Theory & Frameworks
What do we need to know about a complex system to decide what to do next?
Rikka Chen’s personal site · Independent researcher and consultant in enterprise digitalization and AI implementationAbout me
Starting from the real world, rebuilding business, systems, and intelligence
Continuously updated
This body of work is still evolving.
The information is dense, so there is no need to read it in sequence. Start with any section that interests you; some concepts will continue to be revised through practice.
The Chinese version is the primary source of truth; English content may lag behind as the site evolves.
Intelligence and its environment need to be designed together and co-evolve.
What an intelligent system can accomplish depends on its own capabilities and on the environment in which it works. My research and practice explore their relationship: how to build intelligent systems suited to different ecological niches, and how to reshape business and digital environments so that people, agents, and existing systems can work together.
Read: Intelligent systems, ecological niches, and agent networksIntelligence and its environment need to be designed together and co-evolve. Starting from enterprise business, I study how systems represent reality and how people, agents, and existing systems work together within it. I put these methods into practice in manufacturing and other settings, and use this perspective to understand agents, harnesses, and world models.
General Proposition
From digital representations of complex systems to digital environments for human–AI collaboration
Digitalization brings real objects, relationships, states, and rules into the Digital World. As these representations are continuously updated to reflect reality and support what-if analysis, action, and feedback, they gradually develop into digital twins: environments in which people and intelligent systems work together.
Digitalization should ultimately create a digital twin that can participate in real-world operations. Improvements to the environment change what intelligent systems can accomplish, while their actions drive further updates to the environment. The two thus co-evolve.
Research & Applications
From enterprise operations to personal decisions, use these methods to clarify real problems, design working systems, and test and improve them through practice.
Industry OS
Software as a complex system
Personal decision-making
Questions & Judgments
Some questions about putting AI into practice need to be answered again
Common topics
Underlying issue
Related article
Underlying issue
Tasks have their own origins and goals. Understanding why they exist provides a basis for deciding whether to accelerate, redesign, or eliminate them, and which part of the work an agent should undertake.
Common topics
Why do businesses still rely on group chats and shared Excel files despite having so many systems?
Underlying issue
Systems already cover much of the business, but people still fill in the business logic across them. Understanding this manual work helps identify where systems need improvement and where agents are appropriate.
Related articleRepresentation
Common topics
What should a harness actually do?
Underlying issue
A harness does more than assemble prompts, context, and tools for a model. More importantly, it lets an agent reliably perceive the business environment, understand the current state, and call the right capabilities to act.
Related articleAgent harness / AI & digitalization
Common topics
How can an agent truly make decisions instead of merely executing a workflow?
Underlying issue
Executing predefined steps and judging among goals, constraints, unknown information, and multiple choices are two different problems. Real decision-making requires explicitly organizing the factors, constraints, and feasible paths that shape a choice.
Related articleDecision Space
Common topics
Why do AI agents stall when put to work?
Underlying issue
Agents need more than stronger reasoning and execution capabilities. Complexity left unresolved in business definitions, product design, and technical implementation keeps becoming a burden on their context management, reasoning, and execution. Putting agents to work also requires improving the environments they operate in.
Related articleWhy do AI agents stall when put to work?
Digitalization & AI Implementation
Enterprise AI implementation starts with rebuilding the digital world
Understandable → computable → actionable.
Digitalization is the infrastructure for AI. Earlier digital systems were designed primarily for people. They relied on people to interpret information, supply missing rules, coordinate processes, and act as the “glue” between systems. As AI begins to participate in business operations, complexity that these systems have not represented and managed becomes an additional burden on agents’ understanding, reasoning, and execution.
Much agent engineering today aims to improve “autonomous driving”: helping AI adapt to complex external environments. Yet an enterprise’s digital world still resembles separate maps drawn to different standards. Objects, states, and rules are scattered across systems, without a shared foundation for navigating across them.
Every domain will need a continuously updated Business Atlas that connects fragmented business knowledge and systems, enabling people and AI to understand the business, assess problems, and act within the same digital world.
Rebuilding the enterprise digital world means addressing four layers together
The four layers depend on one another, while each presents distinct problems of its own.
Why enterprise AI must start with the business01Business · define the world
Turn real business into a computable world
A subway map does not show every building. It retains only the stations, lines, connections, and transfer rules useful for riding the subway.
An enterprise is the same. Its stable backbone consists of the objects, relationships, events, states, and rules within the business. A Business Atlas is a compressed representation of these structures.
02Organization · define responsibility and division of labor
When silicon intelligence enters an organization, work and responsibility must be redefined
Collaboration among multiple actors—human or agent—is fundamentally a matter of designing the division of labor well. Those rules must resolve into concrete business objects, actions, and responsibilities rather than stopping at departments, roles, and permission settings.
03Systems · represent and maintain the world
Product requirements define the gap between the business and its digital twin
Business keeps changing, while systems usually represent only part of it.
The work of enterprise product design is to identify what the system still fails to express, connect, or execute correctly in the business world, then progressively close that gap.
04Agents · use the environment
Let agents navigate enterprise environments with the Business Atlas
Every time an agent acts, it needs business semantics, live state, rules, and executable interfaces matched to the current situation.
Together, these form the agent’s real working environment—not unlimited context.
Enterprise OS | Organize the whole business across software boundaries
Business does not follow the boundaries of software such as OA, ERP, or MES. Enterprise OS connects existing systems through shared business objects, relationships, states, and rules, enabling people and AI to continue one another’s work in the same digital environment without repeatedly reconstructing the whole picture across systems.
Industry OS | Extend collaboration beyond the enterprise
Enterprises share product definitions, transaction relationships, industry rules, and collaboration needs. When these structures can be reused across companies, the digital environment can extend along the value chain, supporting ongoing collaboration among people, AI, and different enterprises.
Read more
Core questions
- Data & semantics
- What does “poor data quality” actually mean in an enterprise?
Case Studies & Reflections
Cases and methodological reflections grounded in real business problems.
Why did a project to generate BOMs from drawings begin by redefining the product?
The customer initially wanted AI to generate BOMs automatically from drawings. After reassessing the underlying need, we first clarified the product definitions and parameter-to-material configuration rules for six standard models, then built a small tool and put it into use. The technical department reported an improvement of approximately 25%–30% in efficiency for this work.
View caseWhy are equipment records still hard for people and AI to use as more fields are added?
Sales, asset management, maintenance, and quality teams all keep adding fields to the same equipment table. Each request makes sense in isolation, yet the page grows longer and the information becomes harder to use. The problem is not merely the number of fields: different business domains, perspectives, and purposes have been compressed into one representation. Once those perspectives are separated, equipment information can form scenario-specific summaries and even generate different working interfaces dynamically.
View caseWhy must ‘approval’ be unpacked before an agent can automate it?
Many OA systems treat form records, notifications, object creation and change, and decisions that truly require judgment as the same thing: approval. Approvals and notifications then multiply, while people still open the system without knowing what they actually need to complete. Before an agent can participate, the first step is not to make it run the workflow faster, but to distinguish what these actions really are.
View case