AWS Summit Taipei 2026: Decomposing Ontology and Agentic AI - Using Amazon Bedrock to Bring Living Water to Manufacturing ERP

by Ernest Chiang

Embed

Start on current slide

Slide 1

Slide 1 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. © 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Ernest Chiang He/him Managing Director Kyklosify Decomposing Ontology and Agentic AI on Amazon Bedrock for Manufacturing ERP L T - 0 0 1

Slide 2

Slide 2 text

Slide 3

Slide 3 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Ernest Chiang Director of Product & Technology Integration, PAFERS ex-Sports & Fitness Working Group, Bluetooth SIG ex-Sr. Process Integration Engineer, TSMC

Slide 4

Slide 4 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Make reality computable. We are a boutique consulting firm serving 3+ publicly traded companies and 10+ small and medium-sized enterprises.

Slide 5

Slide 5 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. • The Problem and a Live Demo • The Humbled Ontology • Core Concepts and Compliance on Amazon Bedrock • Agentic AI in Practice • Actions and resources

Slide 6

Slide 6 text

Slide 7

Slide 7 text

Slide 8

Slide 8 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Buried in Menus and Codes The ERP runs on a legacy terminal or dated Windows interface, with functions hidden across nested menus and codes—new hires must memorize countless function paths and transaction codes. No Easy Way to Ask Questions Every Improvement Is a Project Improving anything means deploying BI or outsourcing a custom build—easily two months plus a dedicated project team, and every change in requirements kicks off another mini-project. The Problem There's no natural language search, so a simple question like "who did we last buy this from, and at what price" means exporting a report and filtering in Excel—and on the shop floor, where computers and access are scarce, workers often just call the office and wait.

Slide 9

Slide 9 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Ask in Plain Language What if anyone could ask "who did we last buy this material from, and at what price" in plain language—and get an answer in seconds? An agent on Bedrock understands the question, queries your ERP directly, and returns the result. No menu paths, no transaction codes, no training required. Answers Anywhere, for Anyone Change Requests, Not Projects What if a new question didn't require a new project? Because the agent reasons over your data instead of relying on pre-built reports, evolving needs become a conversation—not another two-month BI build with a dedicated team. What-if What if the shop floor never had to call the office again? The same agent works from any phone or tablet, so workers get answers on the spot—no dedicated workstation, no waiting, no lost time.

Slide 10

Slide 10 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. The Setup • A manufacturer with four plants, each running the same ERP • Material module exported as four separate Excel files—exactly the kind of data that's stuck in spreadsheets today The Build The Scenario • Agentic AI (Kiro, or Claude Code) reads the Excel files and understands the columns and data on its own • Processes the data through a pipeline into a relational database (SQLite or Amazon RDS) • Generates a text-to-SQL app with a chat UI, powered by Claude models on Amazon Bedrock The Result • A tool floor and procurement teams can use every day

Slide 11

Slide 11 text

Slide 12

Slide 12 text

Slide 13

Slide 13 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Nothing New—Or Is It? You've probably seen chat UIs like this before, and you might be thinking: what's the big deal? Fair enough. On Your Own Data, in an Hour But what if this same demo ran on your own production line—on the Excel or CSV files exported from your plant's ERP? And what if you could build a prototype to validate it yourself, in under an hour? One Prompt to Budget Conversation What if a single prompt was all it took—so the very next day, you could walk into your manager's office and start the budget conversation? What If You Could Build This Yourself?

Slide 14

Slide 14 text

Slide 15

Slide 15 text

Slide 16

Slide 16 text

Slide 17

Slide 17 text

Slide 18

Slide 18 text

Slide 19

Slide 19 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Knowledge Why You Need Both Workflow The nouns—what the system knows: your materials, suppliers, orders, plants—the entities and facts that describe your business. This is the "knowing." A noun alone is just a record; a verb alone has nothing to act on. Only when knowledge and workflow come together do they form a loop—and a loop is what lets the system iterate and get better over time. The verbs—what the system does: query, compare, decide, act. This is the "doing"—the actions taken on top of what's known. Ontology (The Simple Version)

Slide 20

Slide 20 text

Slide 21

Slide 21 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Generative AI Agentic AI Responds to prompts by generating content—text, code, images. It answers. You ask, it replies, and the next step is up to you. Plans, calls tools, and takes multi-step action to reach a goal. It doesn't just answer—it acts. Give it an objective, and it figures out the steps, uses the tools it needs, and gets the job done. Generative AI vs. Agentic AI

Slide 22

Slide 22 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. RAG Agentic RAG Retrieve, then generate—a single question, a single answer. The system pulls relevant context first, then responds. One pass, and it's done. The agent decides for itself when to retrieve, what to retrieve, and whether to look again. It routes between SQL and vector search, chains multiple steps, and checks its own work—looping until the answer holds up. RAG vs. Agentic RAG

Slide 23

Slide 23 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Data-Isolation Secure Foundation Every call stays inside your own AWS account. Your data is never used to train the underlying models. And nothing crosses your boundary—your data stays where it belongs. Amazon Bedrock runs on the AWS Nitro System with zero operator access, and content safety is enforced through Guardrails. Compliance — Amazon Bedrock

Slide 24

Slide 24 text

Slide 25

Slide 25 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. On Development On Operation During the build phase, agentic AI runs through Kiro or Claude Code paired with Claude models on Amazon Bedrock. The point isn't that "AI can write code"—it's the team's prompting style: give context, define done, trigger long autonomous runs, and get it right in one pass. Ideal for spinning up prototypes fast—for proposals, interviews, and mapping out processes. Once in production, agentic AI runs on Amazon Bedrock AgentCore—Runtime, Memory, Identity, Gateway, and Observability. This is where it graduates from answering to acting: checking inventory, comparing BOMs, generating shortage lists, and drafting re- quote requests. Built for the long haul— operation, maintenance, and everything that keeps running after launch. Agentic AI — On Development & On Operation

Slide 26

Slide 26 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Data Knowledge Option The raw material— the exported Excel files, CSVs, and ERP records before anything is made sense of. It's the foundation everything else is built on, and it lives in storage. The nouns, structured. Once data is organized into entities and relationships— materials, suppliers, orders, plants—it becomes something the system can reason over, not just rows in a sheet. The connective layer between knowing and doing. Before acting, the system weighs context and decides what matters: which supplier, which threshold, whether this case is even worth acting on. Ontology (The Full Version) Workflow Memory The verbs—the actions taken on what's known: query, compare, generate, draft. This is where judgment turns into concrete steps that move work forward. The system checks its own output and feeds the result back in—did the action hold up? This is what closes the loop, letting Data through Workflow improve with every cycle instead of repeating the same mistakes. Information Namespace

Slide 27

Slide 27 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Ontology (The Full Version) The Implementation Lens Every stage needs data and storage; every action needs compute to run—whether it's calculation or logic. In practice, the whole ontology maps cleanly onto two AWS primitives: storage + compute.

Slide 28

Slide 28 text

Slide 29

Slide 29 text

Slide 30

Slide 30 text

Slide 31

Slide 31 text

Slide 32

Slide 32 text

Slide 33

Slide 33 text

Slide 34

Slide 34 text

Slide 35

Slide 35 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Agentic AI on Development Workflow Data Knowledge Knowledge Option / Memory Option Workflow Knowledge Workflow Workflow Option Kiro / Claude Code

Slide 36

Slide 36 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Agentic AI on Operation Agent Workflow Data Knowledge Knowledge Option / Memory Option Workflow Knowledge Workflow Workflow Option

Slide 37

Slide 37 text

Slide 38

Slide 38 text

Slide 39

Slide 39 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Fill out the feedback form, and we'll send you the file pack to practice on your own. The pack includes sample Excel files from all four plants, plus a multi-thousand- word prompt from our meeting notes— everything you need to recreate the live demo from today's talk yourself, in under an hour, using Kiro or Claude Code. Take-Home Workshop Sample Pack

Slide 40

Slide 40 text

Slide 41

Slide 41 text

Slide 42

Slide 42 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. • Ontology ~= Knowledge + Workflow • Kyklosify Business AI Stack • Focus on WHY, not HOW. • Small build, small deploy, small win. • Start small, scale fast. Takeaways

Slide 43

Slide 43 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. © 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. Q&A Traveller of Decomposition and Integration https://www.ernestchiang.com/ AWS User Group Taiwan (AWSUG-TW)

Slide 44

Slide 44 text

Slide 45

Slide 45 text

© 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. © 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. D e v e l o p e r L o u n g e | J U L Y 1 5 , 2 0 2 6