Understanding How AI Thinks (and Where It Doesn’t) — Part 2 From Reasoning to Cognition

Introduction: What Was Missing

Quick recap (from Part 1): We saw that LLMs are very good at understanding meaning (semantics) and even reasoning step‑by‑step, but they still don’t decide or act on their own.

With that context in mind, this part continues the story by asking the next natural question: if understanding and reasoning aren’t enough, what actually enables intelligent behavior?

In Part 1, I shared understanding and reasoning alone do not decide or act. That realization naturally raised a follow-up question — if reasoning isn’t enough, what actually enables intelligent behavior?

In this post I wanted to share about what I learnt about the missing layer: cognitive capability, and how AI agents introduce it through architecture rather than model intelligence.

Cognitive Capability (Knowing What to Do Next)

This question brought me to the concept of cognitive capability — in simple terms, the ability of a system to decide and act, not just explain or understand.

Unlike semantic understanding or reasoning, cognitive capability is not about explaining information — it is about using information.

In simple terms:

It answers the question: “What should I do with this information?”

Cognitive capability includes:

Setting goals
Making decisions
Taking actions
Learning from results

Humans do this seamlessly, often without realizing it.

AI systems, however, do not gain this capability just by becoming better at language or reasoning. They require a different kind of design.

This distinction made the gap between humans and AI much clearer — and it naturally pointed to the concept of agents as the missing architectural layer.

AI Agents (Adding a Brain Around the Model)

Once cognition became the focus, AI agents entered the picture naturally. At this point in my thinking, agents stopped feeling like a buzzword and started feeling like a design necessity.

An AI agent is not just a smarter model — it is a system where a model is embedded inside a control loop.

That loop:

Observes information
Decides what to do
Takes action using tools or systems
Checks the outcome
Adjusts its next step

In this arrangement, roles become clear:

The LLM understands language
The agent owns decisions and actions

This is when I was able to realized the importance of various concepts I kept reading and hearing through podcasts played an important role in building full AI system. I was able to understand why adding tools, memory, and feedback suddenly makes AI systems feel more capable, not because the model changed, but because cognition was introduced at the system level.

With this in mind, I again started wondering how closely this maps to human thinking — and whether humans use a similar separation between fast understanding and deliberate control.

System 1 and System 2 (How Humans Think)

To make sense of this comparison, it helped to borrow a well-known model from psychology that explains how humans think at different speeds.

Psychologist Daniel Kahneman described two ways humans think:

System 1 – Fast Thinking

Automatic
Intuitive
Pattern-based

Example: Instantly recognizing a familiar face.

System 2 – Slow Thinking

Deliberate
Logical
Effortful

Example: Carefully solving a math problem.

Mapping This to AI

LLMs behave like System 1 — fast, fluent, intuitive
Agents behave like System 2 — slow, deliberate, controlling decisions and actions

This mapping helped me clarify why agents feel qualitatively different from standalone models — they introduce control, not just intelligence. That control is what allows systems to pause, decide, and act intentionally.

Conclusion: In AI Systems Cognition Is Architectural

Bringing these ideas together helped solidify the story so far: better models improve understanding, but better architecture enables cognition.

This part reinforced a key insight for me: cognition does not emerge automatically from better reasoning. It emerges from architecture — from systems that can observe, decide, act, and learn.

In the final part, I will share my understanding around why humans still outperform AI in ambiguity, where agents fall short of human cognition, and why this does not diminish the value of today’s AI systems.

Author’s Note: AI-assisted writing tools were used to support the creation of this post. All concepts, perspectives, and the underlying thought process originate from me; the AI served only as a drafting and refinement aid.

The ERP Awakening: Surviving Day 1: The Truth of GenAI Implementation

Introduction: From Vision to Reality

In Post 1: The ERP Awakening, the journey started with the promise of moving from static records to actionable intelligence. That vision is inspiring, but the real test comes on Day 1—when the system meets the real world. This post explores what it takes to move from vision to execution, focusing on the practical data challenges and the first steps in implementing GenAI in an enterprise context.

Context: The Demo Room vs. The Real World

The journey often starts in a demo room. The screen glows, the answers are instant, and the optimism is contagious. This is “Day 0”—the promise of transformation. But the real world is not a demo. When the system is switched on for actual business, the cracks start to show. Data is scattered, processes are inconsistent, and the system struggles to deliver the same clarity seen in the demo. The real work begins here, where vision meets reality.

Problem: Why Day 1 Hurts—The Data Challenge

Most business systems were built to keep records, not to explain them. Over the years, notes piled up, customer names got duplicated, and old process documents stuck around. When GenAI is introduced, it tries to make sense of all this information. The result can be confusion: the system might give an answer that sounds right but is built on mismatched records or outdated information. The real problem isn’t just “messy” data—it’s that the data was never organized for analysis and learning.

Root Cause: Data Standardization and Readiness for GenAI

To get real answers, the data must be organized and standardized. This means:

Merging duplicate records (e.g., “Acme Corp” and “Acme Corporation” become one)
Retiring old process documents that no longer apply
Making sure important details aren’t buried in free-text notes or scattered emails

If these basics are skipped, GenAI will only repeat the confusion. Standardizing and aligning information is the first real step toward clarity and reliable automation.

Insight: What GenAI Actually Does with Enterprise Data

GenAI does not fix data inconsistencies; it surfaces and reflects them. When data is fragmented or non-standardized, GenAI will generate outputs that mirror these limitations. The system is only as good as the information it can access and understand. For GenAI to provide useful insights, the underlying data must be structured, current, and accessible.

Solution: Practical Approaches to GenAI Implementation

There are three practical ways to start implementing GenAI in an enterprise, each matching a stage of maturity:

Stage 1: The Chat Window (Sidecar)

What it is: A simple chat box that sits on top of the system, letting users ask questions about business data. It is best for getting started quickly, answering simple questions, and testing the waters.
Limits: Can only access surface-level information—no deep dives into complex business logic or historical context.

Stage 2: The Built-in Assistant (Platform Native)

What it is: GenAI features built into the ERP platform, with access to more business context and data relationships. Answers are richer and more connected to the business.
Best for: Organizations ready to move beyond basics, using the system’s built-in tools for deeper insights.
Limits: Follows the platform’s rules—custom requests or unique business logic may be out of reach.

Stage 3: The Custom Knowledge Layer (RAG Pipeline)

What it is: A custom solution that connects GenAI to all business data, documents, and records, enabling complex questions and advanced use cases.
Best for: Enterprises with unique needs, lots of documents, or special business rules.
Limits: Building and maintaining this solution takes time, effort, and ongoing care.

Implications: Trust, Transparency, and Change Management

No matter which approach is chosen, trust is built by showing the work. Every answer should come with a source or reference. If the answer isn’t certain, the system should say so. And for important decisions, a human should always have the final say. GenAI works best when everyone can see how the answer was found and understands its limitations.

Conclusion: Day 1 is Just the Beginning

Moving from vision to reality is not a one-day project. The first step is organizing and standardizing the data, then choosing the right approach for GenAI, and finally connecting all the necessary information. The journey is about making the system work for the business—clear, transparent, and ready for the next question. Along the way, each step introduces new concepts and practical learning about how GenAI can be implemented and trusted in the enterprise.

How We Help Enterprises @ 1CloudHub

At 1CloudHub, we help enterprises in adopting GenAI to transform ERP platforms into Systems of Intelligence. We help enterprises to navigate the journey from demo room optimism to Day 1 reality. We work with you to assess your data readiness, choose the right GenAI approach for your business, and build the governance frameworks that turn experimental pilots into sustainable competitive advantages. Whether you need to assess data readiness, platform selection, or a custom RAG solution, we’ve guided organizations through each phase to unlock real value from GenAI in their ERP environments.

The Applied AI Thoughts for Realization Blog Post 3

The Impact Layers — Why AI Models Alone Don’t Deliver Value

Introduction

This is the third post in the Applied AI Thoughts for Realization series.

In the first post, Why AI Feels Overwhelming, we tackled the problem of “AI Fatigue” and the trap of Tactical Thinking—chasing the latest tools without a plan. We argued for a shift to Structural Thinking, focusing on the architecture of problems rather than just the features of models.

In the second post, A Simple Mental Model — 4 Pillars, we established the horizontal dimension of our mental model. We categorized the AI landscape into 4 Domain Pillars—Consumer, Enterprise, Science, and Physical AI—and used the “Engine vs. Vehicle” analogy to show why a “Sports Car” strategy (Consumer) fails when you need a “Cargo Train” solution (Enterprise).

With the understanding of the 4 Pillars of different AI ecosystem and how the application of AI varies across the pillars its critical to understand what is required to deliver value in ech of the pillars. This brings us to the vertical dimension of our framework: The Impact Layers which play critical role in delivering value for each pillar.

Why a Model is not Everything in a AI Solution?

There is often a disconnect when we try to adopt AI. On one hand, we see headlines about models acing exams and writing code. It makes us feel model is what you need to role out a AI driven solution. On the other hand, when Enterprises or Consumers try to start deploying AI based solutions, they end up realizing its much more than just having a model. Enterprises have to address multiple challenges since they face multiple challenges like:

The solution is not responsive and impacts user experience.
Does not produce consistent or reliable results.
It’s not quite fast enough, or it produces results that need a second look.
Integration with existing systems is complex and time-consuming.
The cost of running the solution at scale becomes prohibitively expensive.
It works in demos but fails on real-world, messy data.
Users don’t trust the output enough to act on it without verification.
Models occasionally “hallucinate” or make confident errors.

The question then arises Why is there such a gap between the Intelligence we hear and read about around the Model and realizing the actual Capability we like to experience?

The answer lies in understanding that a “Model” is not the final “Product” or “Solution”.

A model is just raw potential—like a powerful engine sitting on a factory floor. To turn that potential into actual value, it is dependent on several layers of translation. It needs to be hosted on hardware, connected to tools, wrapped in an interface, and integrated into a workflow.

If any one of those layers is weak, the entire experience fails.

The “Iceberg” Theory of AI

A helpful way to visualize this is the Iceberg Theory of AI.

When you interact with an AI application—whether it’s a chatbot, a recommendation engine, or a robot—you are only seeing the tip of the iceberg.

Above the Water (Visible): The Application Layer. This is the user interface, the buttons, and the response time. This is what we judge.
Below the Water (Invisible): The massive infrastructure that supports that tip. This includes the Agents (logic), the Models (intelligence), and the Hardware (compute).

Most of the hype focuses on the “Model” layer deep underwater. But most of the failure happens in the layers between the model and the user. To understand why an AI project succeeds or fails, we need to look below the surface and examine the 5 Layers of AI Impact.

The 5 Layers of AI Impact

Progress in AI doesn’t happen all at once. It moves up this stack, layer by layer.

Layer 1: Hardware (The Foundation)

This is the physical reality of AI. It includes the GPUs (chips) that train models, the data centers that host them, and the edge devices (phones, robots) that run them.

Why it matters: Hardware dictates feasibility. You might have a brilliant AI model, but if it requires $10,000 of compute per hour to run, it cannot be a consumer product. If it takes 5 seconds to respond, it cannot be a self-driving car.
The Constraint: Cost, Energy, and Latency.

Layer 2: Models (The Intelligence)

This is what we typically call “AI.” It includes Large Language Models (LLMs), diffusion models (images), and predictive models. This layer provides the raw reasoning and pattern-matching capability.

Why it matters: Models dictate potential. A smarter model can solve harder problems.
The Constraint: Context Window (memory), Hallucination (accuracy), and Reasoning capability.

Layer 3: Agents & Tools (The Orchestration)

This is the bridge between thought and action. A model can only output text; an Agent can use that text to call a tool—like searching the web, querying a database, or clicking a button.

Why it matters: Agents dictate utility. Without this layer, AI is just a chatbot. With this layer, AI becomes a coworker that can book flights, write code to disk, or control a robot arm.
The Constraint: Reliability. If an agent gets confused and clicks the wrong button, it causes chaos.

Layer 4: Applications (The Interface)

This is the software layer where the human meets the machine. It includes the UI/UX, the workflow integration, and the “vibe” of the product.

Why it matters: Applications dictate adoption. A powerful agent wrapped in a confusing interface will be ignored. The best AI applications often hide the AI completely (e.g., Netflix recommendations).
The Constraint: Friction and Trust. Users must feel in control.

Layer 5: Impact (The Value)

This is the final result. It is not software; it is the change in the real world. Does this tool save time? Does it cure a disease? Does it increase revenue?

Why it matters: Impact dictates sustainability. If an AI project doesn’t generate real value (ROI or societal good), it will eventually be shut down, no matter how cool the technology is.
The Constraint: Human Behavior and Economics. Just because a tool exists doesn’t mean people will change their habits to use it.

The Bottleneck Theory: Why Progress is Non-Linear

The most important thing to understand about these layers is that they must work in coherence.

We cannot simply “upgrade” one layer and expect the whole system to improve. In fact, the system is always limited by its weakest link.

Historical Example: In the 1960s, AT&T invented the Picturephone. It was a brilliant Layer 4 (Application) idea. But Layer 1 (Network Bandwidth) wasn’t ready. The product failed spectacularly.
Current Example: Today, we have incredible Layer 3 (Agent) concepts—AI employees that can do everything. But often, Layer 2 (Model Reliability) isn’t quite there yet; the models still hallucinate occasionally. As a result, the “AI Employee” fails to be reliable enough for critical work.

This interdependence creates a “hurdle for adoption.” You might have the budget and the desire, but if one layer in the stack is immature, your project will stall.

Guidance: The Incremental Approach

So, how do you build when the stack isn’t perfect? You adopt an Incremental Approach.

Instead of trying to build the “Ultimate AI System” that relies on every layer being perfect, you build for the layers that are ready today.

A sample scenario how to Approach incremental Build:

Start with “Human-in-the-Loop” (Layer 3 Lite): Don’t try to build fully autonomous agents yet. Build “Copilots” where the AI drafts the work, and a human reviews it. This mitigates the Layer 2 (Accuracy) risk.
Focus on “Low-Risk” Applications (Layer 4 Safety): Deploy AI in internal brainstorming or draft generation before putting it in front of customers.
Scale as Layers Mature: As models get cheaper (Layer 1 improves) and smarter (Layer 2 improves), you gradually remove the human guardrails.

Advantages:

Immediate Value: You get ROI now, rather than waiting 5 years for “AGI.”
Learning: Your organization learns how to work with AI data and workflows.
Safety: You avoid catastrophic failures by keeping humans involved.

Disadvantages:

Maintenance: You have to constantly update your system as the underlying layers change.
Process Change: It requires changing how people work (training them to use Copilots), which is often harder than just installing software.

By respecting the bottleneck, you build systems that actually work, rather than science fiction that breaks on day one.

Summary

In this post, we explored the vertical dimension of AI execution: the 5 Layers of Impact. We saw how a seemingly simple AI application is actually supported by a complex stack of Hardware, Models, Agents, and Interfaces, and why the “weakest link” in this chain often determines success. But understanding the pillars and layers is only half the picture. In the next post, “How Pillars and Layers Work Together,” we will merge the horizontal Pillars and vertical Layers into a unified perspective. This approach will allow you to predict the behavior, timeline, and constraints of any AI project by understanding how technical layers interact differently across each distinct domain pillars.

Previous Post : The Applied AI Thoughts for Realization Blog Post 2

Understanding How AI Thinks (and Where It Doesn’t) – Part 1 Are LLMs Really Understanding?

From a DeepSeek Article to trying to Understand Semantics vs Reasoning Cognitive concepts in AI

Introduction:

This first part captures the beginning of my thought journey. What started as reading an article about DeepSeek’s long-text technique slowly turned into a more fundamental question about what we really mean when we say an AI system “understands.”

A Simple Article That Led to a Big Question

I recently read an article about a research study that questioned a technique used by DeepSeek to help AI models read very long texts. The idea sounded impressive: compress large amounts of text so an AI can process more information at once.

But the researchers found something surprising.

The AI seemed to perform well not because it truly understood the text, but because it relied on patterns it had seen before. When those patterns were disrupted, the model struggled badly.

Even though I already had a working understanding of how LLMs work and Transformer architectures function, something about this finding triggered my interest to learn deeper. If these models were struggling the moment patterns broke down, what exactly were they doing when we say they “understand” text?

This thought triggered a deeper line of questioning in my mind — not about DeepSeek specifically, but about how we interpret progress in GenAI as a whole.

That curiosity naturally led me to ask:

Are modern AI systems really understanding, or are they just very good at guessing?

Once that question formed, it became clear that I needed to first separate two ideas that are often mixed together: semantic understanding and cognitive capability.

Semantic Understanding (Knowing What Something Means)

The first concept I needed clarity on was semantic understanding — a term frequently used but rarely unpacked.

Semantic understanding simply means understanding the meaning.

In everyday language:

It answers the question: “What does this mean?”

Large Language Models (LLMs) are exceptionally strong in this area.

They can:

Read a paragraph and explain it
Summarize documents
Translate languages
Recognize relationships between ideas

For instance, when an AI explains a legal document or summarizes a report, it is exercising semantic understanding. In many ways, this mirrors how humans comprehend words and sentences.

However, as I reflected on the DeepSeek article, an important limitation became obvious.

Semantic understanding stops at meaning.

It explains what is being said, but it does not decide what should happen next.

That realization naturally pushed me toward the next question: if understanding meaning is not enough, what role does reasoning actually play?

Reasoning Models (Thinking Better, Not Acting Better)

At this point, my attention shifted to reasoning models, often marketed as “thinking” AI.

These models are designed to show their work. They break problems into steps, apply logic, and produce more structured explanations.

On the surface, this feels like a major leap forward — and in many ways, it is.

But when I looked more carefully, I noticed that reasoning models still revolve around a single question:

“What is the best response to this input?”

Even with better logic, they still do not:

Choose goals (which is critical for decision-making — without goals, outputs remain just well-organized facts)
Take responsibility for outcomes
Act independently in the world

So while reasoning models think better, they don’t actually decide.

This insight clarified something important for me: reasoning improves semantic structure, but it still operates within the same boundary.

That naturally led to the next question — if neither understanding nor reasoning decides action, then what does?

Part 1 Conclusion: A Boundary Becomes Visible

By the end of this first part, one boundary had become very clear to me.

Understanding meaning and reasoning about it — even in sophisticated ways — does not automatically lead to decision-making or action. Something else is required.

In the next part, I will share my learning about the missing layer: cognitive capability, and why AI agents represent an important architectural shift rather than just a smarter model.

The ERP Awakening: From System of Record to System of Intelligence

The Foundation of Stability

For the last 30 years, the enterprise software industry has focused on one massive engineering achievement: Stability.

Enterprises have implemented SAP, Oracle, and Microsoft Dynamics to serve as the bedrock of their operations. They optimized for the “System of Record”—an immutable, reliable vault where every transaction is stamped, stored, and secured. In this regard, the strategy succeeded. The foundation is solid.

The Challenge: Data Rich, Insight Constrained

However, a vault is designed to keep things in, not necessarily to let insights out.

Today, the modern ERP operates like a massive, well-organized reference library. It contains all the answers—”Why is margin down?”, “Which supplier is late?”—but finding them requires users to walk the aisles, pull specific files (T-Codes), and decode complex rows of data. This architecture creates three distinct layers of operational friction:

The Insight Latency: Business leaders cannot ask questions directly. They often rely on technical intermediaries to build reports, leading to a “time-to-insight” gap of days or weeks.
The Productivity Burden: Skilled professionals spend hours on high-volume, manual tasks—drafting standard emails, visually verifying invoices against purchase orders when there is an exception, or creating requisition forms.
The Execution Variance: Critical workflows can experience delays due to minor “micro-stops”—like a pricing discrepancy of a few cents—that require manual human intervention to clear.

While the enterprise possesses the data, it often lacks the agility to act on it instantly.

Moving from System of Record to System of Intelligence

If the modern ERP is a comprehensive library, the operational bottleneck lies in the absence of a guide. Users are currently forced to act as their own researchers—navigating complex schemas and table structures just to retrieve basic facts.

Hence the strategic value of Generative AI lies not in replacing the library (the ERP), but in providing an intelligent Librarian to navigate it. By layering cognition over storage of records, enterprises can transition from a passive System of Record to an active System of Intelligence.

The “Three Stages” of Change

To make this transition actionable, organizations should view the evolution from a System of Record to a System of Intelligence not as a single leap, but as three distinct stages of maturity. Each stage builds trust and capability, moving from passive insight to active orchestration.

Stage 1: Synthesizing Intelligence (The Conversational Analyst)

Key Objective: To democratize access to complex ERP data, enabling “self-service” analytics without technical dependency.
Strategic Rationale: The primary bottleneck in most enterprises is “Insight Latency.” Business users face a barrier to entry—they do not know the technical schema required to query the ERP. The first step is to remove this friction by allowing natural language interrogation of the data.
Execution Strategy: Enterprises implement Text-to-SQL layers that act as a “universal translator.” Instead of navigating menus, users query the database using natural language. The system translates the intent into a precise SQL or OData query.
Tangible Impact:
- Use Case: A Regional CFO needs to understand a sudden variance in APAC logistics costs. Instead of commissioning a BI report (3-day lag), they ask the system directly and receive a visual breakdown of freight surcharges in seconds.
- Outcome: Zero time-to-insight for ad-hoc queries.

Stage 2: Augmenting Operations (The Generative Assistant)

Key Objective: To standardize communication and documentation while significantly increasing workforce velocity.
Strategic Rationale: Once users have insight, they must act on it. Often, this action involves creating content—emails, contracts, or summaries. This stage focuses on removing the “Blank Page” fatigue that drains high-value human talent on low-value drafting tasks.
Execution Strategy: This involves Content Generation thru Context Injection. The architecture feeds specific transaction data (such as open Purchase Orders or vendor contracts) into the LLM prompt, instructing it to draft content based on that specific reality for human review.
Tangible Impact:
- Use Case: A procurement team needs to send dunning emails to 50 suppliers regarding late shipments. The Assistant auto-drafts 50 unique emails, each referencing the specific PO number, delay duration, and relevant penalty clauses from the master contract.
- Outcome: Massive productivity gains and strict legal/policy compliance in external communications.

Stage 3: Autonomous Orchestration (The Process Agent)

Key Objective: To achieve “Zero-Touch” processing for routine variances, freeing human capital for complex problem-solving.
Strategic Rationale: Speed is often lost to minor details. Traditionally, any error—no matter how small—halts the process for human review. This stage shifts the paradigm to “Management by Exception,” where the system autonomously resolves routine problems, leaving only complex strategic decisions for human experts.
Execution Strategy: Deploying Agentic Automation. Autonomous agents are granted write-access to specific API endpoints and governed by strict policy logic (e.g., “If variance < $5, then approve”).
Tangible Impact:
- Use Case: The Accounts Payable close is stalled by hundreds of “micro-variances” where invoice totals differ from POs by cents due to rounding errors. The Orchestrator scans, verifies the tolerance policy, and posts the clearing documents automatically.
- Outcome: A faster financial close and a shift of human effort from data entry to strategic relationship management.

The Engineering Challenge: Building Trust

While this transition unlocks immense potential, it forces IT departments to confront a fundamentally new maintenance paradigm: the shift from managing deterministic code to governing probabilistic behaviors.

In traditional systems, if a report generates a wrong number, it is usually a bug in the code that can be traced, patched, and redeployed. In the era of AI, systems face Probabilistic outcomes. A model might generate a slightly different answer depending on context.

This requires new “safety rails”:

Glass Box UI: Systems must always show the user where the answer came from (citations).
Human-in-the-Loop: For high-stakes actions (like paying a vendor), the AI should draft the proposal, but a human must execute the final approval.

The Path Forward

The journey to a GenAI-augmented ERP is an architectural evolution, not a “rip-and-replace” project. To manage risk and ensure successful adoption, enterprises should align their implementation roadmap with the three-stage maturity model defined above.

By starting with Stage 1 (Insight), organizations can validate data accuracy and build user trust in a safe, read-only environment. Once confidence is established, they can advance to Stage 2 (Creation), introducing productivity gains while maintaining human oversight. Finally, only after proving stability, should they progress to Stage 3 (Action) for autonomous processing. This measured evolution ensures that capability grows alongside governance, minimizing operational risk while maximizing business value.

At 1CloudHub we are closely working with Enterprise customers to help them navigate the path to maturity through our consulting services and our solutions and products that help Enterprise to accelerate the pace of adoption to augment GenAI with ERP systems.

Coming Up – Navigating Day 1 Challenges

In the next post, the focus will shift to the foundation. Before building these intelligent layers, enterprises need to ensure their data is ready to support them. The discussion will cover practical strategies for Data Hygiene and how to start small with “Sidecar” pilots.

Coming Up: Post 2 – Navigating Day 1 Challenge : The Practical Reality of Implementation.

The Applied AI Thoughts for Realization Blog Post 2

A Simple Mental Model — How I Break the AI World into 4 Pillars

Introduction

In my previous post, I had shared about the need to shift from Tactical Thinking (chasing tools) to Structural Thinking (understanding the landscape) to understand the AI Landscape. In this post, we will build the foundation of that structure.

When we talk about “Applied AI,” it is easy to get fixated on the “AI” part—the models, the algorithms, the neural networks. But in the real world when we try to adopt AI, the model is often just one part of the equation.

Applied AI is not just about Models; it is a system.

To make AI work, you need more than just intelligence. You need data pipelines, user interfaces, safety guardrails, integration logic, and hardware infrastructure. You need to consider the human who uses it and the environment where it operates. When you look at the full picture, you realize that “AI” is just one ingredient in a complex recipe. And just like in cooking, the same ingredient (AI) produces a completely different result depending on what else you mix it with and how you serve it.

The Core Concept: Why “AI” Is Not One Thing

The biggest mistake organizations and individuals make is treating AI as a monolithic wave—assuming that the same rules, timelines, and strategies apply everywhere. They ask generic questions like “When will AI replace jobs?” or “Is AI safe?”

These questions do not have a simple straight forward answer because since adopting AI is not just focusing on one thing ie “AI”.

The Analogy: The Engine vs. The Vehicle

Consider an AI model (like GPT-4 or Claude) as a high-performance engine. An engine is a sophisticated core component, yet it provides no transportation utility on its own. To function effectively, it requires a chassis, wheels, a steering system, and an operator. It must be integrated into a complete vehicle.

Imagine attempting to solve every transportation challenge with a single strategy: “Install a high-performance sports car engine.”

On a racetrack (Consumer): This approach works perfectly; speed is the primary objective.
Plowing a field (Enterprise/Industrial): A high-revving engine is ineffective; the requirement is torque, traction, and sustained power under load.
Transporting cargo across an ocean (Logistics): Raw speed is irrelevant compared to fuel efficiency, durability, and massive scale.
Exploring the surface of Mars (Frontier/Science): A standard combustion engine will fail instantly due to environmental constraints; the need is for rugged autonomy and specialized engineering.

This is exactly how Applied AI works. The “Engine” (the intelligence) might be similar across different use cases, but the “Vehicle” (the application) must be radically different depending on the terrain. Some times even the Engine has to be modified for some use cases.

This principle applies directly when rolling out AI driven applications. Different applications require fundamentally different architectures, not just different features. Cotninuing with the vehicle anology below section talks about how we can map the 4 AI pillars to different vehicle type:

Consumer AI (The Sports Car): Optimized for high velocity, agility, and individual engagement. The priority is reducing user friction and maximizing experience.
Enterprise AI (The Freight Locomotive): Engineered for massive scale, unwavering reliability, and strict governance. The priority is secure, consistent throughput on defined rails.
Science AI (The Deep-Sea Submersible): Purpose-built for extreme precision in unexplored environments. The priority is navigating high-complexity domains to extract novel insights rather than speed.
Physical AI (The Industrial Rover): Designed for real-world interaction where the cost of failure is physical. The priority is safety, sensor integration, and navigating dynamic, unstructured environments.

If you try to apply “Sports Car” thinking to a “Cargo Train” problem, you will crash. This is why we need to break the AI landscape into 4 Pillars.

The 4 Pillars of Applied AI

Now that we have explored the vehicle analogy, it is clear why AI cannot be treated as a single entity when adopting and applying it. The architecture, stack, and strategy must vary based on fundamentally different challenges: speed vs. reliability, user delight vs. regulatory compliance, and digital outputs vs. physical safety. We can categorize these adoption patterns into four distinct pillars.

Pillar 1: Consumer AI

This is the AI that touches our daily lives. It is fast, personal, and often creative.

The Goal: Enhance individual productivity, creativity, or entertainment.
The Constraint: User Experience (UX) and Latency. If it takes 10 seconds to reply, users walk away. If it’s hard to use, they ignore it.
The “Vehicle Anology”: The Sports Car. It’s about speed, style, and the driver’s feeling.
Real-World Examples:
- ChatGPT / Claude: Chatbots that help you write emails or plan trips.
- Midjourney: Tools that generate art from text.
- Siri / Alexa: Voice assistants that manage your home.

Pillar 2: Enterprise AI

This is the AI that powers businesses and organizations. It is serious, governed, and integrated.

The Goal: Automate processes, analyze data, and augment knowledge work at scale.
The Constraint: Accuracy, Security, and Integration. A chatbot that hallucinates a discount code is annoying; a financial AI that hallucinates a revenue number is a lawsuit. It must connect securely to internal data.
The “Vehicle Anology”: The Cargo Train. It carries a heavy load, runs on fixed rails (processes), and reliability is more important than 0-60 mph speed.
Real-World Examples:
- Customer Support Bots: Systems that handle thousands of refund requests automatically.
- Code Copilots: Tools that help developers write secure code faster.
- Legal Document Analysis: AI that reviews contracts for risks.

Pillar 3: Science & STEM AI

This is the AI that pushes the boundaries of human knowledge. It is precise, computationally expensive, and transformational.

The Goal: Accelerate discovery in biology, physics, chemistry, and math.
The Constraint: Precision and Complexity. “Good enough” isn’t acceptable here. The AI must model the laws of physics or biology accurately.
The “Vehicle Anology”: The Deep-Sea Submersible or Space Rover. It goes where humans physically cannot, exploring the unknown depths of data.
Real-World Examples:
- AlphaFold: AI that predicts protein structures, revolutionizing biology.
- Weather Forecasting Models: AI that predicts extreme weather events with higher accuracy than traditional physics models.
- Material Science Discovery: AI finding new battery materials.

Pillar 4: Physical AI

This is the AI that leaves the screen and enters the real world. It is the hardest pillar because the real world is messy and unforgiving.

The Goal: Interact with physical objects, navigate environments, and perform manual tasks.
The Constraint: Safety and Physics. If a chatbot makes a mistake, you get bad text. If a robot makes a mistake, it breaks something or hurts someone.
The “Vehicle Anology”: The Industrial Robot or Autonomous Truck. It must be rugged, aware of its surroundings, and fail-safe.
Real-World Examples:
- Waymo / Tesla FSD: Autonomous vehicles navigating traffic.
- Warehouse Robots: Amazon’s robots moving packages.
- Humanoid Robots: Emerging robots designed to fold laundry or work in factories.

Why This Distinction Matters

You might ask, “Why not just categorize AI by what it does—like Text AI vs. Image AI?”

Categorizing by modality (text, image, video) tells you what the tool is, but it doesn’t tell you how to manage it. A text model used to write a poem (Consumer) behaves completely differently from a text model used to summarize a medical record (Enterprise).

By categorizing by Pillar, you gain a clearer understanding of what to expect. You can immediately identify the constraints, timelines, and success metrics that apply to your specific AI project.

1. Different Speeds of Innovation

Consumer AI moves at the speed of software. New apps launch weekly.
Physical AI moves at the speed of hardware and safety regulation. It takes years to certify a robot or a self-driving car.
Mistake to Avoid: Don’t get frustrated that your warehouse robots aren’t improving as fast as ChatGPT. They are in a different pillar with different friction.

2. Different Measures of Success

Consumer AI is measured by engagement and delight.
Enterprise AI is measured by ROI, accuracy, and cost-savings.
Science AI is measured by breakthroughs and new knowledge.
Mistake to Avoid: Don’t judge a scientific model by its user interface, or an enterprise tool by how “fun” it is to chat with.

3. Different Risk Profiles

If a Consumer image generator makes a weird picture, it’s a meme.
If an Enterprise legal bot hallucinates a clause, it’s a liability.
If a Physical robot fails, it’s a safety hazard.

When you know which pillar a project belongs to, you can immediately anticipate:

What constraints will dominate (speed? safety? accuracy?)
What stakeholders will be involved (users? regulators? scientists?)
What timeline is realistic (weeks? months? years?)
What failure modes to expect (bad UX? compliance issues? physical harm?)

Instead of discovering these answers the hard way—through trial and error—the pillar framework lets you predict them upfront. This is the “predictive power” of structural thinking: you’re not just reacting to problems; you’re anticipating them before they occur.

When you understand which pillar you are operating in, you stop applying the wrong rules to the game. You stop trying to drive a tractor like a Ferrari.

This clarity transforms how you approach any AI initiative. Rather than asking the vague question “How do we adopt AI?”, you can now ask the precise question: “Which pillar does this project belong to, and what does that tell us about how to execute it?”

For example, if your company wants to build an internal knowledge assistant for employees, you know immediately that you are in the Enterprise pillar. This means:

You will need to prioritize data security and access controls from day one
The AI must integrate with your existing identity management and document systems
Hallucinations are not just annoying—they could spread misinformation across your organization
Your success metric is not “how engaging is the chat” but “how much time did we save” and “how accurate are the answers”
You should expect a 3-6 month rollout, not a weekend prototype

Contrast this with building a creative writing assistant for novelists, which sits in the Consumer pillar. There:

Speed and personality matter more than perfect accuracy
Users expect a delightful, intuitive interface
Your success metric is user retention and satisfaction
You can iterate weekly based on user feedback

The same underlying language model could power both applications, but the vehicles you build around that engine are completely different. The pillar framework gives you this insight before you write a single line of code or sign a single vendor contract.

Summary

In this post, we established the first fundamental layer of structural thinking: the 4 Pillars of Applied AI—Consumer, Enterprise, Science, and Physical. We explored how the same underlying AI “engine” produces radically different outcomes depending on the “vehicle” it powers. Most importantly, we learned that knowing which pillar your project belongs to allows you to predict its constraints, stakeholders, timelines, and failure modes before you begin.

But identifying the right pillar is only half the story. Even within a single pillar, AI projects succeed or fail based on how well the underlying layers—from hardware to models to applications—work together. In the next post, “The Impact Layers — How AI Progress Actually Happens,” we will dive beneath the surface to explore the 5-layer stack that determines whether AI potential translates into real-world value, and why even the smartest model can fail if a single layer is weak.

References

Previous Blog in Series : Blog 1: Why AI Feels Overwhelming — And Why That’s the Wrong Way to Look at It
Next in Series : Blog 3: The Impact Layers — How AI Progress Actually Happens (Coming Soon)

The Applied AI Thoughts for Realization Blog Post 1

Why AI Feels Overwhelming — And Why That’s the Wrong Way to Look at It

Introduction

In case of people who closely follow AI related developments as well as for people who are in early stages of understanding the AI landscape its become very hard to track the developments in the space and understand how new technologies, tools, techniques and solutions can be applied in their respective domains and use case ideas they have. For the ones closely following developments in AI space every morning, it feels like the landscape of Artificial Intelligence has shifted overnight. You wake up to a barrage of headlines: a new Large Language Model (LLM) that crushes previous benchmarks, a new image generator that renders reality perfectly, or a new agentic tool that promises to automate your entire workflow.

For engineers, leaders, and decision-makers, this constant acceleration often triggers a mix of excitement and anxiety. There is a pervasive fear of falling behind—the sense that if you don’t master this specific tool released today, you will be obsolete tomorrow. This is “AI Fatigue,” and it is the natural result of trying to drink from a firehose without a cup.

The objective of this blog series, Applied AI Thoughts for Realization, is to help the readers to put down the firehose and step back. The objective is not to cover the latest news or review the newest tools. Instead, the goal of the blogs in this series is to provide you with a structured mental model—a way to organize the chaos into a coherent map.

Over the course of this series, I will try to avoid the hype cycles and focus on a first-principles approach to understanding the AI landscape. I will help you to explore how to categorize AI into distinct “Domain Pillars” based on where it is applied, and how to understand the dependencies and progress within the Domain Pillars through specific “Impact Layers.”

By the end of this series, you won’t just have more information; you will have a mental model framework. When you encounter any news about new AI developments or innovations—whether it’s a breakthrough in consumer gadgets, an enterprise platform launch, or a scientific research milestone—you will be able to instantly map it to its specific domain pillar and identify which layer it operates within. This clarity will help you understand not just what the announcement is, but where it fits in the broader landscape, why it matters in that context, and whether it’s relevant to your work.

The Problem the Series helps to Solve : The Trap of Tactical Thinking

Imagine you decide to build a house. You walk into a massive hardware store, credit card in hand.

On Monday, you buy a power drill because the salesperson says it’s the fastest one ever made. On Tuesday, you see a new type of saw that uses lasers, so you buy that too. On Wednesday, you hear about a revolutionary type of hammer, so you rush back to the store.

By the end of the week, your garage is full of cutting-edge tools. You are exhausted from researching specs and comparing brands. But when you look at your empty lot, you realize a painful truth: You haven’t laid a single brick. You have a collection of tools, but you don’t have a blueprint.

This is exactly where most of us are with Artificial Intelligence today. We are stuck in Tactical Thinking.

We treat AI as a shopping list of features and vendors. We obsess over the “tools”:

“Did you see the context window on that new model?”
“Is OpenAI better than Google for coding?”
“Should we use RAG or fine-tuning?”

While these questions aren’t irrelevant, asking them first is a trap. When you focus solely on the tools, you become reactive. You are constantly pivoting based on the latest press release. You judge AI progress by how fast the “drill” spins (model benchmarks), rather than whether it can actually help you build the “house” (solve a specific problem).

This tactical approach leads to two major issues:

Paralysis: You are afraid to commit to a solution because something better might come out next week.
Misalignment: You try to force a tool into a job it wasn’t meant for—like trying to frame a house with that laser saw just because it was expensive.

To escape this cycle, we need to stop looking at the tools and start looking at the architecture.

The Shift: From Tools to Structure

The antidote to tactical paralysis is Structural Thinking.

If tactical thinking asks “What tool should I use?”, structural thinking asks “Where does this problem live, and what are the constraints of that environment?”

When you shift your mindset from tools to structure, you stop chasing every new announcement. You realize that AI is not a single, monolithic wave washing over everything in the same way. Instead, it is a set of capabilities that behaves radically differently depending on the context.

Why Structure Matters for Scalability and Flexibility

The biggest advantage of structural thinking is that it future-proofs your strategy.

In the tactical world, your strategy is brittle. If you build your entire workflow around a specific vendor’s model, and that vendor changes their pricing or a competitor releases a better model next month, your strategy breaks. You are constantly rebuilding.

In the structural world, your strategy is flexible. You define the architecture of your solution—the data flows, the safety guardrails, the user interaction patterns—independent of the specific engine powering it.

If a new, faster model comes out? You simply swap it in as a component.
If a regulation changes? You adjust your governance layer without tearing down the whole application.

Structural thinking allows you to build systems that last, rather than prototypes that expire. It moves you from being a consumer of technology to an architect of solutions. It forces you to acknowledge that a “good” AI system for writing a marketing email is fundamentally different from a “good” AI system for controlling a robotic arm—not just because the tools are different, but because the structure of the problem (risk, speed, cost, accuracy) is different.

The Solution: A Preview of the Framework

To navigate this landscape effectively, we need a map. Over years of working with AI across various domains, I have developed a mental model that breaks the AI world down into two distinct dimensions. Think of it as a coordinate system for understanding any AI development.

Dimension 1: The 4 Domain Pillars (Where AI Applies) First, we must recognize that “AI” is not a single thing. It is a set of technologies applied in radically different environments. We divide the landscape into four vertical pillars:

Consumer AI: The AI we use in our daily lives (chatbots, image generators).
Enterprise AI: The AI that powers businesses (automation, data analysis).
Science & STEM AI: The AI that accelerates discovery (drug discovery, material science).
Physical AI: The AI that interacts with the real world (robotics, autonomous systems).

Dimension 2: The 5 Impact Layers (How AI Progresses) Within each pillar, progress doesn’t happen in a vacuum. It moves through layers of maturity, from the raw silicon to the final societal change:

Hardware: The chips and infrastructure.
Models: The algorithms and intelligence.
Agents & Tools: The orchestration that makes models useful.
Applications: The interfaces we actually touch.
Impact: The real-world value and behavioral change created.

The Power of the Grid When you combine these, you get a grid. You can place any news story, any tool, or any project onto this grid. Suddenly, the chaos disappears. You aren’t just looking at “AI”; you are looking at “Layer 2 (Models) within Pillar 3 (Science).” Below is a diagram that helps understand the framework.

This framework allows you to ignore the noise that doesn’t affect your specific coordinates and focus deeply on the areas that do.

What to Expect from This Series

This blog post is just the starting point. Over the coming weeks, we will unpack this framework piece by piece, giving you the tools to apply it to your own work.

Here is the roadmap for the series:

Part 1: Foundational Thinking We will dive deeper into the mental models. We’ll explore the 4 Domain Pillars in detail to understand their unique characteristics, and we’ll break down the 5 Layers to see how innovation actually flows from hardware to impact.
Part 2: Pillar-by-Pillar Deep Dives We will dedicate specific articles to each of the four domain pillars—Consumer, Enterprise, Science, and Physical AI. We will analyze the specific trends, constraints, and opportunities within each domain.
Part 3: Applying the Framework Finally, we will turn theory into practice. We will discuss how to use this framework to make better decisions, whether you are evaluating a new vendor, planning an internal AI project, or simply trying to stay ahead of the curve.

By the end of this journey, you will have a clear, reusable lens through which to view the AI landscape—one that turns information overload into actionable insight.

References

Next in Series: Blog 2: A Simple Mental Model — How I Break the AI World into 4 Pillars

Data Privacy vs Safety of Society : Which is important ?

Recently there has been lot of debates happening in the area of data privacy especially with regards to companies trying to take a high moral ground in protecting customer data from being exposed to Government Agencies. I am not just talking about the Apple fiasco happening but there have been many such cases in the recent past. Before letting know my view I would like to share my thoughts on the origins (as i perceive) of this problem.

Lets start with Internet (which people say was a short form for inter-networking), it all started with the invent of computers in 1950 followed by the evaluation of closed network computing in 1960’s and 70’s with only government and educational institutions using the networks to exchange information. Client Server computation was the approach with thin dummy clients used as interaction terminals and all computation and data storage centralized at the Servers.

With the invention of TCP/IP in 1978 the whole world of connectivity and applications using the technology exploded with it the users. Closed networks owned by government, defence companies and educational institutions evolved into commercial networks offered by ISP’s and other service providers.

Along the way other evolution were happening fast in the key areas like the computing power (Moore’s law effect), Memory and Storage that enabled the manufacturing of personal computers which were smaller in size and were affordable.

The killer combination of Personal Computers and the Invention of TCP/IP (Ethernet) enabled fast growth in the usage of computers and internet which became the back bone of all computational needs.

Internet Users evolved from few thousands in 1950’s to few millions in 1970’s and few hundred millions through 1980’s and 90’s. Turn of the century saw a major boom in internet users which currently stands at around 3.5 billion users at the end of 2015(approximately 45+% of world population).

This phenomenal increase in people using an open network has helped make the world to be very connected and has led to a information revolution. At the same time there are people who have also taken advantage of the openness and used it for sinister means.

Currently almost half the globe is using the internet not all of them are aware of the dangers involved using the internet due to which they fall prey to various types of attacks. Either getting virus into their computer or getting hacked into their banking accounts or trapped in a fishing fraud and the list goes on. My objective in this blog is not to talk about the impact Internet can cause to the privacy of illiterate internet users who get hacked and personal data taken away but about the people who use the same internet to hide what their doing and harm societies and why giving importance to the privacy of such people will bring harm to society.

As high lighted early the deadly combination of computing power in small devices added to it the access to an open internet provides a deadly weapon for people who would like to cause harm to societies around the world. With such capabilities people are able to communicate with others across the world to plan and execute sinister acts without getting noticed. People are now able to tap on to the computational capabilities available in smart phones or a Tabs or a Laptops to secure their data and communications without much difficulty. Comparing this, a decade back when the computational power of PC’s were equivalent to the power in a feature phone.

People started developing software and services that would help protect information and communications that became handy for the evil minded people. Sharing and Access to information became easy due to the pervasiveness of internet. Policing such a network has become a nigh mare. One can just use a Smart phone to access information, communicate, do commerce etc, they do not require to own a PC or Laptop to communicate and share information. They just need a device in the palm ie. which is the mobile phone or a tablet.

To tap on to the growth of internet and pervasiveness of the internet enabled devices companies started building various types of services to offer to the internet users. Among them the key being communication and social media services. This enabled people to create virtual communities to the extent that the entire online business model and valuation of such companies were based on the number of the user base they have. Facebook for example has almost 1.6 billion active users which is more than the population of China or India. Combined user base of top 10 such service providers is more than the total world population. If we look it from a consumer lens this is a very good news as they have lot of choices. But this has also started creating lot of social challenges. I again do not want to get into a judgement as there are contradictory views prevailing in this subject.

One of the key reason why people end up using such services in large numbers is that they have full freedom to be anonymous in using such services. Even though these are virtual services due to the reach it has its also has become a medium to be used by people to influence other peoples easily. Getting out ones message, irrespective of the stature and background, to huge mass of people has become very easy. At the same time, it is has become very hard to monitor and moderate such communications for various reasons and one of them being the protection given to users by the service providers in the name of privacy.

Definitely there are benefits that the consumers get by using such services the challenge is when people take advantage of the anonymity and try to cause harm to individuals or to society. We have heard of stories at personal level were people impersonate and try to spoil the reputation of other people, we have heard about people who spread rumors that cause major riots, kids preyed on by strangers and more. Since we do not have a structure and mechanism to monitor such acts of crime, protecting from such ill effects has become very difficult. Even agencies with money and resource power are only able to monitor specific leads.

In recent years the most worrying trend has been the act perpetrated by people taking advantage of the privacy protection offered by companies offering communication and social services has been act of terrorism. I do not want to provide you with examples of terrorist incidents that have happened around the world in the last few years but definitely the investigations have provided good insights into how people have used technology and services to carry out their acts of violence. People have effectively used the internet services to propagate messages, recruit people, organize and plan resources and communicate with each other. The security agencies in various countries were unable to monitor such activities as all the communication to perform the activities were protected by the relevant service providers.

The sad part of all this is how the technology providers and services providers are falling over each other to defend they stand on privacy. My question is how would they answer to the thousands of people who have lost their life’s due to these terror acts and in which the technology and services have provided a major support to carry out these acts.

There are many questions for which we have no answers yet to effectively address the challenge. Below are few of them,

How do we avoid people taking advantage of the anonymity provided by technology and service providers ?
How can the social media and communication services providers can self regulate, moderate and monitor their user base to look for people using it for sinister means ?
What kind of structure and regulatory policies governments can work out so that their citizens can have faith in the process when they access citizen data ?
How to control activities performed across boarders ?

Below are some of my thoughts on the subject of balancing privacy vs safety to societies,

People have to realize and recognize internet and usage policy of the country from where they use the internet. Until citizens follow the same one need not worry about any persecution or impact to privacy.
People need to be careful when sharing information in the internet. They need to realize its a open world there and they dont have any control on how the information they share will be used. Hence the key is that they are aware of what they share and who they share with.
People need to realize governments have to do the duty of protecting their citizens. As described above with so much stacked up against them its not fair to blame them if they try to monitor or access citizens data for intelligence gathering pro active action or post incident investigations.
Governments need to put in place dedicated agencies which will be responsible to perform the monitoring of citizens internet activities and ensure proper process and security mechanism in place. The agencies need to be regulated by an independent organization which have representation of people and organization trusted by citizens.
Companies providing social and communication services are obliged to ensure that people using their services are complying with the policies and guidelines of the country from where the service is used. This can kill the viability of some of these business but it does not mean that it can cause harm to individuals and communities of any country by people using their services.

With the strong feelings around privacy, freedom of expressions, being anonymous etc my above thoughts may not be welcomed my majority of internet citizens. My thoughts are purely based on my belief that if a government or agency can save my life and that involves getting access to my personal data I am ok for such compromise on privacy.Either my internet activity being monitored or I am body scanned or Tap Checked in Airport or My Bags being scanned I always feel its been done to save my life !

Big Data Tools One Size Does Not Fit All when it comes to Enterprises

We have heard enough about Big Data in the last few years. The technologies, the benefits, the big ideas etc. Now we are moving into the era of Application ie. how to realize the real benefits. Yes the Interprisers (Internet Companies) are in the fore front of using the various technologies and the data effectively since most of the current set of tools have their origins from these Interprising companies. The sucess of the current set of tools in the Interprises are because of the nature of the data they have to handle which is predominently unstructured. But when it comes to the real enterprises the challenge arises because of the nature of the data and the operations they perform on their data. The enterprise data are mostly structured and the operations performed on them are mostly relational. It does not mean Enterprises cannot benefit from using the Big Data technologies. The benefits a traditional enterprise would get using the current set of big data tools and technologies would not be same as what an Interprising company would be looking at. Hence enterprises need to have a different approach in using the technologies and tools. One of the key aspects will be looking at interoperability of the traditional big data tools sets like hadoop or nosql db’s etc with traditional Enterprise data management tools used for datawarehousing, ETL, BI related functions. Its critical that Enterprises have a clear principles to achieve the same.

Connected Devices Juggernaut

Google’s purchase price of 3+ billion for a start up that has a solution for Connected device is big considering that the solution is yet to prove its worth. Looking at the bigger picture connected devices and the information / data it will be generating is going to be a major area of focus in coming years. The path for success for companies focused in providing products/solutions is not going to be smooth. The reason for the same is the dependencies and challenges prevailing in the existing environment/ecosystem. Current solutions available out in the market are still a closed loop ones and proprietary. Including the Nest solution which Google has spent its monies for. As we have seen and has been proven again and again its critical to have an open platform and ecosystem for such technologies for mass adoption. Below are a list of challenges that i see is required to be addressed for the connected devices story to really take off.

1. Limitation related to connectivity. The solutions should make use of pervasive wireless technologies like the mobile networks rather than using fixed wireless technologies like WiFi or low range technologies like zigbee.

2. Identity and Authentication : There needs to be a programmable identity module framework that will help seamless connection to networks.

3. Data Privacy and Security : This has been a hot topic of the current internet services and would still prevail and be a strong barrier to data generated by connected devices. If we do not have a trusted model and framework for data sharing its hard to convince people of using such devices and sharing data.

4. Data Management (Collection, Management and Sharing) : Huge amounts of data are expected to be generated by connected devices. There needs to be solutions and services in place that can help collect such data, manage the data and help share the data.

5. Standards : Connected devices can be deployed by an enterprise, individual, government agencies for various purposes. The categorization and organization of such data in a standards based approach is critical. There are some standards based activity carried out by some organization and in certain technology areas. We need to have a single body to frame the end to end standards related to connected devices.

6. Cost and Revenue : One of the key barrier for connected devices is going to be the cost. One of the factors that will impact the same is the revenue model and the mass adaptability of the solution.

There is going to be a great potential for connected devices (imagine a petrol pump sensor alerts a user of low gas and the user immediately gets a phone alert with a google maps link to the nearest petrol pump with directions. To add on to this use case imagine if the back-end application that receives constant updates on the gas levels and based decides to send the notification based on the time of the day, gas pump distance from current location etc.). They key now is Organizations need to quickly address the various barriers that would hinder the adoption of connected devices.