Advanced Reasoning
- 17. H2O.ai Confidential Telco AI Customer Retention Analyst Data Analysis & Modelling Agent Customer Profile ● Demographics ● Tenure & History Churn Data Services ● Current Subscriptions ● Usage Patterns Financial Data ● Monthly Charges ● Payment History h2oGPTe Agents Python Coding Data pre-processing & analysis. H2O Driverless AI Churn Modelling using AutoML and interpretability. Advanced Reasoning Open AI’s GPT-o1 for planning and validating methodology. Insights Visualisations ● Churn contribution by segment ● Feature importance Key Findings Churn rate breakdown per demographic, service and financial factors Recommendations Recommended intervention for reducing churn based on analysis.
- 18. H2O.ai Confidential AI Power Plant Performance Analyst Data Analysis & Modelling Agent Sensor Inputs ● Temperature ● Ambient Pressure ● Relative Humidity ● Exhaust Vacuum Power Plant Data Target Variable Net hourly electrical energy output. h2oGPTe Agents Python Coding Data pre-processing & analysis. H2O Driverless AI Churn Modelling using AutoML and interpretability. Advanced Reasoning Open AI’s GPT-o1 for planning and validating methodology. Insights Visualisations ● Effect of each ambient variable on power output ● Seasonality and trends DAI Model Machine Learning Model artifact to forecast hourly power output. Recommendations Suggest optimal operating parameters, alert thresholds and control system adjustments.
- 19. H2O.ai Confidential AI Agents in Practice Caveats & Tips Data Integration & Access Core systems and databases may not be easily accessible via API. May need heavy engineering support to uplift legacy systems. Implementing Guardrails Establish input and output guardrails to help ensure the AI agent acts within defined constraints This may be important for regulatory compliance. Testing & Validation Implement red teaming strategies and continuously monitor the behaviour of your AI agent through traces. SME Oversight Engage subject matter experts (SMEs) throughout development to ensure alignment with regulations and internal procedures.
- 20. H2O.ai Confidential Agentic vs Generative AI Agentic AI actively interacts with systems and takes autonomous actions, while generative AI focuses on content creation. Recap h2oGPTe Capabilities h2oGPTe sets new benchmarks in agentic AI performance, featuring advanced tool calling, multimodal processing, and enterprise-grade guardrails for deployment. h2oGPTe Use Cases Researcher and coding agents in h2oGPTe can decrease the time to discover useful insights and recommendations whilst remaining transparent in its thought process. Your turn! Keep learning & building your skills on h2o.ai/university
- 21. H2O.ai Confidential Agentic AI to autonomously support ticket management processes for faster resolution The agent, powered by H2OGPTe capabilities, executes a seamless workflow that includes sequential tasks: ticket tagging, routing, and suggesting recommendations. It leverages historical ticket data, conversations, and product documentation as the knowledge base. This approach enhances efficiency, improves accuracy, and adapts over time. Classifier Using h2oGPTe, add product tag to the ticket to point it to the right knowledge base Task A new ticket gets raised on the support portal Researcher Agent Case Study 1 h2oGPTe Router Ticket Management Ticket gets routed and resolution is recommended to human agent along with metadata.. Support Ticket & Product Knowledge Base Resolution
- 22. H2O.ai Confidential H2O Model Mate An intelligent assistant that helps you improve your H2O Driverless AI models using AI agents. The application provides guided suggestions and automated improvements for model optimization, interpretability, and custom objectives. Case Study 2
- 23. H2O.ai Confidential Part 2: Whiteboard Workshop
- 24. H2O.ai Confidential Post-Call Action Orchestrator The agent automatically captures commitments and required actions from customer service calls, creates structured work items, assigns them to appropriate teams, and monitors their completion while keeping customers updated on progress through their preferred communication channels. Agentic AI Use Case Ideas Document Processing Automation An AI agent manages the end-to-end processing of standard banking documents like loan applications and account openings, intelligently routing documents to appropriate departments, following up on missing information, and maintaining audit trails of all actions taken. Intelligent Customer Service Routing The agent evaluates customer inquiries in real-time, considering factors like urgency, complexity, and customer history to route queries to the most appropriate service channel or representative, while proactively gathering relevant information to expedite resolution. Regulatory Compliance Monitor An AI agent continuously monitors transaction patterns and account activities, identifying potential compliance issues and generating detailed reports for review. It adapts its monitoring parameters based on feedback from compliance officers and regulatory updates. Account Maintenance Assistant The agent proactively identifies and resolves routine account maintenance issues, such as updating expired card information, managing recurring payments, and alerting customers about potential service disruptions before they occur. Financial Education Coach An AI agent provides personalized financial education by analyzing a customer's banking behavior and knowledge level, then delivering targeted educational content and practical exercises to improve their financial literacy and decision-making skills.
- 25. H2O.ai Confidential Brainstorming & Prioritisation Whiteboarding Template 30 mins ACTIVITY 1 LOW EFFORT HIGH EFFORT HIGH REACH LOW REACH Common Pain Points Widespread issues that impact many teams but are simple to solve. Local Opportunities Single team or process challenges with clear paths to improvement. Deep-rooted Challenges Complex problems affecting specific departments or functions. Strategic Projects Critical business-wide challenges requiring major transformation. 01 Identify pain points, business goals and other ideas. 02 Rank ideas using a prioritisation matrix. 03 Focus on low effort, high reach use cases.
- 26. H2O.ai Confidential Current State Whiteboarding Template Use Case Motivation Key Stakeholders Expected Impact Current Implementation Repetitive customer queries consume agent time that could be spent on complex issues Support Agents and Training Managers More efficient use of agent expertise, allowing them to focus on high-value customer interactions Customers initially interact with a chatbot that tries to answer simple questions. … … … 30 mins ACTIVITY 2 01 Describe the motivation for the use case idea. 02 Identify key stakeholders or process owners. 03 Describe the expected impact from solving the problem. 04 Describe the solution that is currently implemented.
- 27. H2O.ai Confidential Opportunity Areas Whiteboarding Template Agentic Solution Business Impact System Integration Guardrails Deploy an autonomous agent that analyzes incoming queries, automatically resolves common issues through system actions, and only routes complex cases to human agents. Reduction in average handling time for common queries. Integration with ticketing system and relevant backend systems that can execute common customer requests (e.g., password resets, balance inquiries, statement requests). Strict verification of customer identity before any action, mandatory human review for any system changes. … … … … 01 30 mins ACTIVITY 3 Describe how agents might be leveraged to solve pain points 02 Define ROI based on business impact 03 Identify any system integration required 04 Define any input and output guardrails
- 28. H2O.ai Confidential Agentic vs Generative AI Agentic AI actively interacts with systems and takes autonomous actions, while generative AI focuses on content creation. Recap & Conclusion h2oGPTe Capabilities h2oGPTe sets new benchmarks in agentic AI performance, featuring advanced tool calling, multimodal processing, and enterprise-grade guardrails for deployment. Workshop Outcomes Through collaborative whiteboarding sessions, we identified key pain points, prioritized use cases based on ROI and effort, and developed concrete implementation strategies. Next Steps Schedule follow-up meetings to refine prioritized use cases.
- 29. H2O.ai Confidential Appendix - Contact Centers + Agentic
- 30. H2O.ai Confidential Key Ideas : Agentic AI for Contact Centers Automate customer conversation Automate during-call action items - Conversations [Data] - Meta Data [Documents] - Intent Identification Agentic Actions - Conversation Intent Validation + Metadata Matching - Action Identification - API / Function call identification - Executing the Function call (ie. automated actions) - Database Update - CRM update - Notifications, Logging Automated Quality Assurance
- 33. H2O.ai Confidential Autonomous agentic AI: execute multi-step workflows autonomously h2oGPTe Agents bring true autonomy to your workflows, enabling LLMs to handle multi-step tasks like web research, data modeling, database access, and iterative code execution. Programmatic and continuous, these agents reduce manual workload by executing tasks requiring sequential logic, real-time decision-making, and data handling. h2oGPTe Agents can autonomously generate multi-page PDFs with charts, tables, and flowcharts based on real-time data—complete with source code for full transparency.
- 34. H2O.ai Confidential h2oGPTe Retrieval Tool Legend: Experiment Autodoc H2O DAI Documentation Stage 1 Stage 2 [optionally] Researcher’s suggestions passed to engineer Launch experiment and report failures to agents Stage 3 Chat Chat Manager User-proxy agent Researcher Assistant agent Manager User-proxy agent Engineer Assistant agent User proxy LLM Access
- 35. H2O.ai Confidential Agenda (draft) 1. Goals (5 mins) a. Workshop purpose 2. h2oGPTe Agents Demo (30 mins) 3. Agentic Apps (30 mins) 4. Domain Specific Case Study Discussion (60 mins) a. e.g. Business Banking Contact Centre Flow i. Current State (CBA to prefill template) ii. Identify areas for agents iii. Metrics for success (domain specific e.g. NPS) iv. Prioritisation (ROI vs Complexity)
- 36. H2O.ai Confidential 1. Goals: Identify and prioritise use case ideas where agents can deliver the greatest ROI. 2. h2oGPTe Agents demo a. GAIA benchmark results b. Tool calling e.g. RAG Text, Web Search, Audio-Video Transcription 3. Agentic Apps a. H2O Model Mate b. Telecom Help Desk 4. Domain Specific Case Study Whiteboarding a. Describe current state i. Identify key pain points in decisioning processes ii. Identify key stakeholders or process owners iii. Describe the expected impact from solving the problem b. Identify areas where agents can be applied i. Describe how agents might be leveraged to solve pain points ii. Define ROI based on business metrics (e.g. Improve identification and rectification of compliance lapses by X%) iii. Identify any system integration required iv. Define guard rails c. Prioritise i. Prioritisation matrix to rank use cases based on ROI (low effort, high value first) ii. Define clear owners and next steps
- 37. H2O.ai Confidential Prioritisation Whiteboarding Template 01 30 mins ACTIVITY 2 Prioritisation matrix to rank use cases based on ROI. 02 Define clear owners and next steps LOW EFFORT HIGH EFFORT HIGH VALUE LOW VALUE Quick Wins Minimal risk with few guardrails and system integration is available. Fill-ins Simple automations that provide marginal ROI but have few guardrails and system integration is available. Time Sinks Complex integrations and guardrails that solve difficult business problems or benefit few customers. Strategic Projects Complex integrations, guardrail and business problem but delivers substantial ROI. 1. Assign ownership 2. Owners should arrange follow-up meetings to refine and expand use cases
- 38. H2O.ai Confidential Current State Whiteboarding Template Pain Point Key Stakeholders Expected Impact Repetitive customer queries consume agent time that could be spent on complex issues Support Agents and Training Managers More efficient use of agent expertise, allowing them to focus on high-value customer interactions … … … 01 30 mins ACTIVITY 2 Identify the Motivation Identify key pain points in decisioning processes Wish list - ideas ? Business Goals 02 Identify key stakeholders or process owners 03 Describe the expected impact from solving the problem
- 39. H2O.ai Confidential Demo ideas: ● Financial services ○ Investment Analysis ○ Risk Assessment Predictive Modeling ○ Regulatory compliance real-time monitoring ○ Portfolio optimisation market analysis ○ fraud detection pattern recognition ● Government ○ Policy Analysis document processing ○ public service optimiser ○ emergency response real-time coordination ○ grant management application processing ○ regulatory updates compliance tracking ● Telco ○ network performance real time monitor ○ customer support ○ infrastructure planning ○ service quality automated assessment ○ resource optimisation load balancing ● Energy ○ grid management real time control ○ predictive maintenance equipment monitoring ○ environmental impact emissions tracking ○ resource planning demand forecasting ○ safety protocols automated compliance ● Healthcare ○ Clinical decision support evidence based analysis ○ patient care optimiser ○ research analysis literature review ○ equipment management predictive maintenance ○ compliance management regulatory tracker Finally, reveal all use cases developed by AI