From AI Models to AI Architectures

🧭 Dojo Compass

Module: Finance, Risk Management and Long-Term Resilience

Focus Area: Technology, AI and Future Readiness

Learning Format: Article

Artificial intelligence is evolving from a collection of tools and models into a diverse ecosystem of computational architectures with fundamentally different capabilities. As these technologies develop, companies will need to reconsider how they approach AI adoption.

This article examines the transition from selecting AI functionality and individual models to designing broader AI architectures around business objectives. It explores the principal emerging architectural approaches, their respective advantages and limitations, and the strategic implications of adopting a more comprehensive perspective on AI integration.

🎯 Key Challenge

Much of the current discussion about corporate AI adoption focuses on relatively narrow questions:

• Which AI tools should we use?
• Which prompts produce the best results?
• Which model is most capable?
• Which business processes can we automate?
• How can we incorporate AI into our existing systems?

These are legitimate questions, but they increasingly risk restricting how companies think about the possibilities of artificial intelligence.

AI is not a single technology with a progressively improving set of capabilities. It is an expanding field encompassing different approaches to learning, representing information, understanding environments, predicting outcomes, reasoning, planning and taking action.

The distinction matters because different architectures can produce fundamentally different solutions to the same business problem.

Consider a company seeking to improve its strategic planning.

An approach based primarily on large language models (LLMs) might involve using AI to research markets, analyse competitors, evaluate strategic alternatives, generate business plans and support management decision-making.

An approach incorporating a state-based world model might instead seek to represent the company’s operating environment, identify the variables and relationships that determine its performance, learn how those variables change over time and simulate the potential consequences of different strategic decisions.

These approaches are not mutually exclusive. Indeed, combining them might produce a substantially more powerful strategic capability.

However, they begin with different conceptions of what the problem is and what intelligence is required to address it.

The same distinction applies across virtually every business function.

A manufacturer might use an LLM to analyse production reports, a digital twin to represent its manufacturing processes, and a world model to predict how the factory might respond to changing conditions. An agentic system might then use these capabilities to recommend or initiate operational adjustments.

The question is no longer simply which model can perform a particular task. It is which combination of technologies, representations and computational processes can best achieve the desired outcome.

This introduces a broader strategic challenge.

Companies need to move beyond thinking about AI as a collection of tools and start thinking about AI as an architectural capability that can reshape how their businesses operate.

🥋 Dojo Solution

Design AI architectures around business objectives, rather than designing business applications around whichever AI models happen to be available.

This requires a change in perspective, progressing through three levels of AI consideration.

1. From functionality to models

At the first level, companies identify useful AI functionality.

This might involve summarization, translation, classification, information extraction, image recognition, code generation or conversational assistance.

The next consideration is which model is best suited to the task.

An LLM might be appropriate for interpreting natural language, while a vision model might be better suited to analysing images. A smaller specialized model might deliver faster, cheaper and more reliable results for a repetitive, narrowly defined task.

This is an important level of decision-making, but it remains largely focused on individual capabilities.

2. From models to architectures

The second level considers how different models and computational approaches can be combined.

A business application might incorporate:

• An LLM for interpretation and reasoning.
• A retrieval system for accessing company knowledge.
• A predictive model for forecasting.
• A knowledge graph for representing relationships.
• An agent for coordinating activities.
• Conventional software for executing deterministic processes.

The resulting system is more than a model connected to an application. It is an architecture in which different components perform different functions.

This introduces opportunities for greater specialization, flexibility and reliability.

3. From architectures to business transformation

The third level is the most strategically important.

Rather than asking how AI can improve an existing business process, companies ask what operating model, organizational capability or commercial opportunity becomes possible through the availability of new AI architectures.

For example:

• Could a company move from periodic forecasting to continuous predictive management?
• Could it move from responding to operational problems to anticipating and preventing them?
• Could it develop products that interact intelligently with their operating environments?
• Could it replace sequential human workflows with coordinated networks of specialized AI agents?
• Could it simulate strategic alternatives before committing significant capital?
• Could it develop entirely new services based on predictive intelligence that was previously unavailable?

This is where AI adoption becomes a strategic question rather than simply a technology investment.

Understanding the principal AI architectural approaches

To make these decisions intelligently, management needs a working understanding of the principal approaches currently available or emerging.

The following categories are not mutually exclusive. Some describe the underlying architecture of a model; others describe how models are organized into larger systems.

1. Large Language Models (LLMs)

LLMs are trained on enormous volumes of information to predict and generate sequences of tokens. Their ability to recognize patterns in language and other representations makes them useful for understanding instructions, synthesizing information, generating content and performing a wide range of reasoning tasks.

Advantages

• Exceptional flexibility across a broad range of knowledge-intensive tasks.
• Natural language interfaces that make sophisticated capabilities accessible.
• Ability to synthesize information from different sources.
• Strong capabilities in explanation, summarization, analysis and code generation.
• Can coordinate tools and other computational resources.

Disadvantages

• Can produce plausible but inaccurate information.
• Do not automatically possess a reliable, continuously updated representation of a company’s operating environment.
• May struggle with precise numerical, temporal and causal reasoning.
• Can be expensive to operate at scale.
• Knowledge may be incomplete, outdated or insufficiently grounded in company-specific information.

Strategic applications: Research, management support, customer interaction, knowledge work, software development and business process assistance.

LLMs are likely to remain a central component of enterprise AI, but they should not be confused with the entirety of AI.

2. Retrieval-Augmented Generation (RAG) and Knowledge-Based Architectures

RAG combines generative models with external information retrieval. Rather than relying exclusively on information embedded in a model’s training, the system retrieves relevant material from external knowledge sources and makes it available when generating a response.

Knowledge graphs and semantic models provide additional ways of representing entities, concepts, relationships and organizational knowledge.

Advantages

• Allows AI to work with current and company-specific information.
• Reduces dependence on retraining large models whenever information changes.
• Can provide traceability to underlying documents and information sources.
• Enables the integration of knowledge distributed across different systems.
• Knowledge graphs can represent relationships that simple document retrieval might overlook.

Disadvantages

• Retrieval quality depends on the quality and organization of underlying information.
• Poorly structured or incomplete knowledge can produce misleading results.
• Retrieval does not automatically establish whether information is accurate.
• Requires ongoing maintenance, access controls and information governance.
• A knowledge representation describes what is known but does not necessarily explain how the environment behaves.

Strategic applications: Enterprise knowledge management, compliance, research, customer intelligence, institutional memory and decision support.

An important distinction is that representing knowledge and predicting the behaviour of a system are different capabilities.

3. Predictive Machine Learning Models

Predictive machine learning encompasses models trained to identify patterns in data and estimate future outcomes or classify observations.

Examples include forecasting models, classification systems, anomaly detection and statistical learning approaches.

Unlike general-purpose LLMs, these models are often designed around relatively specific prediction objectives.

Advantages

• Can perform very well on clearly defined prediction tasks.
• Often more computationally efficient than general-purpose models.
• Can be evaluated against measurable outcomes.
• Useful for identifying patterns that may not be readily apparent to human analysts.
• Can be integrated into existing operational systems.

Disadvantages

• Performance depends heavily on data quality and relevance.
• Models may struggle when underlying conditions change.
• Often require significant domain-specific data and training.
• Predictions do not necessarily explain underlying causes.
• A model trained to predict a particular outcome may have limited usefulness outside its intended application.

Strategic applications: Demand forecasting, credit assessment, fraud detection, customer churn, predictive maintenance and financial planning.

These approaches remain important even as more sophisticated generative and agentic architectures emerge.

4. Digital Twins and Executable Models

A digital twin is a digital representation of a real-world asset, process or system, typically connected to operational data and capable of representing its behaviour according to known rules, relationships and constraints.

A factory digital twin, for example, might represent machinery, production lines, inventory, operating schedules, resource constraints and production processes.

It can be used to monitor performance, test scenarios and understand the consequences of changing operating conditions.

Advantages

• Provides a structured representation of complex operating environments.
• Can incorporate established engineering knowledge and physical constraints.
• Allows companies to test scenarios without necessarily intervening in real operations.
• Can connect operational data with management and planning systems.
• Offers a foundation for more sophisticated simulation and predictive capabilities.

Disadvantages

• Requires substantial work to construct and maintain.
• Depends on the accuracy of its underlying representations and rules.
• May not capture unexpected behaviours or previously unidentified relationships.
• Can become outdated if the real-world system changes.
• Modelling complex environments can be expensive and technically demanding.

Strategic applications: Manufacturing, infrastructure, logistics, energy, construction, supply chains and asset management.

A digital twin is particularly valuable when the company understands the underlying system sufficiently well to represent its processes and constraints.

However, representing what is already understood is different from learning what is not yet understood.

5. World Models

<text color=”secondary” weight=”medium” size=”xs”>LEARNED DYNAMICS, STATE PREDICTION AND FUTURE SCENARIOS</text>

World models represent one of the most interesting emerging developments in AI.

A world model seeks to learn how an environment behaves and evolves. Rather than simply generating a plausible description of the world, it develops an internal predictive representation of relevant states, relationships and dynamics.

In simplified terms, an LLM might estimate what words or concepts are likely to follow a given sequence of information. A world model seeks to estimate how the state of an environment might change under particular conditions or actions.

A simplified representation is:

LLM:

P(next token | preceding tokens)

World model:

P(future state | current state, action, relevant history)

The distinction is important.

A company might use a world model to examine how changes in production schedules, machine availability, material supply or environmental conditions could affect future factory performance.

More advanced systems could evaluate different interventions and support the selection of actions that produce desirable outcomes.

Some emerging approaches, including Joint Embedding Predictive Architecture (JEPA), seek to predict representations of future states rather than reconstructing every detail of the underlying observations. Other approaches learn dynamics through model-based reinforcement learning, robotics and predictive simulation.

Advantages

• Can learn patterns and dynamics beyond explicitly programmed rules.
• Offers the possibility of modelling complex, changing environments.
• Can support scenario evaluation and planning.
• May identify previously unrecognized relationships in operational data.
• Can provide a foundation for more sophisticated predictive and autonomous systems.

Disadvantages

• Requires substantial relevant observational data and appropriate training.
• Learned representations can be difficult to interpret.
• Predictions can become unreliable outside the conditions represented in training.
• Errors may accumulate over longer prediction horizons.
• Prediction does not automatically establish causation.
• Reliable evaluation, uncertainty management and safety remain significant challenges.

Strategic applications: Advanced manufacturing, robotics, autonomous systems, logistics, infrastructure, resource management and complex operational planning.

An important qualification is that world models are still an emerging and diverse research area. Not every system described as a world model provides reliable simulation or decision-making capabilities.

Nevertheless, the potential strategic significance is considerable. A company that can develop a useful learned representation of how its operating environment behaves may acquire capabilities that go well beyond conventional reporting and forecasting.

6. Agentic and Multi-Agent Architectures

Agentic AI systems extend the capabilities of foundation models by providing them with objectives, memory, tools and the ability to execute multiple steps towards a desired outcome.

A multi-agent architecture distributes responsibilities among multiple agents, potentially allowing different agents to specialize in research, analysis, planning, verification or execution.

For example, a company’s procurement system might incorporate agents responsible for monitoring suppliers, forecasting requirements, evaluating prices, identifying risks and coordinating purchasing activities.

Advantages

• Can automate complex workflows rather than isolated tasks.
• Enables coordination across different systems and business functions.
• Allows specialized agents to focus on different responsibilities.
• Can operate continuously and respond to changing information.
• Creates opportunities for new forms of organizational productivity.

Disadvantages

• Multiple agents can compound errors rather than eliminate them.
• Coordination and communication introduce additional complexity.
• Uncontrolled autonomy can create operational and security risks.
• Tool access and decision-making permissions require careful governance.
• Performance can be difficult to evaluate across long-running tasks.

Strategic applications: Procurement, customer service, research, finance, software development, compliance and operational management.

The important distinction is between an AI assistant that provides information and an AI system that can coordinate and execute work.

Agentic architectures move AI closer to the operational core of a business.

7. Multimodal and Embodied AI Architectures

Multimodal AI systems work across different forms of information, including text, images, video, audio and sensor data.

Embodied AI extends this further by connecting intelligence to systems that interact with the physical world, such as robots, autonomous vehicles and industrial equipment.

These approaches are important because business environments are not composed exclusively of documents and language.

A manufacturing operation generates visual, acoustic, spatial, numerical and temporal information. A logistics operation involves physical objects, locations, movement and changing environmental conditions.

Advantages

• Allows information from different sources to be combined.
• Enables richer understanding of physical and operational environments.
• Creates opportunities for more natural human-machine interaction.
• Supports robotics, inspection and autonomous operations.
• Can connect digital decision-making with physical execution.

Disadvantages

• Integrating different modalities can be computationally demanding.
• Different sensors and data sources may have different levels of accuracy and reliability.
• Physical actions introduce safety and control requirements.
• Real-world conditions can be unpredictable.
• Systems must often operate within strict latency and hardware constraints.

Strategic applications: Robotics, quality control, logistics, autonomous equipment, industrial inspection, healthcare and spatial computing.

These architectures will become particularly important as AI moves beyond the digital office and into physical operations.

The case for hybrid architectures

The preceding approaches should not be viewed as competing alternatives from which a company must select just one.

In many circumstances, their greatest value will emerge through combination.

Consider again the manufacturing example.

A comprehensive architecture might incorporate:

1. Physical data from machinery, sensors and production systems.
2. A semantic model representing equipment, materials, processes and relationships.
3. A digital twin representing known operating rules and constraints.
4. A world model learning how the factory behaves under different conditions.
5. An LLM providing interpretation, explanation and higher-level reasoning.
6. AI agents coordinating analysis, planning and authorized actions.
7. Conventional software and control systems executing deterministic processes.
8. A visualization and interaction layer allowing people to understand and manage the system.

Each component performs a different function.

The semantic model provides structured knowledge. The digital twin represents known behaviour. The world model learns predictive dynamics. The LLM provides language-based reasoning and coordination. Agents organize work and interact with external systems. Deterministic software and controllers provide reliable execution.

Importantly, the architecture should also support feedback. New observations can improve predictive models, while newly discovered relationships can inform and update the company’s knowledge representations.

This is a fundamentally different conception of AI from simply connecting a language model to a collection of company documents.

The architecture is the strategic choice

The underlying lesson is that the choice of AI architecture can influence the nature of the business capability being developed.

Consider the following examples:

Business objective	Narrow AI approach	Broader architectural approach
Improve strategy	LLM-assisted research and planning	Knowledge models, predictive models, scenario simulation and reasoning agents
Improve manufacturing	AI analysis of production reports	Digital twin, world model, predictive maintenance and controlled operational agents
Improve customer relationships	Conversational AI	Customer knowledge graph, multimodal intelligence and coordinated service agents
Improve supply chains	AI-assisted forecasting	Predictive models, supplier networks, simulation and autonomous coordination
Improve organizational learning	AI-powered document search	Semantic knowledge systems, organizational memory and agents supporting knowledge discovery

The broader approach may require more investment and technical sophistication. It is not necessarily the right answer for every application.

But it can also create capabilities that cannot be achieved simply by making an existing AI tool more effective.

🏗️ Putting It into Practice

Companies can begin adopting this broader approach through a structured process.

Step 1: Start with the business objective

Avoid beginning with a particular AI model, product or technology.

Instead, identify an important business objective or operational challenge.

For example:

• Reduce production downtime.
• Improve strategic forecasting.
• Increase customer retention.
• Improve capital allocation.
• Reduce procurement costs.
• Accelerate product development.

Define what the company wants to achieve and how success will be measured.

Step 2: Examine the problem through different AI perspectives

Consider how different architectural approaches might address the same objective.

Ask:

• Would an LLM provide useful analytical or reasoning capabilities?
• Is the problem primarily one of accessing and organizing information?
• Would a predictive model provide greater value?
• Does the problem involve a complex system whose behaviour needs to be represented?
• Would a digital twin or world model provide useful scenario-testing capabilities?
• Could agents coordinate activities that currently require substantial human intervention?
• Does the solution require multimodal or physical-world interaction?

The purpose is to avoid prematurely restricting the solution to the capabilities of a particular technology.

Step 3: Identify the required architecture

Determine which components are necessary and how they should interact.

For example, a predictive maintenance solution might require:

• Operational sensor data.
• A structured representation of equipment.
• A model that predicts equipment failure.
• An LLM that explains findings and assists maintenance personnel.
• An agent that coordinates inspections and maintenance scheduling.
• Existing operational software to execute approved actions.

Not every component needs to be sophisticated or AI-driven.

In fact, conventional software may be preferable where deterministic execution is required.

The objective is to allocate each function to the most appropriate component.

Step 4: Assess data, infrastructure and organizational readiness

A sophisticated architecture is only as useful as the environment supporting it.

Evaluate:

• Availability and quality of operational data.
• Ability to integrate information across existing systems.
• Data ownership, access rights and governance.
• Computational infrastructure and operating costs.
• Availability of technical expertise.
• Security, privacy and regulatory requirements.
• Ability to monitor, test and validate outputs and decisions.

Particular attention should be given to the distinction between explicit knowledge and knowledge that must be learned from observed behaviour.

A company may have excellent records of what happened without having sufficient information to understand why it happened or predict what will happen next.

Step 5: Develop a limited but meaningful prototype

Select a business problem where the proposed architecture can be tested against measurable outcomes.

Avoid attempting to build an entire enterprise AI architecture at once.

Develop a limited implementation that demonstrates the essential components and their interaction.

Test:

• Whether the architecture produces better results than the existing approach.
• Whether its predictions and recommendations are sufficiently reliable.
• Whether different components communicate effectively.
• Whether the system performs adequately under changing conditions.
• Whether the economic benefits justify the additional complexity.

For world models and digital twins, this should include testing against real observed behaviour, not simply evaluating whether simulated outputs appear plausible.

Step 6: Reconsider the operating model

Once an architecture demonstrates value, consider its broader organizational implications.

Which activities can be automated? Which decisions can be delegated? Which capabilities should remain under human control? What new responsibilities emerge?

This may require changes to:

• Organizational structures and responsibilities.
• Information management and data ownership.
• Decision-making authority.
• Employee skills and training.
• Risk management and internal controls.
• Technology governance and investment priorities.

AI architecture is not simply an IT matter. It can affect how work is organized and how the company creates value.

Step 7: Treat AI architecture as a continuing strategic capability

AI architectures will continue to evolve.

New models, techniques and computational approaches will emerge. The economics of different architectures will change. Capabilities that are currently experimental may become commercially practical.

Companies should therefore avoid designing themselves around a single model or technology provider.

Develop an architecture that is modular, measurable and capable of incorporating new capabilities as they become relevant.

Periodically reassess whether the existing architecture remains appropriate for the company’s objectives.

The ultimate objective is not to build the most sophisticated AI system. It is to maintain the ability to select and combine the most appropriate forms of intelligence as the business and its environment evolve.

📌 Key Takeaways

• Think beyond functionality: Selecting the right prompts, tools and model features is only the beginning of AI adoption.
• Think beyond individual models: Different models and computational approaches offer fundamentally different capabilities.
• Think architecturally: The strategic question is increasingly how different forms of intelligence should be combined to achieve business objectives.
• Understand the distinctions: LLMs, knowledge systems, digital twins, predictive models, world models and agents perform different but potentially complementary functions.
• Recognize the potential of world models: Learning how an environment behaves and predicting how it might evolve could create capabilities beyond conventional information processing.
• Embrace hybrid architectures: The greatest value may emerge from combining generative, predictive, agentic, simulation-based and deterministic systems.
• Consider organizational implications: AI architecture can change workflows, decision-making, responsibilities and the underlying operating model.
• Let business objectives determine technology choices: The architecture should follow the problem, not the other way around.
• Maintain architectural flexibility: AI is evolving too quickly for companies to assume that today’s preferred model or architecture will remain optimal.

🌿 Reflection

Imagine that your company is starting again today, with access to the AI capabilities that are becoming available.

Would you design its operating model around the processes, systems and organizational structures that currently exist? Or would you first consider what becomes possible when different forms of artificial intelligence can represent knowledge, understand environments, predict outcomes, simulate alternatives and coordinate actions?

Perhaps the more important question is whether your company is using AI to make its existing business more efficient, or whether it is beginning to discover entirely new ways of creating value.

There is a significant difference between the two.

The companies that ultimately derive the greatest strategic benefit from AI may not be those that adopt the most powerful individual models, but those that develop the clearest understanding of how different forms of intelligence can be combined to achieve their objectives.

⚔️ Dojo Mission

Choose one important business objective or operational challenge within your company.

Conduct a short AI Architecture Review:

1. Identify how you might address the challenge using conventional LLM functionality.
2. Consider whether predictive models, knowledge systems, digital twins, world models or agents could provide additional capabilities.
3. Sketch a potential architecture showing how the different components might work together.
4. Identify one limited application that could be developed and tested within the next 30 days.

Do not begin by asking which AI product you should purchase.

Begin by asking:

What kind of intelligence does this business problem actually require, and what architecture would best provide it?

That is the beginning of thinking about AI strategically.