AI Needs a Theory of Constraints: How Do Governments & Enterprises Can Create Higher Societal Value from AI?

      AI is advancing at an extraordinary pace.

Hundreds of billions of dollars are being invested across the AI ecosystem, and the demand for more investment continues to grow. AI is becoming pervasive, touching almost every aspect of our personal and professional lives.

Competition is intense. Development cycles are accelerating. At the same time, some industry leaders and policymakers are increasingly calling for appropriate safeguards and regulatory frameworks.

But there is another question that deserves more attention:

Are we getting proportional societal value from the enormous investment and effort going into AI?

I believe there may be a systemic reason why the answer is not always yes.

AI is a chain, not a single technology

Investment is happening across the entire AI value chain covering Data centers, storage and cloud infrastructure, Energy and communications, Chips and semiconductors, Foundation models, Data infrastructure, Engineering and coding, Cybersecurity, Robotics and industrial AI, Autonomous systems, Business-specific applications, AI skills and capability building, and Governance and regulatory frameworks

A significant proportion of investment is currently concentrated in infrastructure and foundational capabilities.

Yet, at the same time, we repeatedly encounter situations such as:

• AI is hungry, but the grid isn't ready.
• The pilot is successful, but the data isn't ready.
• The AI agent is ready, but cybersecurity and guardrails aren't ready.
• The capital is available, but the required compute or chips aren't available.
• The application is ready, but the regulatory framework isn't ready.
• The technology is ready, but the organization isn't ready to redesign its processes around it.

Each of these represents a mismatch somewhere in the AI value chain. And that mismatch can limit the value that the entire system can deliver.

Could this be a Theory of Constraints problem?

The Theory of Constraints (TOC) by Goldratt is based on a relatively simple principle: The throughput of a system is constrained by its most limiting constraint.

Instead of trying to optimize every component independently, TOC asks us to identify the constraint that is limiting the overall system, make the best use of that constrained resource, align the rest of the system around it, and then elevate the constraint.

Once the constraint moves, we identify the next one.

I believe this thinking can be applied to the AI ecosystem.

But there is one important difference. A manufacturing system typically optimizes throughput for a business. For AI, the ultimate objective should be broader i.e. Maximize value AI creates for humanity

Applying TOC to AI

Classic TOC-- AI ecosystem interpretation

Identify the constraint-- Find what is currently preventing AI from creating more societal value

Exploit the constraint-- Get maximum utilization from the existing scarce capacity

Subordinate everything else-- Align other resources and activities around the constraint

Elevate the constraint-- Invest, innovate or change policy to increase the constraint's capacity

Repeat-- Once the constraint moves, identify the next constraint

What could the constraints be?

The constraints can exist at different points in the AI ecosystem.

• Capital: Is there sufficient financing for infrastructure, applications and deployment?
• Compute: Are sufficient GPUs, accelerators and computing resources available?
• Energy: Can electricity generation and availability keep pace with AI infrastructure?
• Physical infrastructure: Are data centers, grids, cooling systems and networks available?
• Data: Is the required data available, accessible, high-quality and usable?
• Talent: Do we have enough AI engineers and domain experts?
• Organizational capacity: Can organizations redesign processes quickly enough to take advantage of AI?
• Economics: Does the cost of AI deployment make economic sense relative to the benefits?
• Human adoption: Do employees, customers and citizens have the capability and willingness to use AI effectively?
• Problem selection: Are we applying enormous AI capabilities to the problems that can create the greatest value?

The constraint doesn't have to be technological.

In fact, the most important constraint may sometimes be organizational, economic, regulatory or human.

Let's consider two examples

1. Electricity availability

Imagine that the AI industry can build 100 GW of additional compute capacity, but the power system can support only 60 GW.

The additional 40 GW of compute cannot generate its expected value until the energy constraint is addressed.

In this situation:

Compute is not the binding constraint; Electricity availability is.

Investing more money in compute without addressing the energy constraint would therefore increase capacity without proportionally increasing societal throughput.

The same principle applies to data centers.

A data center may be constructed in 2–3 years, while energy generation, transmission and grid connection can require substantially longer planning and development cycles.

Therefore:

Building more data centers does not necessarily solve an AI capacity problem if the energy system cannot support them.

2. Capital allocation

Consider a hypothetical situation in which there are 10,000 potentially valuable AI applications, but only 1,000 have access to adequate capital.

Capital becomes a constraint.

But simply adding more capital isn't necessarily the answer. The more important question becomes:

Where will the next dollar of capital create the greatest increase in societal AI throughput?

Should it go toward: another foundation model or AI infrastructure or energy infrastructure or healthcare AI or education AI or cybersecurity or industrial AI or AI agents?

The answer should depend on which investment removes a binding constraint and enables the greatest additional societal value.

The multiplier effect of constraint removal

Some constraints may be more important than others because removing one constraint can unlock several others.

For example: Energy infrastructure enables Data-center capacity which enables Compute availability which enables AI applications which enables Enterprise adoption which creates economic and social value

In such a situation, investment in the original constraint can have a multiplying effect throughout the system.

This leads to an interesting question:

Should we measure investments not only by their direct return, but also by the additional AI capacity and societal value they unlock elsewhere in the system? I believe we should.

Constraint prioritization becomes critical

The AI ecosystem will almost certainly have multiple constraints simultaneously. Therefore, we need to prioritize them.

One possible approach is to evaluate each constraint based on:

• How much additional throughput will its removal create?
• How early is it in the AI value chain?
• How many downstream constraints will it unlock?
• How much investment is required to remove it?
• How quickly can it be removed?

And the constraint will keep moving

This may be one of the most important aspects of applying TOC to AI.

Today's constraint may be Compute, tomorrow it could be Energy then it could be Talent. Therefore, AI policy & action cannot be a one-time exercise. Constraint identification & removal needs to become a continuous process.

What role should governments and enterprises play?

I believe both have an important but different role.

Governments and Policy institutions

Governments can address constraints that affect entire ecosystems through:

• Infrastructure investment
• Energy and grid policy
• Education and workforce development
• Incentives
• Data policy
• Standards
• Regulatory frameworks
• Public-private partnerships
• Research and development
• International cooperation

The objective should not necessarily be for governments to pick individual AI winners. Instead, governments can help identify and remove systemic constraints that affect many participants simultaneously.

Enterprises

Enterprises should apply the same thinking internally.

Their constraints may be:

• Data availability
• Legacy systems
• Talent
• Capital
• Process design
• Integration
• Cybersecurity
• Change management
• AI governance
• Employee adoption

The question becomes:

What is the single most important constraint preventing our organization from obtaining more value from AI?

Rather than implementing AI everywhere simultaneously, organizations could focus their transformation efforts around that constraint.

The bigger question

AI capability may not ultimately be limited by the intelligence of the models.

It may be limited by our ability to build the infrastructure, energy systems, data ecosystems, organizations, skills, policies and human capabilities needed to convert that intelligence into value.

Perhaps the next AI challenge isn't simply building more AI.

Perhaps it is identifying and removing the constraints that prevent AI from delivering its full potential to society.

This is a concept I am currently developing, and I would genuinely value perspectives from people working across AI, infrastructure, investment, enterprise transformation, government policy and organizational change.

What do you think is the biggest constraint preventing AI from delivering greater value to society today?

The Conversational AI Paradox: Why Great Introductions Are Still Ending in Frustrated Customers

 We have officially entered the era of daily interactions with Conversational AI Agents. They are on the front lines of customer service, evolving rapidly, and each sitting at a different stage of human-like sophistication.

Having interacted with several AI voice agents across Retail, Banking, Telecom, and Insurance over the past few weeks, I’ve experienced a full roller-coaster ride of Customer Experience (CX).

Here is what is working exceptionally well—and where the experience breaks down.

Phase 1: The Good News (Greetings & Identity Verification)

1. The Voice & Opening Greetings The voice quality, natural cadence, and greeting mechanisms are nearly perfected. There is no longer that jarring robotic delay, making the initial connection comfortable for the caller. I will give it an A+ grade

2. Upfront Transparency Most agents immediately state: "I am an AI agent here to assist you." Setting expectations at the beginning which builds trust and establishes a strong CX foundation. I will give it an A Grade

3. Understanding Intent Instead of rigid interactive voice response (IVR) menus ("Press 1 for Account Balance…. "), AI agents ask open questions in natural prose. In 80–85% of my tests, the agent accurately reflected my intent back to me in plain English, adapting effortlessly to my accent and vocabulary. I will rate them at A- grate

4. Frictionless Authentication Identity verification across industries was seamless. In one case, the AI identified me by my phone number, confirmed my name, and thanked me for being a premium club member. At this stage, the CX felt modern, hyper-personalized, and impressive. This is at A+ Grade

Phase 2: Where the Experience Breaks Down

While the beginning of the journey feels seamless, the actual resolution phase reveals critical gaps. I am sharing two cases for better understanding of the challenge

Case 1: Repair Appointment (Hardware /Appliance store) - Scheduling Disconnect

• The Goal: Book a repair appointment.
• The Interaction: The AI agent picked up immediately, verified my details, and walked through morning/afternoon options for the next few days. We booked the appointment for tomorrow. Appointment details were sent to me over an e-mail.  I was thrilled.
• The Flaw: The confirmation email arrived for the day after tomorrow, rather than tomorrow as requested. The whole interaction with agent seemed to be a waste. Driven by urgency, I walked into the physical store directly. To my surprise, there were multiple open slots available for that same day and I could complete my work on the same day. I was now doubting the capability of Agent to bring in right information resulting in to a really bad impression
• The CX Takeaway: A delightful conversation means nothing if the underlying logic delivers inaccurate outcomes. It leaves the customer questioning the utility of the AI altogether.

Case 2: Obtaining Clarification (Mortgage / Financial Services) - The Infinite Loop & Abrupt Hang-Up

• The Goal: Seek clarification on a complex notice.
• The Interaction: Authentication and intent-capture went smoothly. Very Happy
• The Flaw: After asking about the clarification, the agent forced the conversation into a rigid option-tree loop. When I explicitly said, "I am unable to find the right option, please connect me with a human agent," there was a brief pause—and the line went dead. I was shocked and frustrated
• The CX Takeaway: Getting stranded without a human fail-safe leaves customers feeling undervalued. It conveys that the organization prioritizes cost reduction over actual service quality.

The Conclusion: Bridging the "Midway" Gap

Conversational AI implementations currently excel at starting a great experience, but often fail to sustain it through resolution. To take AI agents from midway execution to a true human-AI collaboration, organizations should focus on three focus areas:

1. Train on Complex Edge Cases: Base models must be fine-tuned on real-world, non-standard customer scenarios rather than perfect script flows.
2. Implement Real-Time Sentiment Analysis & Handoffs: Systems need background monitoring to evaluate call "pulse." If an interaction stalls or sentiment degrades, it should trigger an instant, context-aware escalation to a live human agent or recommend alternative path for conversation to the AI agent. This will help retaining customer faith and their experience stays high
3. Orchestrate Specialized Agents: Rather than relying on a single generalist AI, deploy specialized sub-agents (e.g., scheduling, billing, complex troubleshooting) that pass context seamlessly behind the scenes.

I remain immensely optimistic about the future of Conversational AI as technology continues to mature.

Over to you: What has your experience been with live AI voice agents recently? Are you seeing smooth resolutions, or hitting similar execution walls? Let’s discuss in the comments below! 👇

Customer Experience: Most Critical Context for Future AI Systems

 The AI conversation today is dominated by discussions around models, reasoning capabilities, automation, and autonomous agents. Organizations across industries are rapidly deploying AI-powered assistants, copilots, and workflow automation platforms in pursuit of higher efficiency and lower operational costs.

But as enterprises move from experimentation to real-world implementation, one realization is becoming increasingly clear: The effectiveness of AI systems depends not only on the intelligence of the model, but on the quality of the context available to it.

Even highly advanced AI models can produce poor outcomes when they lack situational awareness. Conversely, a well-contextualized AI system can deliver highly relevant, personalized, and operationally effective experiences.

The future of enterprise AI may therefore depend less on “who has the smartest model” and more on:

• who manages context best,
• who integrates enterprise knowledge deeply,
• and who operationalizes customer understanding effectively.

What Is Context in AI Systems?

In simple terms, context is the information that helps an AI system understand:

• what is happening,
• who the user is,
• what objective is being pursued,
• what has already occurred,
• and what constraints or expectations matter in that moment.

Context gives AI systems situational awareness.

Without context, AI interactions become transactional and disconnected. With context, they become intelligent, adaptive, and continuous.

Common Contexts Used in Today’s AI Systems

Modern AI applications and AI agents already leverage several forms of context.

1. Conversational Context

This includes the history of interactions between the user and the AI system. E.g. previous questions, prior responses, unresolved issues, ongoing workflows, etc.

This enables continuity in conversations instead of forcing users to repeat information.

2. User Context

AI systems increasingly personalize responses using: user preferences, role, expertise level, behavioral patterns, and historical interactions.

A finance executive, a clinician, and a customer support representative may all receive different responses to the same question because their contexts differ.

3. Task and Workflow Context

AI agents often need to understand the current task, workflow stage, business priorities, SLAs, and dependencies across systems.

This is especially important in enterprise environments where AI is expected to participate in operational processes rather than simply answer questions.

4. Business and Domain Context

Enterprise AI systems are increasingly grounded in organizational policies, compliance rules, pricing structures, clinical workflows, payer rules, and operational procedures.

Without domain context, AI outputs may sound intelligent but remain operationally unusable.

5. Real-Time Environmental Context

AI systems also use dynamic signals such as location, device, channel, system state, real-time events.

This enables adaptive and situationally aware interactions.

The Missing Layer: Customer Experience Context

While many organizations focus on technical and operational context, one of the most critical context layers is often underdeveloped: Customer Experience (CX) context.

Most current AI systems understand transactions - Far fewer truly understand experiences.

A customer interaction is rarely just a single event. It is part of a broader journey shaped by:

• expectations,
• emotions,
• prior interactions,
• friction points,
• trust levels,
• urgency,
• and relationship history.

This is where CX context becomes essential.

What Is CX Context?

CX context is the collective understanding of:

• where the customer is in their journey,
• what they are trying to achieve,
• how they feel,
• what problems they previously encountered,
• and what experience the organization aims to deliver.

It goes far beyond CRM data.

Traditional CRM systems may know:

• who the customer is,
• what they purchased,
• and when they interacted.

CX context additionally understands:

• frustration signals,
• repeated failures,
• escalation history,
• communication preferences,
• sentiment trends,
• loyalty indicators,
• and journey progression.

Why CX Context Matters for Future AI Systems

As AI agents become more autonomous, they will increasingly influence customer trust and brand perception directly.

Without CX context:

• AI interactions feel robotic,
• customers repeat themselves,
• personalization remains shallow,
• and frustration escalates quickly.

With CX context:

• interactions become continuous,
• empathy improves,
• resolutions accelerate,
• proactive engagement becomes possible,
• and experiences feel genuinely personalized.

In industries such as healthcare, banking, telecom, insurance, and retail, this distinction can significantly impact both customer satisfaction and business outcomes.

The Risk of AI-Driven Efficiency Without CX Awareness

Many current AI implementations are heavily focused on primary objectives such as automation, Operational efficiency, and cost reduction. While these are valid business goals, organizations often overlook an important consequence: AI systems optimized only for efficiency can unintentionally degrade customer experience. This is already becoming visible across industries.

Customers increasingly encounter:

• difficult-to-escape chatbots,
• repetitive automated interactions,
• fragmented journeys,
• excessive self-service loops,
• lack of empathy,
• and delayed access to human assistance.

In many cases, AI implementations are measured primarily on:

• call deflection,
• reduced handle time,
• lower support costs,
• or workforce reduction.

However, these metrics alone do not capture the broader business impact.

An AI system may successfully reduce operational costs while simultaneously:

• increasing customer frustration,
• reducing trust,
• lowering loyalty,
• increasing churn,
• and damaging brand perception.

The result is a dangerous tradeoff: short-term cost savings at the expense of long-term customer and revenue erosion.

The Hidden Cost of Poor AI Experiences

When organizations fail to incorporate CX context into AI systems, automation can become mechanically efficient but experientially ineffective.

Customers often perceive such systems as impersonal, rigid, difficult to navigate, and disconnected from their actual needs.

Over time, this creates:

• customer fatigue,
• declining satisfaction,
• lower retention,
• and reduced lifetime value.

Ironically, organizations may save money operationally while losing significantly more through:

• lost customers,
• negative word of mouth,
• declining renewal rates,
• and reduced revenue growth.

This is particularly risky in industries where trust and relationships matter deeply, such as healthcare, banking, insurance, telecom, and travel.

AI Should Optimize Both Efficiency and Experience

The next generation of AI systems cannot be designed solely around automation metrics.

Future-ready AI systems must balance:

• operational efficiency, with
• experience quality.

This requires a shift in thinking: from “How many interactions can AI eliminate?” to “How can AI improve outcomes while strengthening customer relationships?”

Organizations that succeed will likely treat CX not as a secondary consideration, but as a foundational design principle for AI systems. In the future, the most successful AI implementations may not be the ones that automate the most interactions —but the ones that create the most trusted, seamless, and contextually intelligent experiences.

AI Can Become a CX Enabler — Not Just a Cost Reduction Tool

The conversation around AI often assumes that automation and customer experience are competing priorities. They do not have to be.

Organizations have an opportunity to use AI not only to reduce costs, but also to elevate customer experience in ways that were previously difficult to scale economically.

One of the most promising approaches is using AI to optimize internal, non-customer-facing operations while redeploying the resulting capacity toward high-value human engagement.

For example:
AI can streamline back-office activities such as: documentation, workflow coordination, data validation, claim processing, scheduling, internal knowledge retrieval, and operational decision support.

This can significantly reduce administrative burden and free up employee bandwidth.

Instead of viewing these savings purely as workforce reduction opportunities, organizations can reinvest part of this newly available capacity into improving customer experience.

In healthcare, banking, insurance, and other service-intensive industries, this could enable:

• faster human assistance,
• proactive outreach,
• personalized guidance,
• reduced wait times,
• concierge-style support,
• and better handholding during complex journeys.

In many situations, customers do not necessarily want fewer human interactions. They want:

• fewer frustrating interactions,
• faster resolutions,
• and meaningful assistance when it matters most.

AI can help make this economically viable.

This creates a more balanced AI strategy:

• AI handles repetitive and operationally heavy work,
• while humans focus on empathy, trust, judgment, and relationship-building.

The result is not simply automation —it is augmentation of customer experience.

Organizations that adopt this mindset may discover that the real value of AI is not just cost efficiency, but the ability to deliver higher-quality experiences at scale.

Example: AI in Revenue Cycle Management

Consider a patient contacting a healthcare organization regarding a denied insurance claim.

A basic AI system may simply ask for claim details and follow a scripted workflow.

A CX-aware AI system, however, may understand:

• the patient has already called twice,
• the issue is delaying treatment,
• previous promises were missed,
• sentiment is increasingly frustrated,
• and the patient prefers proactive updates via text.

This changes how the AI behaves:

• it avoids repetitive questioning,
• prioritizes empathy,
• accelerates escalation,
• coordinates across systems,
• and proactively communicates next steps.

That is the power of CX as context.

How Can Organizations Build CX-Aware AI Systems?

Building CX-aware AI requires more than deploying a language model. It requires creating an integrated experience intelligence layer.

Some key capabilities may include:

Unified Journey Intelligence

Connecting interactions across: channels, touchpoints, business units, and systems.

Sentiment and Emotion Signals

Capturing behavioral and conversational cues that indicate: frustration, urgency, satisfaction, or confusion.

Persistent Memory

Allowing AI systems to maintain continuity across interactions rather than treating each engagement independently.

Context Orchestration

Dynamically combining: customer data, workflow data, operational signals, and experience signals
in real time.

Experience-Centric Governance

Defining not just what AI can do, but what experience it should deliver.

 

The Next Competitive Advantage

As foundational AI models become increasingly commoditized, competitive differentiation may shift toward:

• proprietary context,
• operational integration,
• and experience intelligence.

Organizations that best operationalize customer experience as a contextual layer may ultimately build AI systems that are not only more efficient, but also more trusted, empathetic, and effective.

The future of AI may therefore not be defined solely by intelligence.

It will be defined by contextual understanding — and customer experience could become its most important form.