Chapter 37: Emerging AI Technologies

"Your job is not to predict the future. Your job is to build an organization that can adapt to any future."

--- Professor Diane Okonkwo, MBA 7620: AI for Business Strategy

The Hype Cycle, Revisited

Professor Okonkwo puts a chart on the screen. It is Gartner's AI Hype Cycle from five years ago --- the 2021 edition.

"Look at the technologies at the Peak of Inflated Expectations," she says. "Generative AI. AI-augmented development. Composite AI. Autonomic systems."

She lets the class absorb it.

"Now think about what actually happened. Generative AI is real --- it reshaped entire industries. AI-augmented development is embedded in every engineering team worth its name. But autonomic systems? Still mostly a research concept. Composite AI? Quietly useful in a few enterprise platforms, but nobody writes breathless blog posts about it."

She clicks to the next slide. It is the same hype cycle from 2023. The Peak of Inflated Expectations is dominated by foundation models, generative AI for everything, and --- sitting right at the peak --- AI agents.

"Half the technologies at the peak are now either abandoned or quietly delivering value in the Plateau of Productivity," Okonkwo continues. "The other half are stuck in the Trough of Disillusionment. Every year, a new cohort of technologies climbs the peak. Every year, executives are tempted to bet their companies on whatever is at the top."

She looks at the class.

"You will be tempted. Your competitors will be. Some will be right. Most will waste millions. The skill isn't predicting which technologies will win. It's building the organizational capability to evaluate, pilot, and adopt new technologies systematically."

NK, sitting in her usual seat, has been thinking about this. Her marketing instincts --- honed through years of separating genuine consumer trends from fads --- tell her that the signal-to-noise ratio in AI technology is worse than in any field she has encountered. "How do you separate the signal from the noise?" she asks. "Because right now, every vendor pitch I see claims their technology is transformative."

"Excellent question," Okonkwo says. "And the answer is not intuition. It is framework. Ravi, would you like to introduce the class to Athena's approach?"

Ravi Mehta stands. He has been sitting in the back, as he usually does when he visits campus --- a practitioner among students, uncomfortable with the attention but generous with his experience. "At Athena, we use what we call the AI Technology Radar," he says. "It is shamelessly stolen from ThoughtWorks' Technology Radar, adapted for AI. And I can tell you, it has saved us from spending at least $30 million on technologies that were not ready."

He walks to the whiteboard and draws four concentric rings.

"The radar has four rings: Hold, Assess, Trial, and Adopt. Every emerging technology gets placed on the radar based on a structured evaluation. The ring tells the organization how much investment is appropriate."

Definition: An AI Technology Radar is a structured evaluation framework that categorizes emerging technologies into four rings: Hold (monitor but do not invest), Assess (investigate and build understanding), Trial (run bounded pilots with clear success criteria), and Adopt (deploy at scale with production support). The framework prevents both premature commitment and paralysis.

Tom leans forward. He has been itching to talk about agentic AI for three weeks. "Where does agentic AI sit on Athena's radar right now?"

Ravi pauses. "It just moved from Assess to Trial. And the reason it moved is not because the technology matured. It is because a competitor forced our hand."

The room goes quiet. Ravi has just introduced the elephant in the room --- the NovaMart crisis that has consumed Athena's leadership for the past three months.

"We will get to that," Okonkwo says. "But first, let us survey the landscape. Tom, NK --- I want you both thinking about each technology through two lenses. Tom, your lens is technical feasibility: can this actually work at enterprise scale? NK, your lens is business viability: if it works, does anyone care? Both lenses are necessary. Neither is sufficient."

Agentic AI: When AI Stops Waiting for Instructions

The most significant near-term shift in artificial intelligence is the transition from AI as a tool --- something you prompt, query, or invoke --- to AI as an agent: something that plans, reasons, uses tools, and executes multi-step tasks with varying degrees of autonomy.

What Agentic AI Actually Is

The concept is deceptively simple. Traditional AI systems, including the large language models we explored in Chapter 17, operate in a request-response paradigm. You ask a question; the model generates an answer. You provide a prompt; the model produces output. The human remains in the loop at every step, directing the interaction.

Agentic AI systems break this pattern. An AI agent receives a goal --- "Research the top five competitors in the sustainable packaging market and produce a comparative analysis with pricing data" --- and then autonomously plans the steps required, executes them (including using external tools like web browsers, databases, APIs, and code interpreters), evaluates its own progress, adjusts its approach when things go wrong, and delivers a finished result.

Definition: An AI agent is a system that uses a language model (or other AI) as its reasoning core, combined with the ability to plan multi-step tasks, use external tools, maintain memory across steps, and operate with some degree of autonomy. Agents differ from chatbots in that they act rather than merely respond.

The architecture of an agentic system typically includes:

• A reasoning engine (usually a large language model) that decomposes goals into subtasks
• Tool use --- the ability to invoke external functions: search the web, query a database, execute code, send an email, call an API
• Memory --- maintaining context across a long sequence of actions, including what has been tried, what failed, and what succeeded
• Planning and replanning --- the ability to create a multi-step plan and revise it when intermediate results deviate from expectations
• Self-evaluation --- checking whether the output actually satisfies the original goal

Current Capabilities and Limitations

By early 2026, agentic AI has moved from research demonstration to early enterprise deployment. The capabilities are genuine but bounded:

What agents can do reliably:

• Research and information synthesis (gathering information from multiple sources, summarizing, and producing structured reports)
• Code generation and debugging (writing, testing, and iterating on software --- see Case Study 1 on AI coding agents)
• Structured data workflows (extracting data from documents, populating databases, generating reports)
• Customer service escalation (handling multi-turn customer interactions that require accessing multiple systems)
• Scheduling and coordination (managing calendars, booking travel, coordinating logistics across systems)

What agents struggle with:

• Tasks requiring genuine judgment about ambiguous situations
• High-stakes decisions where errors have serious consequences (medical, legal, financial)
• Tasks that require understanding organizational context, politics, or unwritten norms
• Long-horizon planning (more than 10-15 sequential steps tend to accumulate errors)
• Recovery from unexpected situations that fall outside training distribution

Enterprise Applications

The business case for agentic AI centers on a simple observation: knowledge workers spend enormous amounts of time on structured, multi-step tasks that require judgment but not deep expertise. Competitive analysis. Vendor evaluation. Compliance checking. Data reconciliation. Report generation. These tasks are too complex for simple automation (they require reasoning and adaptation) but too routine to justify senior executive attention.

Agentic AI promises to compress these tasks from hours to minutes. A 2025 McKinsey analysis estimated that agentic AI could automate 25-35 percent of knowledge worker tasks by 2028 --- not by replacing workers, but by handling the structured portions of their work and freeing them for higher-judgment activities.

Business Insight: The business case for agentic AI is strongest in processes that are (1) multi-step, (2) involve structured information from multiple sources, (3) follow somewhat predictable patterns, and (4) currently require skilled but not uniquely expert human labor. If a task takes a competent analyst four hours and follows a recognizable pattern, it is a candidate for agentic automation.

NK raises her hand. "So is this the same as the workflow automation we covered in Chapter 21? Because it sounds like RPA with better language skills."

It is a perceptive question. The distinction matters. Robotic Process Automation (RPA), which we discussed in Chapter 21, follows rigid, predefined scripts: click this button, copy this field, paste it here. RPA breaks when the process changes. Agentic AI, by contrast, reasons about what to do next. It can adapt to unexpected data, try alternative approaches, and handle variations in the task. The difference is analogous to the difference between following a recipe step-by-step and cooking a meal when you understand the principles of cooking --- you can improvise when the recipe calls for an ingredient you don't have.

But NK's skepticism is well-placed. The hype around agentic AI in 2025-2026 echoes the hype around RPA in 2018-2019, when vendors promised that software robots would automate 40 percent of enterprise processes. The actual impact was significant but more modest, and adoption was slower than projected. History suggests similar caution is warranted with agentic AI.

Safety and Governance

Agentic AI introduces governance challenges that go beyond those we explored in Chapters 27-30. When an AI system acts --- placing orders, sending emails, modifying databases, executing code --- the stakes of errors increase dramatically. A chatbot that gives bad advice is a customer service problem. An agent that places an incorrect $500,000 purchase order is a financial problem.

Key governance requirements for agentic AI include:

1. Defined authority boundaries. What actions is the agent permitted to take without human approval? What actions require human confirmation? These boundaries must be explicit, configurable, and enforced by the system architecture --- not merely documented in a policy.

2. Audit trails. Every action an agent takes must be logged, along with the reasoning that led to the action. When things go wrong (and they will), the organization needs to understand what the agent did and why.

3. Kill switches. The ability to halt an agent's execution instantly, without data loss or corruption, is a non-negotiable safety requirement.

4. Scope limitations. Agents should be designed with the narrowest set of permissions necessary for their task. An agent that researches competitor pricing does not need write access to the procurement system.

5. Human-in-the-loop escalation. For decisions above a defined risk threshold, the agent should pause and request human review rather than proceeding autonomously.

Research Note: The alignment problem --- ensuring AI systems pursue the goals their operators intend, not distorted proxies of those goals --- becomes more acute in agentic systems. A customer service agent optimized to "resolve tickets quickly" might learn to close tickets without actually solving problems. Defining the right objectives for agentic systems is as important as the technical architecture. See Chapter 25 for background on how misaligned objectives create bias.

Multi-Agent Systems: Teams of AI

If a single agent can handle a multi-step task, what happens when you orchestrate multiple agents working together? This is the frontier of multi-agent systems --- and it is evolving rapidly.

The Architecture of Collaboration

In a multi-agent system, specialized agents with different capabilities collaborate to accomplish complex goals. Rather than building one monolithic agent that can do everything, you design a team:

• A research agent that gathers and synthesizes information
• An analyst agent that evaluates options against defined criteria
• A writer agent that produces polished output
• A reviewer agent that checks the work of other agents for errors, inconsistencies, or policy violations
• An orchestrator agent that coordinates the team, assigns tasks, and resolves conflicts

This mirrors how human organizations work. You do not hire one person who is simultaneously a researcher, analyst, writer, and quality assurance specialist. You build a team with complementary skills and a manager who coordinates.

Emerging Platforms

Several platforms have emerged to support multi-agent development:

Microsoft AutoGen provides a framework for building multi-agent conversations where agents with different roles interact to solve problems. Its core abstraction is the "conversable agent" --- an agent that can send and receive messages from other agents, enabling back-and-forth collaboration.

CrewAI takes a more structured approach, allowing developers to define agent "crews" with specific roles, goals, and backstories. Each agent in a crew has defined responsibilities and the framework manages task delegation, execution, and handoffs.

LangGraph (from LangChain) enables developers to build agentic workflows as directed graphs, where each node represents an agent action and edges define the flow of control. This approach provides more fine-grained control over agent behavior and supports complex branching and looping patterns.

Amazon Bedrock Agents and Google Vertex AI Agents offer managed platforms for building and deploying agents in cloud environments, with built-in integrations to enterprise data sources and tools.

Business Insight: The multi-agent space is evolving too rapidly for any specific platform to be a safe long-term bet. The underlying patterns --- role specialization, orchestration, tool use, and structured communication --- are more durable than any particular framework. When evaluating platforms, prioritize portability and avoid deep lock-in.

Business Applications

Multi-agent systems are most compelling for tasks that are too complex for a single agent but too structured for a fully human team:

• Due diligence automation. A team of agents researches a potential acquisition target: one gathers financial data, another analyzes legal filings, a third evaluates market position, and a reviewer agent checks for consistency and flags risks.
• Content production pipelines. A research agent gathers source material, a writer agent produces drafts, a fact-checker agent verifies claims, and an editor agent polishes the output.
• Supply chain optimization. Agents monitor different segments of a supply chain, share information, and collaboratively adjust inventory, routing, and supplier selections.

The honest assessment: multi-agent systems in early 2026 are impressive in demonstrations and useful in narrow, well-defined applications. They are not yet reliable enough for high-stakes, unsupervised enterprise deployment. The coordination overhead between agents --- miscommunication, conflicting actions, cascading errors --- remains a significant engineering challenge.

Edge AI and On-Device Inference

While agentic AI captures headlines, a quieter revolution is happening at the opposite end of the computing spectrum: AI is moving from the cloud to the device.

The Case for Edge AI

Edge AI refers to running AI models directly on end-user devices --- smartphones, IoT sensors, cameras, industrial equipment, vehicles --- rather than sending data to a cloud server for processing. The advantages are compelling:

Latency. A self-driving car cannot wait 200 milliseconds for a cloud server to decide whether the object in the road is a pedestrian. An industrial robot cannot pause for network latency before adjusting its grip. Edge inference happens in single-digit milliseconds.

Privacy. When a smart home camera processes video locally, the footage never leaves the device. When a health monitoring device analyzes biometric data on-device, it stays on the user's wrist. This architectural decision eliminates entire categories of privacy risk (see Chapter 29 for the privacy-by-design principles that edge AI embodies).

Bandwidth. A factory with 500 IoT sensors generating data continuously would overwhelm any network connection if all data were streamed to the cloud. Processing at the edge means only meaningful insights --- anomalies, alerts, summaries --- need to be transmitted.

Reliability. Edge AI works without an internet connection. A medical device in a rural clinic, a mining operation underground, a ship at sea --- all can use AI capabilities regardless of connectivity.

Cost. Cloud inference is cheap per query but expensive at scale. A retailer running computer vision on every security camera in 1,200 stores, 24 hours a day, will pay less for on-device inference than for cloud processing.

Model Compression and TinyML

The challenge is obvious: large language models have billions of parameters and require powerful GPUs. A smartphone has limited memory, processing power, and battery life. How do you fit a useful AI model on a resource-constrained device?

The answer involves several techniques, many of which we touched on in the context of model optimization in Chapter 13:

Quantization reduces the numerical precision of model weights --- from 32-bit floating point to 16-bit, 8-bit, or even 4-bit representations. This reduces memory requirements and speeds up inference with minimal accuracy loss. An 8-bit quantized model uses roughly one-quarter the memory of its 32-bit equivalent.

Knowledge distillation trains a smaller "student" model to mimic the behavior of a larger "teacher" model. The student cannot match the teacher on every task, but for a specific, well-defined task, it can come close at a fraction of the computational cost.

Pruning removes unnecessary connections (weights near zero) from neural networks, reducing model size without significantly affecting performance on the target task.

Architecture search designs model architectures optimized for specific hardware constraints from the ground up, rather than shrinking large models after the fact.

Definition: TinyML refers to machine learning models and techniques designed to run on microcontrollers and other extremely resource-constrained devices --- hardware with kilobytes (not gigabytes) of memory and milliwatts (not watts) of power consumption. TinyML enables AI in devices like hearing aids, environmental sensors, and agricultural monitors.

Business Applications

Edge AI is already generating business value in several domains:

• Retail. Computer vision for inventory monitoring, shelf compliance, and foot traffic analysis --- running on in-store cameras rather than streaming video to the cloud (see Chapter 15 for computer vision fundamentals).
• Manufacturing. Predictive maintenance models running on industrial equipment, detecting anomalies in vibration, temperature, or sound patterns and alerting operators before failures occur (see Chapter 16 for time series approaches).
• Healthcare. On-device analysis of medical images, ECG patterns, and wearable sensor data, enabling point-of-care diagnostics in settings without reliable internet.
• Agriculture. Drone-mounted AI that identifies crop diseases, pest infestations, and irrigation issues in real time during field surveys.
• Automotive. Advanced driver assistance systems (ADAS) and autonomous driving functions that process sensor data locally for real-time decision-making.

Athena Update: Ravi has been piloting edge AI in Athena's stores for the past year. Computer vision models running on in-store cameras provide real-time shelf monitoring --- detecting out-of-stock items, misplaced products, and pricing errors. The system processes video locally and sends only alerts and analytics summaries to Athena's central systems. "No customer video ever leaves the store," Ravi emphasizes. "That was a non-negotiable requirement from our governance team." On the AI Technology Radar, edge AI for in-store operations has moved from Trial to Adopt.

Small Language Models: When Less Is More

The dominant narrative in AI from 2020 to 2024 was "bigger is better." GPT-3 had 175 billion parameters. GPT-4 was rumored to use over a trillion. Each generation of frontier models was larger, more expensive, and more capable than the last.

That narrative is shifting. A counter-movement toward smaller, more efficient models --- Small Language Models (SLMs) --- is gaining momentum, driven by economics, performance, and practical necessity.

Why Small Models Matter

Cost. Running a 70-billion-parameter model costs roughly 10-20 times more per query than running a 7-billion-parameter model. For applications with millions of daily queries, the cost difference is measured in millions of dollars per year.

Speed. Smaller models generate responses faster. For applications where latency matters --- customer-facing chatbots, real-time coding assistance, interactive search --- response time is a competitive advantage.

Customization. Smaller models are easier to fine-tune on domain-specific data. A 7-billion-parameter model can be fine-tuned on a single GPU in hours. Fine-tuning a frontier model requires specialized infrastructure and, often, partnership with the model provider.

Deployment flexibility. Small models can run on edge devices, in private data centers, or on modest cloud instances. They do not require the massive GPU clusters that frontier models demand.

Privacy. Running a small model locally means data never leaves the organization's infrastructure. For regulated industries --- healthcare, financial services, defense --- this is often a hard requirement.

The SLM Landscape

Several families of small language models have demonstrated surprising capability:

Microsoft Phi series (Phi-2 at 2.7B parameters, Phi-3 at 3.8B and 14B) achieved benchmark scores that rivaled models many times their size, particularly on reasoning and coding tasks. The Phi series demonstrated that training data quality and curriculum design could partially compensate for reduced scale.

Google Gemma (2B and 7B parameters) offered an open-weights alternative for lightweight deployment, with particular strength in multilingual applications.

Meta Llama 3 (8B and 70B variants) provided open-source models that community researchers and enterprises could fine-tune freely --- a significant advantage for organizations that need customization without vendor dependency.

Mistral (7B and Mixtral 8x7B) introduced mixture-of-experts architectures to the open-source community, achieving strong performance by activating only a subset of parameters for each query.

When Small Is Better

The critical business question is not "Is a small model as good as GPT-4 or Claude?" The answer to that question is usually no. The critical question is: "Is a small model good enough for this specific task?"

For many enterprise applications, the answer is yes:

• Intent classification (routing customer inquiries to the right department) does not require frontier-model reasoning.
• Named entity extraction (pulling dates, names, and amounts from documents) is a well-defined task where small models excel.
• Summarization of structured documents (meeting notes, reports, emails) can be handled effectively by 7B-parameter models fine-tuned on domain data.
• Code completion for specific programming languages and frameworks benefits more from domain-specific training than raw scale.

Business Insight: The decision between a frontier model and a small language model is not a technology decision --- it is a business decision. Consider the total cost of ownership (inference costs, hosting, fine-tuning, maintenance), the required accuracy threshold, the latency requirements, the privacy constraints, and the customization needs. For many production applications, a fine-tuned 7B model outperforms a general-purpose frontier model at one-tenth the cost. See Chapter 11 for model evaluation frameworks.

Tom, who has been practically vibrating with enthusiasm, finally speaks up. "This is the part that excites me. The democratization angle. A startup with $50,000 in compute budget can now build a competitive AI product by fine-tuning a small open model on their domain data. That was impossible two years ago."

Okonkwo nods. "And what is the strategic implication of that?"

Tom thinks. "That the competitive advantage shifts. It's not about having the biggest model anymore. It's about having the best data for your specific use case --- and the organizational capability to fine-tune, deploy, and maintain a model."

"Exactly. Which brings us back to a theme from Chapter 4: data as a strategic asset."

Quantum Computing: A Reality Check

Few topics in technology generate as much confusion as quantum computing's relationship to artificial intelligence. Headlines promise that quantum computers will supercharge machine learning, break encryption, and solve optimization problems that are intractable for classical computers. The reality is considerably more nuanced.

What Quantum Computers Actually Do

A classical computer processes information in bits --- ones and zeros. A quantum computer uses qubits, which can exist in a superposition of states, enabling certain types of parallel computation that are fundamentally impossible on classical hardware.

The key word is "certain." Quantum computers are not universally faster than classical computers. They offer dramatic speedups for specific categories of problems:

• Factoring large numbers (relevant to cryptography, via Shor's algorithm)
• Searching unsorted databases (Grover's algorithm provides a quadratic speedup)
• Simulating quantum systems (useful for drug discovery, materials science, and chemistry)
• Certain optimization problems (portfolio optimization, logistics routing, combinatorial problems)

For most computational tasks --- including the matrix multiplications that dominate deep learning --- quantum computers offer no advantage over classical hardware.

Quantum Machine Learning: Where We Actually Are

Quantum machine learning (QML) is an active research area exploring whether quantum computing can accelerate or improve machine learning algorithms. The honest assessment as of early 2026:

What has been demonstrated:

• Quantum algorithms for specific linear algebra operations (like solving linear systems via the HHL algorithm) that are theoretically faster than classical equivalents
• Quantum kernel methods for small classification problems
• Quantum-inspired classical algorithms that borrow concepts from quantum computing but run on classical hardware (and sometimes outperform the quantum originals)
• Variational quantum circuits that can learn simple patterns on small datasets

What has not been demonstrated:

• Any quantum machine learning algorithm outperforming classical methods on a practically useful problem at a meaningful scale
• Quantum advantage for training neural networks
• Scalable quantum hardware capable of running the algorithms that theorists have designed

What Business Leaders Should Do

The honest advice for most organizations:

1. Do not invest in quantum computing capability today unless you are in pharmaceuticals, materials science, financial services (certain optimization problems), or logistics --- the sectors where quantum advantages are most likely to materialize first.

2. Monitor the landscape. The technology is evolving, and breakthroughs --- while unpredictable --- are possible. Assign one person (not a team) to track quantum computing developments quarterly.

3. Protect against quantum threats. Quantum computers will eventually break current encryption standards (RSA, ECC). Begin transitioning to post-quantum cryptography now --- not because quantum computers can break your encryption today, but because the transition takes years and sensitive data encrypted today could be stored and decrypted later ("harvest now, decrypt later" attacks). See Chapter 29 for encryption fundamentals.

4. Be skeptical of vendors selling quantum AI. If a vendor claims their quantum computing solution will transform your machine learning today, they are selling hype. The technology is not there yet.

Research Note: The National Institute of Standards and Technology (NIST) released its first post-quantum cryptography standards in 2024, providing concrete guidance for organizations beginning the transition. IBM, Google, and several startups are leading quantum hardware development, but the timeline for "quantum utility" (where quantum computers consistently outperform classical computers on useful tasks) remains uncertain.

NK nods approvingly. "So quantum computing goes in the Hold ring on the radar?"

Ravi confirms. "Hold. We track it quarterly. We have started our post-quantum cryptography transition. But we are not building any AI capabilities on quantum hardware. The technology risk is too high and the timeline too uncertain for a retailer."

Neuromorphic Computing: Computing Like the Brain

If quantum computing reimagines computing at the physics level, neuromorphic computing reimagines it at the architectural level --- designing chips that process information the way biological brains do rather than the way traditional processors do.

How It Differs

Conventional computer processors are based on the Von Neumann architecture: memory and processing are separate, connected by a bus. Data shuttles back and forth between storage and processor. This architecture works beautifully for sequential computation but is inherently inefficient for the massively parallel, event-driven processing that characterizes neural computation.

The human brain, by contrast, processes information through approximately 86 billion neurons connected by roughly 100 trillion synapses. There is no separation between memory and processing --- computation happens in the connections. The brain consumes about 20 watts of power. A data center running an equivalent-scale neural network consumes megawatts.

Neuromorphic chips attempt to close this efficiency gap by mimicking the brain's architecture:

• Spiking neural networks process information through discrete events (spikes) rather than continuous numerical values, enabling event-driven computation that is extremely power-efficient
• Co-located memory and processing eliminate the Von Neumann bottleneck
• Massive parallelism enables thousands of simple operations to occur simultaneously

The Hardware Landscape

Intel Loihi 2 is the most prominent research neuromorphic chip, with approximately 1 million artificial neurons. Intel has demonstrated applications in robotic control, odor recognition, and optimization problems, achieving 10-100x energy efficiency improvements over conventional approaches for specific tasks.

IBM TrueNorth (now succeeded by the NorthPole architecture) demonstrated that neuromorphic design could achieve dramatic efficiency gains for pattern recognition tasks.

SynSense and BrainChip (Akida) are commercializing neuromorphic processors for edge AI applications --- particularly in always-on, low-power scenarios like keyword detection, gesture recognition, and sensor processing.

Business Relevance

Neuromorphic computing is genuinely early-stage for most business applications. Its most promising near-term use cases are in edge AI scenarios where extreme power efficiency is critical:

• Always-on sensor processing (environmental monitoring, security cameras, wearable devices)
• Robotics (real-time sensory processing with minimal power consumption)
• Autonomous systems (drones, vehicles, industrial robots)

For enterprise AI workloads --- training large models, running inference on language models, processing business data --- neuromorphic hardware offers no current advantage.

Business Insight: Neuromorphic computing is a technology to track, not a technology to invest in, for most businesses. Its relevance will grow as edge AI expands and as energy costs become a larger fraction of AI budgets. Place it firmly in the Hold ring unless your business depends on ultra-low-power, always-on sensing.

Hardware Economics: The Engine Room of AI

Every AI model, regardless of its elegance, runs on hardware. And the economics of that hardware --- who makes it, who can access it, how much it costs --- shape the AI landscape as profoundly as any algorithm.

The GPU Bottleneck

From 2023 through 2025, the most significant constraint on AI development was not algorithmic --- it was the global shortage of high-end GPUs, particularly NVIDIA's H100 and its successors. Wait times for H100 clusters stretched to 6-12 months. Prices on secondary markets exceeded $40,000 per chip. Cloud GPU instance costs spiked by 50-100 percent.

This shortage had profound business implications:

• Startups that could not secure GPU access could not train competitive models, effectively creating a hardware barrier to entry
• Enterprises that had not pre-ordered GPU capacity found themselves unable to launch AI initiatives on their planned timelines
• Cloud providers (AWS, Azure, GCP) engaged in a GPU arms race, committing tens of billions of dollars to NVIDIA purchases
• National governments began treating advanced AI chips as strategic assets, with the US restricting GPU exports to China

The Custom Chip Landscape

The GPU shortage accelerated investment in alternative AI hardware:

Google TPUs (Tensor Processing Units) are custom-designed for machine learning workloads. Google uses TPUs internally for Search, YouTube, and its Gemini models, and offers them to external customers through Google Cloud. TPUs offer competitive performance for training and inference at costs that can undercut NVIDIA GPUs for specific workloads.

Amazon Trainium and Inferentia are AWS's custom training and inference chips, respectively. Amazon designed these to reduce its own dependence on NVIDIA and to offer customers a lower-cost alternative. Early benchmarks suggest 30-50 percent cost savings over equivalent GPU instances for supported model architectures.

Groq built a chip architecture (the Language Processing Unit, or LPU) optimized specifically for inference speed. Groq's chips achieve dramatically faster inference than GPUs for certain model architectures, though their training capabilities are limited.

Cerebras built wafer-scale chips --- single chips the size of an entire silicon wafer --- optimized for training large models. The approach is radical and the hardware is expensive, but it eliminates many of the communication bottlenecks that slow down distributed GPU training.

Definition: Inference is the process of running a trained model on new data to generate predictions or outputs. Training is the process of building the model by learning from data. The hardware requirements for training (massive parallelism, large memory) differ from those for inference (throughput, latency, cost per query). Many alternative chips target inference optimization because inference workloads dominate production costs.

The Cost Trajectory

The economics of AI compute are evolving in two seemingly contradictory directions:

Training costs for frontier models are rising. GPT-4 reportedly cost over $100 million to train. Frontier models in 2025-2026 may cost $500 million to $1 billion. Only a handful of organizations --- OpenAI, Google, Anthropic, Meta, a few others --- can afford to build foundation models from scratch.

Inference costs are plummeting. The cost per token for using a large language model dropped roughly 90 percent between 2023 and 2025, and continues to fall as model efficiency improves, hardware competition increases, and economies of scale take effect.

For most businesses, this is excellent news. You do not need to train a foundation model. You need to use one --- through APIs, fine-tuned variants, or self-hosted open models. And the cost of usage is falling rapidly.

Business Insight: Hardware economics matter because they determine who can build AI and who can only use it. The training cost barrier means that foundation model development is consolidating among a handful of firms. But falling inference costs mean that the application of AI is democratizing rapidly. The strategic question for most businesses is not "Can we build a foundation model?" (no) but "Can we apply foundation models more effectively than our competitors?" (yes, with the right data, processes, and organizational capability).

Tom, who spent three years working with cloud infrastructure, sees the implications immediately. "This is why the build-vs-buy decision we discussed in Chapter 6 keeps evolving. The 'build' option used to mean training your own model. Now 'build' means fine-tuning an open model on your data and hosting it on your own infrastructure. 'Buy' means using an API. And the economics keep shifting."

Okonkwo nods. "Which is precisely why your AI strategy cannot be a static document. It must be a living framework that adapts as the cost structure changes."

Open-Source vs. Closed Models: The Great Debate

The AI industry is split between two fundamentally different approaches to model distribution, and the choice between them carries significant strategic implications for every organization deploying AI.

The Landscape

Closed (proprietary) models --- including OpenAI's GPT-4 and successors, Anthropic's Claude, and Google's Gemini --- are accessible only through APIs. Users can prompt the model and receive outputs, but cannot examine the model's architecture, weights, or training data. The provider controls pricing, access, and the model's capabilities.

Open-weight models --- including Meta's Llama series, Mistral's models, Alibaba's Qwen, and Google's Gemma --- make the model weights publicly available. Users can download the model, run it on their own hardware, fine-tune it on their own data, and modify it as they see fit. (Strictly speaking, most "open-source" AI models are more accurately described as "open-weight" --- the weights are released, but the training data, training code, and full reproduction pipeline often are not.)

The Tradeoffs

Strategic Implications

The right choice depends on context, and many organizations will use both:

Use closed models when: - You need frontier-level capability (the most sophisticated reasoning, longest context windows, multimodal processing) - You want minimal infrastructure burden - Your volume is moderate (thousands, not millions, of queries per day) - You are prototyping and need speed to market

Use open models when: - Data privacy is a hard requirement (regulated industries, sensitive data) - You need deep customization for a specific domain or task - Your inference volume is high enough that API costs become significant - You need to run AI in environments without reliable internet connectivity - Vendor dependency is a strategic risk your organization is unwilling to accept

AI and Robotics: Embodied Intelligence

For decades, AI and robotics were separate fields. AI researchers focused on software intelligence --- perception, reasoning, language. Robotics researchers focused on hardware --- motors, sensors, manipulation, locomotion. The convergence of these fields is producing a new category of technology: embodied AI, where intelligent software controls physical systems in the real world.

The Current State

Warehouse and logistics automation is the most commercially mature application of AI robotics. Amazon operates over 750,000 robots across its fulfillment network. These systems handle picking, packing, sorting, and transportation within warehouses, working alongside human workers. The robots use computer vision (Chapter 15) to identify items, reinforcement learning to optimize movement paths, and multi-agent coordination to avoid collisions and balance workloads.

Cobots (collaborative robots) are designed to work safely alongside humans on factory floors. Unlike traditional industrial robots --- which operate in caged enclosures because they are dangerous --- cobots use force sensors, computer vision, and AI-based path planning to work in shared spaces. Universal Robots, FANUC, and ABB are leading cobot manufacturers, with applications in assembly, quality inspection, and material handling.

Autonomous vehicles remain the highest-profile (and most humbling) application of embodied AI. After over a decade of development and tens of billions of dollars in investment, fully autonomous vehicles operate commercially only in limited geographies under restricted conditions. Waymo operates robotaxi services in Phoenix, San Francisco, and Los Angeles. Cruise (GM) suspended operations in 2023 after safety incidents. Tesla's Full Self-Driving remains a Level 2 system (requiring constant human supervision). The lesson: embodied AI in unstructured environments (public roads) is dramatically harder than embodied AI in structured environments (warehouses).

Humanoid robots are the most speculative frontier. Companies including Tesla (Optimus), Figure, Agility Robotics (Digit), and 1X Technologies are developing general-purpose humanoid robots for warehouse, manufacturing, and household tasks. The pitch: a humanoid form factor can navigate environments designed for humans --- stairs, doors, shelves --- without requiring infrastructure modifications. The reality: these robots are in early prototype stages, capable of basic tasks (walking, picking up objects, folding clothes) in controlled demonstrations but far from reliable deployment.

Business Insight: For most businesses, the relevant robotics question is not about humanoid robots. It is about automation of specific physical tasks: warehouse picking, quality inspection, material handling, last-mile delivery. These applications are commercially proven, economically viable, and getting better rapidly. Evaluate robotics investments based on specific process economics, not science fiction ambitions.

NK, who has been listening carefully, asks: "What is the timeline for humanoid robots that actually work in a business setting?"

Ravi is characteristically blunt. "Five to ten years for narrow tasks in structured environments. Twenty or more years for general-purpose robots that can do what a human worker does. And those are optimistic estimates."

Tom is more charitable. "The hardware is improving fast. The issue is the software --- getting AI to handle the unpredictability of the physical world. But transfer learning from simulation to reality is getting better, and embodied foundation models are a real research direction."

Okonkwo mediates. "Both assessments are valid. As a business leader, plan for the five-year timeline, not the twenty-year one. Invest in automation that works today. Monitor the frontier for opportunities."

Synthetic Data: Training AI on AI

As AI models have grown larger and more capable, a paradox has emerged: the demand for training data is outpacing the supply of real-world data. This has driven rapid advances in synthetic data --- artificially generated data used to train AI systems.

Why Synthetic Data Matters

Data scarcity. Some domains have inherently limited data. Rare diseases have few cases. Financial fraud is (fortunately) rare. Autonomous driving scenarios involving crashes happen infrequently. Synthetic data can generate millions of examples of rare events.

Privacy. Real medical records, financial transactions, and customer interactions are subject to privacy regulations. Synthetic data that preserves the statistical properties of real data without containing any actual personal information can enable AI development without privacy risk (see Chapter 29).

Bias correction. If historical data reflects biased outcomes --- as we explored in Chapter 25 --- synthetic data can be generated to create more balanced training sets. This does not eliminate bias (the generation process itself can introduce biases), but it provides a mechanism for controlled adjustment.

Cost. Labeling real data is expensive and slow. Human annotators cost $15-50 per hour. Labeling a million images for computer vision training might cost $500,000 or more. Synthetic data can be generated and labeled automatically.

Techniques

Generative Adversarial Networks (GANs) produce realistic synthetic data by training two neural networks against each other --- a generator that creates synthetic data and a discriminator that tries to distinguish synthetic from real. The result is synthetic data that is statistically indistinguishable from real data in many applications.

Diffusion models (the technology behind Stable Diffusion and DALL-E) generate high-quality synthetic images, and are increasingly used to create training data for computer vision systems.

Large language models generate synthetic text data for training NLP models --- customer service conversations, product reviews, medical notes. The quality has improved dramatically, though verifying the accuracy of LLM-generated training data remains a challenge.

Simulation environments generate synthetic data for robotics and autonomous systems. A simulated warehouse, factory, or city street can produce millions of training scenarios in hours, including rare edge cases that would take years to encounter in the real world.

Self-play trains AI systems by having them compete against themselves, generating their own training data through interaction. This technique, pioneered by DeepMind's AlphaGo and AlphaZero, has been extended to language models and other domains.

The Quality Problem

The critical challenge with synthetic data is verification. If you generate a million synthetic customer service conversations to train a chatbot, how do you know they accurately represent real customer behavior? If you generate synthetic medical images to train a diagnostic AI, how do you verify that the synthetic images reflect actual pathology?

This is not merely a theoretical concern. Research has documented cases where models trained on synthetic data performed well on synthetic test sets but poorly on real-world data --- a form of overfitting to the synthetic data distribution.

Research Note: A 2024 study by Shumailov et al. in Nature demonstrated that training language models on outputs from other language models (a process they called "model collapse") can cause progressive degradation in quality over successive generations. As the internet fills with AI-generated content, this creates a growing risk for future model training. The finding underscores the enduring value of high-quality, human-generated data.

The Competitive Landscape Evolution

Each technology we have surveyed creates specific disruption patterns and strategic opportunities. Understanding these patterns requires connecting the technology assessment to the competitive dynamics we explored in Chapter 31.

How Emerging Technologies Create Disruption

The pattern is consistent across technology generations:

1. A new technology enables a new capability (e.g., agentic AI enables autonomous multi-step workflows)
2. A fast-moving competitor deploys the capability before incumbents do
3. The capability changes customer expectations (customers who experience an AI shopping agent expect it from every retailer)
4. Incumbents face a choice: respond with their own deployment or differentiate along a dimension the new technology does not address
5. The window for response is measured in months, not years, because customer switching costs in digital channels are low

This is precisely the dynamic playing out at Athena.

Athena Update: Three months ago, NovaMart launched an AI-powered personal shopping agent --- a system that browses products across retailers, compares options based on the customer's stated preferences and purchase history, negotiates for the best available price, and completes purchases on the customer's behalf. The agent is built on a frontier language model with agentic capabilities, integrated with NovaMart's product catalog, pricing engine, and payment system.

The impact was swift. NovaMart captured 8 percent of Athena's online market share in twelve months. Customer exit surveys revealed the appeal: "It saves me two hours of comparison shopping." "It finds deals I would never have found on my own." "It remembers my sizes, my preferences, my budget."

Grace Chen, Athena's CEO, convened an emergency strategy session last week. The board is demanding a response. Ravi has been tasked with evaluating the options.

Athena's Technology Radar in Action

Ravi uses the AI Technology Radar to structure the evaluation:

Agentic AI --- moved from Assess to Trial.

Ravi's analysis: "An AI shopping agent is technically feasible for Athena. We can build one in 6-9 months using a frontier language model as the reasoning engine, connected to our product catalog, customer data platform, and payment system. The investment is significant --- roughly $12-15 million in development plus $4-6 million in annual operating costs. But the competitive threat justifies it."

The governance challenge: Athena's AI governance framework, built over the past two years (Chapters 27-30), requires that any AI system making purchasing decisions on behalf of customers meet specific standards for transparency, accuracy, and recourse. NovaMart's agent, by contrast, appears to operate with minimal governance guardrails --- a faster path to market but a potential liability.

Edge AI for stores --- moved from Trial to Adopt.

The shelf monitoring pilot has been running for nine months with strong results: 23 percent reduction in out-of-stock incidents, 15 percent improvement in planogram compliance, and significant cost savings from automated inventory checks.

Quantum computing --- remains at Hold.

No current business case. Post-quantum cryptography transition is underway as a security initiative (led by IT, not the AI team).

Neuromorphic computing --- Hold.

Interesting for future edge AI applications but no current commercial hardware that meets Athena's requirements.

Small language models --- moved from Assess to Trial.

Ravi is piloting a fine-tuned 7B-parameter model for internal customer service routing --- classifying incoming inquiries and routing them to the appropriate department. The small model handles this task with 94 percent accuracy at one-fifteenth the cost of the API-based frontier model it replaces.

Athena Update: After two days of intensive debate, the leadership team reaches a decision. Athena will build a competitive AI shopping assistant. But it will build one that reflects Athena's values: transparent about its reasoning ("I recommend this product because..."), honest about its limitations ("I cannot verify this seller's return policy"), and designed with governance guardrails that NovaMart lacks.

"We will not win the race to market," Grace Chen tells the board. "NovaMart has a twelve-month head start. But we can win the race to trust. Customers will eventually care about whether their AI shopping agent is looking out for them or for the retailer. We will build the one that looks out for them."

Tom Kowalski, sitting in Okonkwo's class, does not know any of this yet. But he is about to learn. After class, Ravi approaches him. "I have a proposition," Ravi says. "We are building something ambitious. And we could use someone who understands both the technology and the business case."

Tom's eyebrows rise. "Are you offering me a job?"

"A consulting engagement. Twelve months. You would help architect Athena's AI shopping assistant while finishing your MBA. Interested?"

Tom does not hesitate. "When do I start?"

Building Organizational Readiness for Emerging AI

The final --- and most important --- section of this chapter is not about any specific technology. It is about building the organizational capability to evaluate, adopt, and integrate new technologies as they emerge. Technologies will change. This capability is durable.

The Technology Radar in Practice

Ravi's AI Technology Radar is not just a classification tool. It is an organizational process:

Quarterly scanning. Every quarter, Athena's AI team scans the technology landscape: academic papers, industry reports, competitor announcements, vendor presentations, conference proceedings. The goal is not comprehensive knowledge --- it is early awareness.

Structured evaluation. For technologies that warrant attention, a brief (2-3 page) evaluation document answers five questions:

1. What is it? A clear, jargon-free description of the technology and what it enables.
2. How mature is it? Research prototype, early commercial, proven at scale?
3. What is the business case for our organization? Specific use cases, estimated ROI, competitive implications.
4. What are the risks? Technical risk (it might not work), organizational risk (we might not be ready), ethical risk (it might create governance issues).
5. What is the recommended action? Hold, Assess, Trial, or Adopt --- with specific next steps for each.

Trial with guardrails. Technologies that move to Trial get a bounded pilot: defined scope, defined timeline (typically 90 days), defined success criteria, defined budget, and a pre-committed decision process (at the end of the trial, we will either stop, expand, or move to Adopt based on these specific metrics).

Adopt with support. Technologies that move to Adopt get production infrastructure, dedicated engineering support, monitoring, and governance integration.

Try It: Build an AI Technology Radar for your organization (or a hypothetical organization). Select five emerging AI technologies from this chapter. For each, write a one-page evaluation answering the five questions above. Place each technology on the radar (Hold, Assess, Trial, Adopt) and justify your placement. This exercise will be central to the capstone project in Chapter 39.

Avoiding the Two Failure Modes

Organizations fail with emerging technology in two symmetric ways:

Failure Mode 1: Chasing every trend. The organization launches pilots for every new technology, spreads resources too thin, and never achieves production deployment of anything. Resources are consumed by experimentation; value is never delivered. This is "shiny object syndrome" --- and it is endemic in organizations where AI strategy is driven by enthusiasm rather than discipline.

Failure Mode 2: Waiting too long. The organization demands proof that a technology works before investing, waits for case studies from other companies, and enters the market only when the technology is mature and competitors have established advantages. By the time they act, the window has closed. This is "analysis paralysis" --- and it is equally endemic in organizations where AI strategy is driven by risk aversion.

The Technology Radar provides a middle path: structured experimentation with clear criteria for advancement or abandonment. Not every technology deserves a trial. Not every trial deserves to become a production deployment. But the organization maintains the muscle of evaluation and experimentation, so that when the right technology appears at the right moment, it can move quickly.

Staying Current Without Drowning

For individual business leaders, the challenge of staying current with AI technology is equally daunting. The field moves faster than any individual can track. A pragmatic approach:

1. Follow three to five high-quality sources. Not Twitter/X, not LinkedIn influencers, not vendor blogs. Reliable sources: MIT Technology Review, The Gradient (Stanford's AI publication), the Import AI newsletter (by Jack Clark), the AI-focused reporting at The Information and Bloomberg, and Gartner/McKinsey/BCG research reports on enterprise AI.

2. Attend one conference per year. NeurIPS, ICML, or a business-focused AI conference (AI World, Transform). The goal is not to understand every paper but to sense the direction of the field.

3. Build a personal advisory network. Know two or three people --- an academic researcher, a startup founder, an enterprise AI practitioner --- who can provide context when you encounter a new technology. A ten-minute phone call with a knowledgeable person is more valuable than ten hours of reading hype cycles.

4. Experiment personally. Use the tools. Build a small project with an AI agent framework. Fine-tune a small language model. Run an edge AI demo on a Raspberry Pi. Personal experience is the best inoculation against hype --- once you have built something, you understand both the power and the limitations.

5. Teach what you learn. Explain new technologies to colleagues who are not following the field. If you cannot explain it clearly, you do not understand it well enough to make strategic decisions about it.

Business Insight: The goal of staying current is not encyclopedic knowledge. It is informed judgment. You need to know enough to ask good questions, evaluate vendor claims, and participate meaningfully in technology strategy discussions. You do not need to know how to implement every technology yourself. That is what your technical team is for.

Chapter Summary

Tom is buzzing after class. The consulting engagement with Athena. The chance to architect an AI shopping assistant. The opportunity to apply everything he has learned.

NK watches him practically float out of the lecture hall and shakes her head. "He is going to work twenty hours a day and love every minute of it."

But NK is thinking about something different. She is thinking about the Technology Radar --- about the discipline of evaluation rather than the thrill of technology. She is thinking about how, in her previous career in marketing, she watched brands chase every social media trend and burn through budgets with nothing to show for it. She is thinking about how the pattern is identical in AI.

"Professor," NK says, as the lecture hall empties, "I want to learn the radar process. Not just the technology. The process of deciding what to bet on and what to ignore."

Okonkwo smiles. "That, Ms. Adeyemi, is the most valuable skill in this chapter. More valuable than understanding any individual technology."

"Because the technologies will change."

"Because the technologies will change."

The technologies surveyed in this chapter --- agentic AI, multi-agent systems, edge AI, small language models, quantum computing, neuromorphic computing, alternative hardware, open-source models, robotics, and synthetic data --- represent the current frontier of AI capability. Some will transform business operations within the next two to three years. Some will take a decade. Some will disappoint.

The common thread is not any specific technology. It is the organizational capability to evaluate, experiment with, and adopt new technologies systematically. The AI Technology Radar is one framework for building that capability. The capstone project in Chapter 39 will ask you to build your own.

But technology does not exist in a vacuum. Every technology in this chapter has societal implications --- for employment, for inequality, for privacy, for power concentration, for the environment. Chapter 38 will examine those implications with the same rigor and balance we applied to the technology itself.

The future of AI in business is not determined by algorithms. It is determined by the choices that business leaders --- people like you --- make about how to deploy them.

In Chapter 38, we will examine the societal dimensions of AI: the future of work, the evidence on job displacement and augmentation, the environmental costs of AI, and the question of whether AI's economic benefits will be broadly shared or narrowly concentrated. The Athena story continues with the workforce implications of the AI shopping assistant --- and the tension between competitive necessity and organizational responsibility.