Two observations frame everything that follows. First, AI adoption has crossed a decisive threshold in 2024–2026. Second, value capture has not. The gap between the two is the central managerial puzzle of the era.

1.1 Two graphs every student should know

The McKinsey State of AI survey, fielded annually since 2017, is the longest continuous series of large-sample enterprise AI adoption data. Its 2025 wave reports adoption figures that would have been called fanciful only three years earlier.

The first three figures come from McKinsey's 2025 State of AI (n = 1,491). The last figure — that only about 1% of companies believe they have reached AI maturity — comes from McKinsey's deeper enterprise AI work (Lamarre et al., 2023; subsequent State of AI publications). The combined picture is unambiguous: adoption has gone mainstream while value capture has not.

1.1.1 What does the gap look like in practice?

Gartner's 2024 forecast that 30% of generative AI projects would be abandoned after proof of concept by end of 2025 — owing to escalating costs of $5–20 million per deployment and unclear value — has been broadly vindicated by Deloitte's Q4 2024 State of GenAI in the Enterprise survey, which found two-thirds of executives reporting that no more than 30% of their experiments will scale within 3–6 months. Gartner's 2025 follow-up extends the same finding to agentic AI: more than 40% of agentic AI projects will be cancelled by end of 2027.

A still more pointed figure: a 2025 MIT Sloan study cited in the trade press found that 78% of organisations using traditional change-management approaches for AI implementation are struggling with adoption beyond basic use cases. The headline McKinsey statistic that “70% of digital transformations fail” has been quoted so often that it has become almost folklore — but the underlying empirical pattern is robust.

1.1.2 The McKinsey 2026 update: top quintile pulling away

McKinsey's January 2026 AI Transformation Manifesto, drawing on engagements across 50+ enterprises in 2025, reports that the gap between leaders and laggards is widening. Top-quintile firms are now capturing 16–30% productivity improvements in functional areas (software engineering, customer service, operations, knowledge work), and 31–45% improvements in software quality. For the bulk of firms in the middle three quintiles, productivity gains remain in the low single digits — which is to say, indistinguishable from generally-improving software tooling.

1.2 What this book treats as “AI in business”

The phrase is contested. We use it inclusively to mean any computational system that performs tasks normally requiring human cognition and is deployed in a commercial or organisational context. This deliberately covers:

• Rule-based expert systems (DENDRAL, MYCIN, XCON, INTERNIST-1, PROSPECTOR) — historically important even if rarely called “AI” in the deep-learning era.
• Classical statistical learning (FICO Falcon fraud detection, Amazon collaborative filtering, Netflix Prize, Google PageRank) — quietly powering most enterprise AI value through 2020.
• Deep learning (AlexNet, ResNet, AlphaGo, AlphaFold, BERT) — the 2012–2022 wave.
• Foundation models and generative AI (GPT, Claude, Gemini, Llama, DeepSeek, Qwen, Mistral) — the post-2022 wave.
• Agentic systems (AutoGPT, Operator, Agentforce, Computer Use, Devin) — the 2024+ wave.

Drawing the line generously matters because most of the value in operational enterprises today still comes from the second and third categories — credit scoring, demand forecasting, recommender systems, image classification — even as headlines and budgets concentrate on the fifth.

1.3 Iansiti and Lakhani: the runtime argument

Iansiti and Lakhani's central claim — that AI is becoming the “runtime” of the firm in the sense Satya Nadella means when he says “AI is the runtime that is going to shape all of what we do” — frames the textbook's argument. When a business is driven by AI, software instructions and algorithms make up the critical path in the way the firm delivers value. Humans may have designed the operational systems, but computers are doing the work in real time: setting a price on Amazon, recommending a film on Netflix, qualifying a borrower for an Ant Group loan, painting the digital Rembrandt at ING.

The implications follow. Digital, AI-driven processes are more scalable than traditional processes — adding the next user adds nearly zero operational complexity. They enable greater scope (variety) because they easily connect with other digitised businesses. And they create powerful learning opportunities through ever more accurate, complex, and sophisticated predictions. In doing so, networks and AI are reshaping the operational foundations of firms, enabling digital scale, scope, and learning — the three properties Iansiti and Lakhani argue erase deep-seated limits that have constrained firm growth for hundreds of years.

1.4 The argument in one paragraph

1.5 What is genuinely new in 2024–2026

Despite the insistence that the bottleneck is organisational, the technology curve really has steepened. Five facts deserve to be memorised:

1. Inference costs have fallen approximately 280× in under three years (Stanford AI Index 2025), driven by smaller models, distillation, quantisation, and hardware specialisation.
2. Open-weight models now match closed frontier models on many benchmarks — DeepSeek-R1 (January 2025, MIT-licensed) matched OpenAI's o1; Llama 4 (April 2025) is competitive with GPT-4-class models; Qwen 2.5 and Mistral Large 2 followed similar trajectories.
3. Multimodality is the default — GPT-4o (May 2024), Gemini 2.5 Pro (March 2025), Claude 4 (May 2025) combine vision, language, code, and audio in single models.
4. Agentic capability has crossed a usability threshold — OpenAI Operator scored 87% on WebVoyager in January 2025; Salesforce Agentforce 3.0 (June 2025) supports cross-platform tool use via the Model Context Protocol.
5. Reasoning models inflected the cost-quality curve — OpenAI o1 (Sep 2024) and DeepSeek-R1 (Jan 2025) introduced inference-time reasoning, expanding what agents can verifiably accomplish on multi-step tasks.

1.6 Where the book lands

The book takes positions on three contested empirical questions, each defended in the chapters that follow:

1. Foundation models are commoditising; AI factories are not. The post-DeepSeek strategic landscape moves the locus of advantage up the stack from algorithms to operational architecture. (Chapters 3 and 5.)
2. The labour effects of generative AI compress the skill distribution. Novices benefit substantially more than experts; the augmentation effect is real and large in customer service, software engineering, and professional writing — but the “jagged frontier” matters. (Chapter 15.)
3. The right minimum unit of AI investment is all six Rewired capabilities at once, applied to 2–5 domains, sustained over 3–7 years. Capability investment in concert outperforms capability investment in isolation by a wide margin. There is no shorter path that works. (Chapters 4 and 16.)

1.7 Outline of the book

Chapter 2 traces the five-era history. Chapter 3 introduces the AI factory — the architectural pattern at the heart of Iansiti and Lakhani's framework. Chapter 4 develops Rewired's six-capability operating model. Chapter 5 takes up strategy, competitive collisions, and the “new meta”. Chapters 6–12 survey thirteen sectors. Chapter 13 covers the agentic frontier; 14, governance; 15, labour and productivity; 16, AI maturity and the suitability-for-machine-learning rubric; 17 consolidates the conceptual frameworks; and 18 presents teaching cases.

A short history of business AI from Feigenbaum's expert systems through the agentic enterprise, organised into five eras whose boundaries are drawn by the technology that defined the dominant deployments of the period.

2.1 A timeline at a glance

2.2 Era I — The pre-ML era (1950s–1990s)

The pre-ML era established the architectural separation of knowledge from inference that still organises regulated AI today. Edward Feigenbaum's Stanford laboratory built DENDRAL (1965) for mass-spectrometry interpretation; Edward Shortliffe's MYCIN (early 1970s, ~600 rules with certainty factors) performed at expert-physician level on bacterial diagnosis but was never deployed clinically owing to liability concerns. This pattern — technical adequacy blocked by institutional risk — would replay forty years later with IBM Watson Health.

The era's commercial flagship was John McDermott's XCON/R1 at Digital Equipment Corporation: built in OPS5, operational from 1980, growing to roughly 2,500 rules processing 80,000 orders annually with estimated savings of $25M/year by 1986. Maintenance, however, eventually demanded eight knowledge engineers — the original case study in the operational cost of brittle AI.

Other expert systems worth knowing: INTERNIST-1 (Pittsburgh, 1974) for internal medicine diagnosis; PROSPECTOR (SRI, 1979) for mineral exploration, famously credited with identifying a porphyry molybdenum deposit at Mount Tolman in Washington; and Symbolics LISP machines (1980–1990s), whose collapse marked the second AI winter.

Adjacent statistical work proceeded quietly. Peter Keen and Michael Scott Morton at MIT Sloan articulated decision support systems in the 1970s; Fair, Isaac & Co. introduced the general-purpose FICO score in 1989, perhaps the single most successful pre-ML statistical risk model ever deployed.

The era ended in two AI winters (~1974–1980 and ~1987–1993), driven by brittleness, escalating maintenance cost, and the collapse of the Symbolics LISP-machine market.

2.3 Era II — The statistical machine learning era (1990s–early 2010s)

The shift from rules to statistics gave business analysts general-purpose classifiers. Vapnik and Cortes's support vector machines (Machine Learning, 1995) and Breiman's random forests (2001) became workhorses; Hochreiter and Schmidhuber's LSTM (Neural Computation, 1997) opened sequential modelling.

IBM Deep Blue defeated Garry Kasparov 3½–2½ on 11 May 1997 — a brute-force symbolic system, but the public proof that “computers beat humans” at a benchmark task.

Three commercial milestones from this era still drive enterprise value today:

The era also saw Google PageRank (Brin and Page, 1998) — a statistical eigenvector computation on the web's hyperlink graph — and Google Search ad ranking (Quality Score, 2002), an early commercial deployment of a learned ranking model that would become a $200B+ business.

2.4 Era III — The deep learning revolution (2012–2022)

The era has a precise inflection point: Krizhevsky, Sutskever, and Hinton's AlexNet won ImageNet on 30 September 2012, with top-5 error of 15.3% versus 26.2% for the runner-up. Their company DNNresearch was acquired by Google in March 2013; the trio later received the 2018 Turing Award, and Hinton shared the 2024 Nobel Prize in Physics with John Hopfield.

The era produced the methodological substrate for everything that followed: Word2Vec (Mikolov et al., 2013), GANs (Goodfellow et al., 2014), seq2seq learning (Sutskever, Vinyals, Le, 2014), and ResNet (He et al., 2016) — whose 152-layer skip-connection architecture is still the backbone of most modern computer-vision systems. Most spectacularly, DeepMind's AlphaGo defeated Lee Sedol 4–1 in Seoul on 9–15 March 2016, watched live by over 200 million viewers; AlphaGo Zero (October 2017) then surpassed it with self-play alone in 40 days, learning Go from scratch without human game data.

Commercial deployments followed: Tesla Autopilot (October 2015), Spotify Discover Weekly (July 2015), Google Smart Reply (May 2015), Stitch Fix's hybrid human-plus-ML personalisation. IBM Watson's Jeopardy! victory over Ken Jennings and Brad Rutter (14–16 February 2011, $77,147 to $24,000 and $21,600) brought NLP into public consciousness — and led directly to the Watson Health misadventure that became the era's most expensive cautionary tale.

The era's quiet commercial workhorse was Google Smart Reply and BERT-powered Search. By late 2019, BERT was used in roughly 10% of US English Search queries; by mid-2020, the figure was “virtually all”. Google has never publicly disclosed the revenue impact, but it is reasonably estimated at tens of billions of dollars annually.

2.5 Era IV — Transformers and large language models (2017–2024)

This era began with one paper: Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin, “Attention Is All You Need,” NeurIPS 2017, now cited over 173,000 times. The architecture dispensed with recurrence and convolution in favour of multi-head self-attention, parallelising training and enabling the scale-by-compute regime that followed.

2.5.1 The papers that defined the era

2.5.2 The ChatGPT inflection

Enterprise infrastructure crystallised rapidly: Microsoft 365 Copilot reached general availability on 1 November 2023 at $30/user/month; ChatGPT Enterprise launched 28 August 2023; and JPMorgan's COiN system, mature by then, was canonically credited with eliminating 360,000 lawyer-hours per year.

2.5.3 The 2024 multimodal turn

The 2024 frontier models were materially different from their 2022 predecessors in being natively multimodal. GPT-4o (May 2024) handled vision, audio, and text in a single model with sub-second voice latency. Gemini 1.5 Pro (February 2024) introduced the long-context regime — 1 million tokens, later 2 million — that made full-codebase and full-document reasoning practical. Claude 3.5 Sonnet (June 2024) introduced the “Artifacts” pattern of structured generative outputs, then in October 2024 the “Computer Use” capability that opened the agentic era.

2.6 Era V — The agentic AI era (2023–2026)

The defining feature of the agentic era is the shift from per-step human-triggered copilots to goal-driven systems that plan, decompose tasks, call APIs, and act with reduced human supervision.

The architectural template is Yao et al.'s ReAct (ICLR 2023), which interleaves reasoning traces with tool-use actions. The public spark was Toran Bruce Richards's AutoGPT (30 March 2023), which hit 100,000 GitHub stars within months. The protocol layer arrived with Anthropic's Model Context Protocol (MCP), open-sourced in November 2024, which has become the de-facto integration standard for agentic systems.

2.6.1 Frontier launches

2.6.2 The DeepSeek shock

The era's defining geopolitical moment came at the start of 2025. DeepSeek-V3 (26 December 2024) achieved frontier performance for a reported $5–6 million in training compute. DeepSeek-R1 (20 January 2025), MIT-licensed, matched OpenAI's o1 on reasoning benchmarks.

The DeepSeek shock did not signal the end of frontier model investment. It signalled, instead, that frontier capability can no longer be taken as a durable moat — a commoditisation finding that reshapes every chapter of this book on competitive strategy.

2.7 What the eras have in common

A useful exercise after surveying the five eras is to ask what is constant. Three things, at least:

1. Capability runs ahead of organisational complement. XCON's brittleness, Watson Health's deployment failure, the 78%-adoption-versus-5%-value pattern, and Klarna's reversal are all instances of the same gap.
2. The most economically valuable deployments are quiet. FICO Falcon, Amazon's recommender, Google's Quality Score, BERT in Search, and DBS's GANDALF transformation produced more measurable value than any of the era-defining headline systems.
3. The architecture template is more durable than the technology. The four-component AI factory pattern (Chapter 3) describes Amazon's pre-2010 deployments as well as Anthropic's 2026 ones.

Iansiti and Lakhani argue that the firm of the AI age is organised around an “AI factory” — a scalable decision engine that transforms data into predictions, pattern recognition, and process automation. This chapter develops the framework, its four components, the virtuous cycle, and its three signature deployments.

3.1 The runtime metaphor

An AI factory is the operational pattern that lets a firm replace human-mediated decision processes with software-mediated ones at scale. It is the answer to the question Iansiti and Lakhani pose at the start of Competing in the Age of AI: how does Ant Group serve 700 million customers with 10,000 employees, while traditional banks of equivalent size employ 200,000?

The answer is not that Ant Group's bankers are 20× more productive. It is that Ant Group's operational critical path is run by software, with humans designing, supervising, and improving the system rather than executing inside it. That structural shift is what the AI factory describes.

3.2 The four components

Iansiti and Lakhani identify four components common to every operating AI factory. Each is necessary; collectively they are the prerequisite for digital scale, scope, and learning.

3.2.1 Data pipeline — the fuel

The data pipeline is where most AI projects succeed or fail. It is also where most of the cost lives. The Iansiti-Lakhani treatment is precise: a data pipeline is more than data engineering. It is a managed flow with provenance tracking, schema enforcement, deduplication, integration across systems, labelling (manual, weakly supervised, or self-supervised), and feature engineering. In modern terminology, it culminates in data products — versioned, governed, owned datasets that downstream teams consume through stable interfaces.

The 70–80% effort fraction is widely cited. McKinsey's 2023 work on AI deployment cost shows the same pattern: data wrangling and feature engineering dwarf model training in time, money, and complexity. Where firms underinvest in this layer, model performance is unstable and reproducibility is impossible.

3.2.2 Algorithm development — the machines

Algorithm development is the part of the factory most visible to outsiders. Iansiti and Lakhani's treatment emphasises that this is the part that has commoditised fastest in the post-2022 period. AutoML platforms, pre-trained foundation models, and open-source modelling frameworks (PyTorch, JAX, TensorFlow) have collapsed the distance between “state-of-the-art research” and “production model.” A team of two engineers using Hugging Face fine-tuning today can deploy what would have required ten ML PhDs five years ago.

The implication is that algorithmic capability is rarely the moat. The moat is the surrounding factory: the data pipeline that produces the labels, the experimentation platform that picks the winning configuration, and the infrastructure that runs the model reliably in production at scale.

3.2.3 Experimentation platform — the valves

The experimentation platform is where the AI factory differs most starkly from classical analytics. A platform like Amazon's Weblab runs over 30,000 simultaneous experiments at any time. Netflix runs hundreds of concurrent A/B tests across UI, content, and recommendation logic. Booking.com reportedly runs more than 1,000 simultaneous experiments. These are not occasional A/B tests; they are continuous discovery infrastructures.

The discipline is what distinguishes a firm that can learn from data from a firm that can merely process data. Without rigorous experimentation, predictions become opinions, and the virtuous cycle in §3.3 cannot turn.

3.2.4 Software infrastructure — the pipes

The fourth component is the operational substrate. MLOps — the discipline of deploying, monitoring, and updating ML models in production — has matured rapidly since 2020. Standard components: model registries (MLflow, SageMaker Model Registry), feature stores (Feast, Tecton, Databricks Feature Store), serving platforms (Kubernetes-based ML serving, NVIDIA Triton), monitoring (Arize, Fiddler, Evidently), and orchestration (Airflow, Dagster, Prefect).

The 2024–2026 evolution adds agent infrastructure — orchestration layers, tool registries, and observability surfaces specifically for agentic systems. These are covered in detail in Chapter 13.

3.2.5 The factory's outputs

The factory's output is consistently three things: predictions (what will happen, what does this contain, what should we recommend), pattern recognition (anomalies, segments, novel events), and process automation (acting on those predictions without a human in the loop). These map onto Davenport and Ronanki's three buckets discussed in Chapter 17.

3.3 The virtuous cycle

Once an AI factory is operating, it produces a self-reinforcing loop that is the central source of competitive advantage in the age of AI:

Once a firm has crossed the threshold where this loop is self-reinforcing, it has constructed a barrier to entry that is fundamentally different from the barriers studied in the pre-digital strategy literature (brand, scale, switching costs, network effects). It is closer to a data-and-learning moat: latecomers face a permanently improving competitor with no obvious way to catch up except by recombining digital capabilities from elsewhere.

3.4 Three reference deployments

3.4.1 Netflix

Netflix's recommendation system is the longest-running commercial AI factory. Its current architecture combines collaborative filtering, content-based filtering, deep neural networks, and continuous A/B testing across virtually every UI element. Approximately 80% of viewing originates from algorithmic recommendation rather than search (Gomez-Uribe and Hunt, ACM TMIS, 2015).

The factory's experimentation platform is the most distinctive component: Netflix runs hundreds of simultaneous A/B tests at any time and famously tests not only thumbnails but the headline image personalised per user. The data pipeline aggregates viewing traces, pause-and-rewind patterns, dwell time on artwork, and cross-device session behaviour. The infrastructure runs on AWS at a scale that makes Netflix one of AWS's largest customers.

3.4.2 Ant Group — the canonical case

Ant Group's 3-1-0 lending model — three minutes to apply, one second to approve, zero human intervention — exemplifies the factory pattern in financial services. The factory ingests transaction data from Alipay's payment graph, social and merchant data from Alibaba's commerce platform, and behavioural signals from the app itself; algorithms produce a continuous credit score for ~700M users; the experimentation platform tests pricing, marketing, and product variants; and the infrastructure runs at hyperscale.

Iansiti and Lakhani's headline statistic — 10,000 employees serving 700M users — is a property of the factory, not of the bankers. The book's first detailed case is precisely this contrast: ICBC, the world's largest traditional bank by assets, employs roughly 425,000 people to serve a comparable customer base. The 40× employee differential is the AI factory's signature.

Ant Group also illustrates the regulatory failure mode of frictionless impact (Iansiti-Lakhani Rule 4 in Chapter 5). The Chinese regulator's 2020 intervention — first the suspended IPO in November 2020, then forced restructuring in 2021–2023 — was the single largest regulatory action against a digital firm in modern financial history. The lesson is sobering: a firm that has built an AI factory has built operational power that may not be politically sustainable.

3.4.3 Amazon

Amazon's deployment is broader but architecturally similar. Its Weblab experimentation platform runs more than 30,000 simultaneous experiments at any given time. The same data pipeline that powers product recommendation also powers ad bidding, dynamic pricing, demand forecasting, supply-chain optimisation, and capacity planning across AWS. The factory's revenue impact is most visible in the third-party seller ecosystem: AI-driven listing tools, sponsored ad placement, and fulfilment-by-Amazon optimisation generate the bulk of seller-side margins.

The Iansiti-Lakhani argument is that Amazon's expansion across categories — books → general merchandise → groceries → cloud → advertising → media → pharmacy → healthcare — is enabled precisely because the AI factory transfers across categories. The capabilities are horizontal: the same pricing, demand forecasting, and recommendation infrastructure that runs Amazon Retail also runs AWS, Prime Video, and Alexa.

3.5 The factory at small scale: the LISH counterexample

One implication of Iansiti and Lakhani's argument is that any firm that builds an AI factory can capture digital scale, even without internet-era origins. The Laboratory for Innovation Science at Harvard (LISH) documented dozens of mid-market firms that built useful AI factories at low cost — typically combining cloud-hosted MLOps, open-source modelling tooling, and small in-house data teams. The lesson is that the architectural pattern matters more than the scale of the inputs; a 200-person logistics firm can and should run the same four-component factory that Ant Group does.

The 2024–2026 evolution makes this even more accessible. Hosted foundation models (Anthropic, OpenAI, Google), serverless inference (AWS Bedrock, Azure AI), and open-source modelling stacks (Hugging Face, LangChain, LlamaIndex) mean the factory's algorithm and infrastructure components can be assembled in weeks rather than years. The bottleneck is now the data pipeline, the experimentation discipline, and the organisational complement.

3.6 Why the factory is the moat, not the model

A crucial implication of the framework — and one borne out by the 2024–2026 commoditisation evidence — is that the AI factory's competitive advantage is not the algorithm itself. Foundation models commoditise; the experimentation platform that lets a firm pick the best model for each task does not. Data pipelines commoditise; the proprietary feedback loop in which user behaviour generates labelled data does not. Cloud compute commoditises; the operational discipline to run thousands of experiments concurrently and learn from them does not.

This insight reframes the post-DeepSeek strategic landscape. The right reading of January 2025 is not that Nvidia or OpenAI are doomed; it is that the locus of advantage moves up the stack from algorithms to operational architecture — exactly where Iansiti and Lakhani placed it five years earlier.

3.7 Operating model implications

Iansiti and Lakhani's later chapters draw out the operating-model implications of the factory. Three are worth memorising:

1. The firm shrinks in headcount per unit of output but grows in capability per employee. Ant Group's 10,000 staff are a different mix from ICBC's 425,000: more engineers, more data scientists, fewer transaction-processing clerks.
2. The boundary of the firm moves. Activities previously performed inside the firm (kept as competitive advantage) are now exposed via APIs to ecosystem partners, and activities previously contracted out are pulled inside (the “30–70 in-source shift” covered in Chapter 4).
3. The CEO's role changes. In an AI-native firm, the CEO's central job is the design and tuning of the factory itself — the architecture, the experimentation discipline, the talent mix — rather than the management of human-mediated operations.

3.8 The risks the framework does not foreground

Iansiti and Lakhani are clear-eyed about the risks but the framework can read as triumphalist. Three counter-points worth holding alongside the framework as you proceed through the rest of the book:

• Regulatory and political risk. Ant Group's restructuring is the canonical reminder. Frictionless impact at scale invites political reaction.
• Concentration and inequality. The framework predicts concentration — and Iansiti and Lakhani themselves acknowledge this in their Rule 5 (Chapter 5). Whether concentration is socially optimal is a question the framework does not answer.
• Quality and reliability ceilings. Klarna's reversal (Chapter 8) shows that operational scale alone does not guarantee quality outcomes at the customer interface.

If the AI factory is the architectural what, the McKinsey six-capability framework — drawn from Rewired (Lamarre, Smaje, and Zemmel, 2023) — is the operational how. This chapter develops each capability in turn, anchors the framework with three transformation cases, and presents the financial evidence for capability investment in concert.

4.1 The Rewired thesis

The McKinsey banking-benchmark study cited in Rewired compared 20 global banks identified as digital and AI leaders against 20 laggards over five years (2017–2022). The leaders outperformed laggards on TSR by approximately 14 percentage points per year. The book argues — and the subsequent State of AI surveys confirm — that the gap is durable, widening, and explained primarily by capability investment rather than by initial conditions, geography, or regulatory environment.

4.2 The six capabilities

4.3 Capability 1 — Roadmap

The roadmap is the strategic frame. Rewired's most distinctive prescription is to organise transformation by domains — coherent, end-to-end customer or operational journeys typically 10–15 in number for a large enterprise — and then to attack 2–5 of them seriously. The book reports that roughly 80% of successful transformations re-anchor on domains after starting elsewhere (capability-led, technology-led, or use-case-led).

Why domains rather than use cases? Three reasons:

1. Domain redesign forces alignment of the other five capabilities. A talent investment without a domain target is unmoored; a technology investment without a domain target is shelfware; a data product without a domain target is a dataset.
2. Domains are the unit at which workflow redesign matters. The Iansiti-Lakhani thesis at the operational level: AI capability is necessary but value capture requires workflow redesign, and workflow lives at the domain level.
3. Domains map to executive accountability. A use case has no natural owner; a domain does. The executive sponsor can be held to the financial and operational outcomes of a domain in a way that is impossible for a portfolio of disconnected use cases.

Examples of well-defined domains in different industries: in retail banking, “mortgage origination”, “onboarding and KYC”, “personal-loan customer journey”; in pharmaceutical R&D, “target identification through Phase II”, “clinical trial site selection and patient recruitment”; in retail, “post-purchase customer service”, “merchandise planning and allocation”, “fulfilment optimisation”.

4.4 Capability 2 — Talent

The talent capability has two distinctive features in the McKinsey treatment.

4.4.1 The 30–70 shift

Leading firms typically bring 30–70% of their digital and AI talent in-house, reversing the 20-year drift toward systems-integration outsourcing. DBS Bank moved from approximately 15% in-house tech in 2009 to roughly 90% by 2018 — the most-cited example. The thesis: critical capability cannot be effectively outsourced because it must be embedded in the business's everyday decisions.

The McKinsey research identifies six roles for which in-sourcing is essentially mandatory: data engineer, data scientist, ML engineer, product owner, scrum master / agile coach, and UX/UI designer. Other roles (DevOps, security, infrastructure, project management) can often remain partially outsourced without serious capability loss.

4.4.2 2-in-a-box platform leadership

Each technology platform is co-led by a business owner (responsible for outcomes) and a technology owner (responsible for the build), reporting jointly to a single executive sponsor. This breaks the long-standing IT-versus-business gap that has defeated many digital transformations. The model is most fully realised at DBS, where 33 platforms operate under 2-in-a-box leadership — and at Spotify and ING, whose “tribes and squads” pioneered the form.

4.4.3 The talent supply problem

The talent capability is constrained by global supply. The Stanford AI Index documents that ML PhD output remains heavily concentrated in a handful of US universities (Stanford, Carnegie Mellon, MIT, Berkeley, Princeton); the largest absolute talent pool is in the United States, with China second and India a fast-rising third. For Asia-Pacific firms outside these three countries, the talent strategy must combine in-sourcing of the most critical roles with capability building (typically through targeted graduate-school sponsorship and partnerships with regional universities).

4.5 Capability 3 — Operating model

The operating-model prescription is convergent across Rewired, the agile literature, and the AI factory framework: cross-functional product pods of 8–10 people, owning a domain end-to-end, releasing software continuously, and measured on outcome KPIs (revenue, cost, satisfaction) rather than activity KPIs (lines of code, tickets closed). The pod is the modern unit of digital execution.

The operating model also implies governance: the platform owners coordinate across pods, the executive sponsors coordinate across platforms, and the CEO coordinates across the portfolio. Rewired's most-cited diagnostic question for boards is: How many of your pods can ship to production this week without C-suite sign-off? The answer in most large firms is “essentially none”; in successful transformations, the answer is “most of them.”

4.6 Capability 4 — Technology

The technology capability is summarised in three architectural moves: decouple (microservices, APIs, well-defined contracts), cloud-native (elastic infrastructure, managed services), self-service (developer platforms, automated provisioning, MLOps for production ML).

The DBS figure of 99% workload migration to cloud is a useful directional benchmark — most large firms today are at 30–60%. The 2024–2026 evolution is hybrid: the pure public-cloud thesis has softened as some workloads (training large foundation models, serving high-throughput inference, regulated workloads) move back to on-premise or sovereign cloud arrangements. The right framing is “cloud-native, multi-cloud, and partially sovereign” rather than “all on AWS”.

A specific technology choice worth memorising is the data-mesh-versus-data-lake-versus-data-warehouse debate. McKinsey's reading by 2025 is that the data mesh has won the architectural argument — federated ownership of data products with consistent governance — but that the operational reality is hybrid: data lakes underneath for cheap storage, data warehouses for SQL-driven analytics, and data products as the consumer-facing interface.

4.7 Capability 5 — Data

The fifth capability is the one that has shifted most in the post-2020 period. The Rewired prescription is the data product: a managed, versioned, governed dataset with a clear product owner, well-defined consumers, SLAs, and quality metrics. A typical large firm has 50–500 candidate data products; the goal is to operationalise the 20–50 highest-value ones.

McKinsey's reported impact for firms that adopt this approach: 90% faster delivery of AI use cases, 30% lower TCO of analytics, with reusability lifting downstream value. The benefits compound: each new use case gets cheaper and faster as the data-product layer matures.

The 2024–2026 development is the integration of retrieval-augmented generation (RAG) infrastructure into the data layer. RAG patterns turn unstructured corporate content (documents, emails, presentations, code) into queryable knowledge surfaces that LLMs can retrieve from. The data engineering required to make RAG work — chunking, embedding, indexing, permission preservation, freshness management — is often misunderstood as “just put it in a vector database”. In practice, it is a full data-product engineering exercise.

4.8 Capability 6 — Adoption

The most-overlooked capability and, by McKinsey's reckoning, the single most common failure mode. Adoption requires change management, measured usage, embedded workflow integration, and ongoing reinforcement through KPIs.

The Rewired framework is unusually emphatic that roughly half of every transformation budget should be spent on adoption — a number that strikes most firms as wildly high until they have failed to capture value from a technically successful deployment. The specific change-management activities: training (role-specific, not generic), incentive realignment (KPIs and performance reviews), workflow redesign (often the hardest), executive sponsorship and communication (consistent, not episodic), and ongoing measurement of usage and impact.

Microsoft's own deployment of Microsoft 365 Copilot internally produced widely-shared learnings that adoption is bimodal: a small cohort of power users adopt aggressively and unlock substantial productivity gains; a long tail of users barely engage. The variance can be reduced by training and incentive design but cannot be eliminated.

4.9 Three anchor cases

4.9.1 Freeport-McMoRan and the Bagdad concentrator

Freeport-McMoRan applied ML on existing sensor data — without new capital expenditure — at its Bagdad concentrator in Arizona, achieving approximately 5% throughput improvement. The case is canonical because it illustrates that the AI factory's data pipeline often runs on the data the firm already collects; the ROI is unlocked by the experimentation discipline, not by buying new sensors.

The follow-on point: the Bagdad pilot funded the broader transformation. The financial value released by the pilot was used to invest in the data-product, MLOps, and pod infrastructure that subsequently scaled across Freeport's North American operations. This is the classic “land and expand” pattern in industrial AI transformation.

4.9.2 DBS Bank's GANDALF transformation

4.9.3 LEGO

LEGO's digital transformation, profiled extensively in the Rewired companion case collection, is a useful counter-anchor to DBS because it operates in a non-financial-services context. LEGO rewired across e-commerce, omnichannel retail, the LEGO Ideas community platform, and the Mindstorms / digital play extensions.

The key empirical point is that LEGO managed the rewiring while maintaining the brand's physical-product identity — which is exactly the challenge incumbents in retail, hospitality, and manufacturing face. The 2020–2025 financial outcomes (revenue compounding above 10% per year through pandemic and post-pandemic; market share gains against competitors; successful expansion into China) are partly attributable to the digital infrastructure built in the preceding decade.

4.10 Why all six capabilities, in concert

The strongest finding in the McKinsey benchmark data is that capability investment in concert outperforms capability investment in isolation by a wide margin. A firm that invests in talent and technology but neglects roadmap, data, and adoption typically fails to capture firm-level financial impact — even when individual deployments are technically successful.

The implication for transformation programmes is sobering: the right minimum unit of investment is all six capabilities at once, applied to 2–5 domains, sustained over 3–7 years. There is no shorter path that works.

4.11 Common transformation failure modes

The chapter ends with the reverse — the failure modes Rewired documents and the McKinsey teams have observed across hundreds of engagements:

• The use-case zoo. Hundreds of pilots, none integrated, none scaled. Symptom: a CEO who can list 50 pilots but cannot point to a single domain that has been transformed.
• The platform replatform. A multi-year technology rebuild with no business-facing outcomes until the end. Symptom: $200M spent on a new core banking system before any new customer-facing functionality ships.
• The hire-and-hope. A Chief Data Officer or Chief AI Officer is appointed; the rest of the operating model is unchanged. Symptom: the CDO/CAIO leaves within 18 months citing lack of authority.
• The pilot-to-production gap. Models perform well in pilot but cannot be operationalised. Symptom: 80% of models built, 10% of models in production.
• The adoption shortfall. Production deployments exist; nobody uses them. Symptom: licence purchase data shows 1,000 seats; usage data shows 100 active users.

Iansiti and Lakhani argue that AI is not just changing how firms operate; it is changing the rules of competition itself. This chapter develops their five rules of the “new meta” alongside Agrawal-Gans-Goldfarb's three solution layers, the post-DeepSeek commoditisation debate, and the AI-native disruption pattern.

5.1 Strategic collisions

The central strategic phenomenon of the AI era is the collision — when an AI-enabled digital operating model meets a traditional one. Iansiti and Lakhani's case studies are organised around these collisions: Amazon vs. Walmart in retail, Ant Group vs. ICBC in banking, Netflix vs. Blockbuster in media, Tesla vs. legacy automakers in cars, Peloton vs. SoulCycle in fitness, Shopify vs. eBay in marketplace commerce.

The pattern is consistent. The traditional firm faces diminishing returns to scale — its operating complexity grows faster than its revenue. The digital firm faces increasing returns to scale, scope, and learning. When they collide, the digital firm's value curve overtakes the traditional firm's, often catastrophically and quickly.

The underlying mathematics is simple. Traditional firms face a U-shaped average cost curve — efficient up to some scale, then increasingly costly to coordinate at greater scale. Digital firms face a downward-sloping average cost curve over the relevant range — every additional user is essentially free to serve, and the data that user generates makes the rest of the system more accurate. The collision is between two fundamentally different production functions.

5.2 Rule 1 — Change is no longer localised; it is systemic

Inventions during the Industrial Revolution pertained to specific industries — the steam engine had more impact in manufacturing and transportation than in banking or healthcare. Digital technology and AI are different: they cut across every industrial environment at the same time. The same transformer architecture that recommends songs also drafts contracts, interprets X-rays, prices freight, and writes code.

The empirical pattern is that no industry has been observably immune. Even the slowest-moving sectors — primary agriculture, defence procurement, ecclesiastical administration — have measurable AI deployment by 2026. The diffusion path is uneven; the diffusion direction is universal.

5.3 Rule 2 — Capabilities are increasingly horizontal and universal

In a dramatic reversal from the Industrial Revolution's vertical specialisation, the age of AI is making vertical, siloed organisations and specialised capabilities less relevant. Competitive advantage is shifting from vertical capabilities (decades of insurance underwriting expertise, deep manufacturing process knowledge) toward universal capabilities in data sourcing, processing, analytics, and algorithm development — the AI factory components from Chapter 3.

The signal: when Uber looked for a new CEO, the board hired someone who had previously run a digital firm (Expedia), not a transportation services company. The same applies to firms entering new sectors: Amazon (retail) entered cloud, advertising, content, and pharmacy; Tencent (gaming and messaging) entered financial services and healthcare; Alibaba (commerce) entered banking and logistics. Each move worked because the AI factory transferred — not the vertical knowledge.

The implication for executive talent is profound. The most-valuable senior leaders are increasingly those who can manage the AI factory rather than those with deep vertical expertise. The latter is being recoded as a feature of the data pipeline rather than a feature of the executive's CV.

5.4 Rule 3 — Industry boundaries are disappearing; recombination is the rule

Industries originally evolved from traditional trades to support vertical specialisation. Those clear boundaries are dissolving. Google entered the auto industry. Alibaba launched a bank. Amazon entered pharmacy. Apple entered financial services. Tesla entered insurance. Costco entered prescription drug fulfilment. Shopify entered logistics. Stripe entered banking.

The structural reason: digital interfaces let operating models cut across old verticals and enter new industries with new, highly connected business models. While traditional organisations suffer diminishing returns to scale or scope, digital networks enjoy increasing returns — both as they grow in size and as they connect to other networks.

The 2024–2026 evolution is that the recombination is becoming bidirectional. NVIDIA, a chip company, has become a major software platform (CUDA, DGX Cloud, Omniverse) and an applied AI firm (autonomous vehicles, drug discovery). Anthropic, an AI lab, is becoming an enterprise software company. OpenAI has become a consumer products company. The boundaries between “chip company”, “model lab”, “enterprise software firm”, and “consumer app” are dissolving.

5.5 Rule 4 — From constrained operations to frictionless impact

Digital operating models remove traditional operating constraints. Ant Group serves an order of magnitude more customers than the largest traditional bank. Facebook reaches an order of magnitude more people than the US postal system. Information moves instantaneously at near-zero marginal cost via networks to infinite numbers of recipients.

But removing friction is not always good. Frictionless systems are prone to instability and have difficulty finding equilibrium. Once in motion, they are hard to stop:

The 2024–2026 evolution adds AI-mediated amplification on top of digital amplification. Generative AI lowers the cost of producing convincing variants of content; agentic AI lowers the cost of distribution; reasoning models lower the cost of crafting messages tuned to specific audiences. The net effect is a multiplicative reduction in the cost of mounting a sophisticated information operation, with downstream consequences for elections, financial markets, and public health.

5.6 Rule 5 — Concentration and inequality will likely get worse

As digital networks carry more transactions, network hubs gain power. Once a hub is highly connected in one sector (Airbnb in home rentals, Alibaba in peer-to-peer retail), it gains advantages as it links to a new sector (Airbnb in travel experiences, Alibaba in financial services). The pattern produces concentration of wealth, power, and relevance across markets, industries, and geographies.

The empirical evidence as of 2026 is that this rule is largely vindicated. Stanford's AI Index 2025 documents US private AI investment of $109.1 billion versus China's $9.3 billion; the seven US hyperscalers (Microsoft, Google, Amazon, Meta, Apple, Nvidia, Tesla) account for the majority of global AI infrastructure spending; and the top quintile of AI-using firms — McKinsey's “rewired” cohort — capture a disproportionate share of the gains (16–30% productivity improvements vs single-digit averages).

The geographic concentration is striking. Of approximately $250 billion in 2024 global private AI investment, more than 60% landed in firms headquartered in California. Of the top 50 AI-native firms by 2025 enterprise value, fewer than five are headquartered outside the United States and China. The implication for sovereignty and policy is taken up in Chapter 14.

5.7 The Agrawal-Gans-Goldfarb framework: three solution layers

If Iansiti and Lakhani describe what AI-enabled competition looks like, Agrawal, Gans, and Goldfarb (Power and Prediction, 2022) explain why most enterprise AI doesn't disrupt anything yet. Their framework distinguishes three layers of AI deployment:

The authors call the current era the “Between Times” — the period after AI's promise has been demonstrated but before its full potential is realised. They argue that we are in a proliferation of point and application solutions; the truly disruptive system solutions remain rare. The historical analogue is electricity: it took roughly forty years from Edison's 1882 Pearl Street station for factories to redesign around electrical power (replacing line-shaft architecture with distributed motors at each workstation).

5.7.1 The insurance system-solution example

Agrawal, Gans, and Goldfarb's clearest worked example is home insurance. Today, insurance is risk transfer: the homeowner pays premiums, the insurer pools and pays out claims. With three super-powerful AIs predicting (1) lifetime customer value × probability of converting, (2) likelihood of filing a claim × claim magnitude, and (3) legitimacy of any claim, the insurer can:

1. Allocate marketing optimally;
2. Price premiums precisely (or decline if expected loss exceeds price);
3. Settle claims in seconds;
4. And — most disruptively — offer risk prevention as a service: subsidise a leak-detection sensor whose installation cost is less than the expected reduction in claim cost.

The fourth move transforms insurance from a financial industry into a hybrid IoT-and-services industry. This is what a system solution looks like.

5.7.2 The healthcare system-solution analogue

The healthcare equivalent is the shift from episodic treatment to continuous monitoring and prevention. With AI predictions of cardiac risk from continuous wearables, an insurer or integrated provider can intervene earlier (medication, lifestyle modification, early treatment) at lower cost than treating an acute event. The deployment requires redesigning every link in the chain: data collection (consumer wearables → clinical data systems), permissioning (HIPAA, GDPR, local equivalents), reimbursement (preventive vs episodic billing codes), and care delivery (continuous vs visit-based). The point-solution version of the same idea — a smart device that warns the patient — captures little of the value.

5.7.3 The pharmaceutical-discovery system-solution analogue

AlphaFold + generative chemistry + automated wet-lab + clinical-trial AI is potentially a system solution that compresses drug discovery from a decade to a year. Each component exists; the system-level integration that captures the full value does not yet. Isomorphic Labs (Chapter 7) is the most credible attempt to build it.

5.8 The commoditisation debate

Iansiti and Lakhani argue that the AI factory is the moat. The 2024–2026 evidence pulls in two directions on this claim.

5.8.1 The case for commoditisation

The DeepSeek-R1 release in January 2025 demonstrated that frontier reasoning capability can be achieved at one-tenth to one-hundredth of the previously assumed cost — and that the resulting model can be released MIT-licensed. Llama 4 (April 2025), Qwen, and Mistral followed similar paths. Open-weight models now match closed frontier models on most benchmarks. The implications: foundation model capability is becoming a commodity input, not a moat.

The supporting evidence: foundation model API prices have fallen 90%+ since GPT-4's launch; the gap between open and closed frontier models has closed to roughly 6–12 months on most benchmarks; switching costs between model providers (in well-architected enterprise deployments) are days rather than years.

5.8.2 The case for AI-native disruption

The Menlo Ventures 2025 State of GenAI report and Foundation Capital's 2026 outlook document a striking pattern: AI-native startups are taking material market share from incumbents in agile departments where speed of iteration matters more than integration depth.

The Menlo Ventures finding: in finance and operations, startups now hold 91% of the AI-native software share; in market research, sales, marketing, and product, the figures are similar. Where AI-native disruption is weak: IT and data science, where reliability and deep integrations outweigh speed.

5.8.3 The synthesis

The two views can be reconciled. Foundation models are commoditising; AI factories are not. Cursor's moat is not the model — it's the editing surface, the repo-level context, the diff approval flow, the developer feedback loop. Glean's moat is the cross-system permission graph, the entity resolution, the personalisation of search to organisational context. These are AI factory advantages, not model advantages. The Iansiti-Lakhani thesis survives — but its locus moves from algorithms to the operational architecture surrounding them.

An additional implication is that AI-native firms can move faster than rewired incumbents in the early years of a market — but the rewired incumbents (DBS, JPMorgan, Walmart) tend to win the long run because they have access to proprietary data, distribution, and customer relationships that the AI-natives must build from scratch. The right framing is “ten-year race” rather than “who's ahead in 2026.”

5.9 The five ethical categories

Iansiti and Lakhani's Chapter 8 — “The Ethics of Digital Scale, Scope, and Learning” — groups the new ethical challenges of AI-enabled firms into five categories, each of which receives extended treatment in Chapter 14 of this textbook:

1. Digital amplification — viral falsehood, polarisation, addictive engagement loops.
2. Bias — algorithmic discrimination at scale, the Buolamwini-Gebru pattern.
3. Cybersecurity — attack surface as networks expand.
4. Control — what is the appropriate locus of decision authority when algorithms drive the operational critical path?
5. Inequality — winner-take-most concentration, geographic disparity, labour displacement.

Microsoft's six AI principles — fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability — are the firm's response to these categories and have become the de-facto template for enterprise AI ethics policies.

5.10 Strategic implications for the 2026 firm

If you take Iansiti-Lakhani, Agrawal-Gans-Goldfarb, and the AI-native disruption evidence seriously together, the strategic playbook for the established firm in 2026 is:

1. Build the AI factory. Architecture beats algorithms.
2. Invest in system solutions, not just point solutions. The 80% of gains comes from the 20% of deployments that redesign workflows.
3. Watch for collisions, not for entrants. The threat is rarely a new bank; it is a tech firm that decides banking is one more vertical where its AI factory transfers.
4. Treat foundation models as commodities. Multi-vendor, model-agnostic, ready to swap when DeepSeek-R2 lands.
5. Take ethics seriously, early. Microsoft's six principles, ISO 42001, the EU AI Act — these are not bureaucracy; they are the operating constraints of digital scale, scope, and learning.
6. Plan for the long run. AI-native disruption is real but partial; the rewired-incumbent path requires a 5–10 year horizon and the full Rewired six-capability investment.

Finance remains AI's largest commercial value pool, with deployments spanning algorithmic trading, credit scoring, fraud detection, robo-advisory, and increasingly central-bank supervisory technology. The sector is also where the “classical” statistical-ML wave still drives most of the value, even as generative AI is layered on top.

6.1 Algorithmic and quantitative trading

Quantitative strategies now drive over 60% of US equity daily volume. The canonical case remains Renaissance Technologies' Medallion Fund, which returned roughly 30% in 2024 on $106B AUM. Two Sigma, D. E. Shaw, and DRW operate at similar quantitative scale, increasingly incorporating deep learning for alternative-data signal extraction (satellite imagery, credit-card panel data, natural language sentiment).

The architectural pattern is a textbook AI factory: tick data and alternative data feed a pipeline; the pipeline trains classical (random forest, gradient boosting) and deep (LSTM, transformer) models; the experimentation platform is the trading book itself, with rigorous attribution of P&L to signals; the infrastructure is co-located, sub-millisecond, and on-premise.

The 2024–2026 evolution adds LLM-driven signal extraction from earnings calls, regulatory filings, and analyst reports. The pattern: open-source LLMs fine-tuned on financial corpora extract sentiment and topic signals from textual sources at scale that was impractical with classical NLP. Several quant firms have publicly disclosed material P&L attribution to LLM-derived signals in 2025.

6.2 Generative AI in major banks

JPMorgan's COiN system — deployed in 2017 and now mature — eliminates 360,000 lawyer-hours/year by reading 12,000 commercial credit agreements in seconds. President Daniel Pinto estimated AI's annual value at the bank at $1–1.5 billion against an $18 billion 2025 tech budget — roughly an 8% productivity dividend on the technology budget alone, ignoring revenue uplift.

The deployment template emerging across major banks: a permissioned in-house LLM platform (often built on Azure OpenAI or AWS Bedrock with private inference), fronted by role-specific applications (research drafting, code generation, customer-support assist, compliance review), with rigorous data-leakage controls and content moderation.

6.3 AI lending and credit scoring

The watershed event in mainstream credit scoring came on 22 April 2026, when FICO Score 10T received GSE approval for use in conforming mortgage underwriting. Trended-data and machine-learning-derived scores now flow through the largest US mortgage market.

The Hemachandran & Rodriguez (2024) volume's finance chapter (Jafar, Alam, and El-Chaarani) emphasises three ML applications that scale particularly well in emerging-market banking: SME credit scoring using alternative data (mobile-phone usage, geolocation, utility-payment history) where bureau data is thin; cash-flow forecasting for working-capital lending; and portfolio-construction tools for retail wealth management at HNW thresholds previously uneconomic.

6.3.1 The fintech-credit landscape in Southeast Asia

The Southeast Asian fintech-credit landscape is particularly illustrative because alternative-data underwriting is a commercial necessity rather than an optimisation:

• Kredivo (Indonesia) — buy-now-pay-later and personal lending. Underwrites on mobile-phone metadata, e-commerce behaviour, and bill-payment history. Now serves over 8 million users.
• Akulaku (Indonesia, Philippines, Malaysia, Vietnam) — consumer credit at scale. Uses smartphone metadata, utility-payment data, and social-graph signals.
• Atome (Singapore, regional) — BNPL with regional spread.
• Aspire (Singapore) — SME banking. Cash-flow-based underwriting on integrated bank-feed and accounting data.
• Funding Societies (Singapore, regional) — SME peer-to-peer lending. ML-mediated risk segmentation across thin-file SMEs.

6.4 Fraud detection

Modern card-not-present fraud detection uses graph neural networks and transformer-based sequence models on transaction streams. Mastercard Decision Intelligence Pro reports up to 300% improvement in detecting at-risk merchants and 200% reduction in false positives. Visa Advanced Authorization is credited with preventing over $40 billion in fraud annually.

The 2024–2026 frontier is cross-institution fraud graphs — graph ML applied to transaction networks that span multiple banks and payment processors via privacy-preserving compute (federated learning, homomorphic encryption, secure enclaves). The BIS Project Aurora results (§6.6) are the canonical demonstration.

6.5 Robo-advisory

Robo-advisory crossed $1.2 trillion in AUM by year-end 2024 (Vanguard $365B, Empower $200B, Schwab $89.5B, Betterment $56.4B). The economics are not universally favourable: Goldman's Marcus Invest, JPMorgan's Automated Investing, and Ellevest exited the segment over scale economics. The teaching point: scale is not the same as competitive advantage. The robo-advisor as commoditised distribution channel has won; the robo-advisor as proprietary moat has not.

The 2024–2026 evolution adds LLM-driven advice surfaces. Morgan Stanley's AI @ Morgan Stanley Assistant is the most-deployed example: not a robo-advisor but a knowledge surface that lets human advisors handle 2–3× the prior client load. The unit economics work because the marginal cost of advisor time has fallen, not because the distribution has been automated.

6.6 Central banking and supervisory AI

The Bank for International Settlements' Project Aurora demonstrated that graph machine learning can detect up to 3× more money-laundering networks at 80% lower false positives than rules-based approaches. Project Aurora's findings have been institutionalised in the BIS Innovation Hub's roadmap and influence supervisory AI design at multiple G20 central banks.

Other notable supervisory AI deployments:

• Bank of England Prudential Regulation Authority — ML for stress-testing scenario generation.
• European Central Bank — NLP on supervisory dialogue and bank disclosures.
• Monetary Authority of Singapore — Project Mojito (suspicious-transaction monitoring), Veritas (fairness-and-explainability framework for FS AI).
• Bank Negara Malaysia — Sandbox-tested supervisory analytics on the Open API data layer.

6.7 Islamic finance and Shariah-compliant AI

Islamic finance AI (Zoya, Wahed, Musaffa, IFG.vc, HalalScreener) automates AAOIFI Standard No. 21 screening across roughly 40,000 instruments. The deployments remain primarily rule-based rather than LLM-grounded — the regulatory tolerance for hallucination in fatwa-adjacent domains is essentially zero — but generative AI is increasingly used for customer-facing explanation rather than core screening logic.

The structural challenge for Shariah-compliant AI is verification. A retrieval-augmented LLM can summarise an AAOIFI ruling, but the user (or the Shariah supervisory board) must be able to verify the citation chain back to authoritative sources. This is a generalisable pattern: in regulated domains, AI's value comes from compressing the time to a verifiable answer rather than from delivering a fully autonomous one.

6.8 The DBS case revisited

6.9 The persistent failure mode: trader-and-quant culture clash

One under-discussed pattern in financial-services AI is the cultural friction between the front-office trading culture (intuition-driven, individual P&L, short time horizon) and the AI factory's culture (data-driven, team P&L, longer time horizon). The most consistent failure mode in bank AI deployments is not technical; it is the front office refusing to adopt model output that contradicts trader intuition, or refusing to surface the data needed to train models in the first place. The successful deployments (Goldman's GS AI Platform, JPMorgan's LLM Suite, DBS's GANDALF) have invested heavily in cultural alignment alongside technical infrastructure.

Healthcare experienced its most consequential AI moment in May 2024 with the publication of AlphaFold 3. The Nobel followed in October 2024. The clinical floor — ambient AI scribes, FDA-cleared imaging devices, hospital virtual nurses — is now in the steepest part of its adoption curve.

7.1 Drug discovery and molecular AI

The field's centre of gravity is now structural prediction extended to molecular interactions. AlphaFold 3 (Abramson et al., Nature 630:493–500, May 2024) extends protein structure prediction to nucleic acids, ligands, and ions in a unified diffusion architecture. The 2024 Nobel Prize in Chemistry went to Demis Hassabis and John Jumper of DeepMind alongside David Baker of Washington — the first time a Nobel had been awarded to AI researchers for an applied scientific breakthrough rather than for foundational AI methods themselves.

Insilico Medicine's INS018_055 became the first drug with both AI-discovered target and AI-designed molecule to reach Phase II, with topline data in November 2024 showing forced vital capacity (FVC) improvement in idiopathic pulmonary fibrosis. As of May 2026, no AI-designed drug has yet received FDA approval — but the pipeline is now thick enough that the first approvals are widely expected within 24 months.

7.2 Ambient AI scribes: the steepest curve in clinical IT

Ambient documentation tools (Microsoft Dragon Copilot, Abridge, Suki, Ambience) have deployed across major US health systems faster than any clinical IT category in recent memory. Two anchoring data points:

The economics work because the clinical-documentation tax on physicians has been the single largest driver of burnout in US medicine, and the ROI is immediate (more time per patient or shorter days), measurable, and felt directly by the highest-paid clinical staff. The same technology pattern is now being deployed in Singapore, Hong Kong, and the UK, with the largest deployments yet to come in Asia-Pacific systems where physician shortage is acute.

7.3 FDA-cleared AI medical devices

The FDA's AI/ML-enabled medical device list reached 1,016 cumulative authorisations by December 2024, with 168 in 2024 alone. Approximately 96% are 510(k) clearances and 76% are radiology devices. Only 16.7% included Predetermined Change Control Plans — meaning most cleared devices remain “locked” rather than continuously learning.

The category leader in clinical workflow impact is Viz.ai's stroke triage platform, which saves an average of 52 minutes per case in door-to-treatment time for large-vessel occlusion — a clinically meaningful gain in a domain where neurons die at roughly 1.9 million per minute.

Other notable deployments: HeartFlow (FFR-CT for coronary disease), IDx-DR (autonomous diabetic-retinopathy screening, the first FDA-authorised autonomous AI diagnostic), Aidoc (multi-modality imaging triage), Paige Prostate (digital pathology AI for prostate cancer detection).

7.4 Hospital virtual care and nursing

Hippocratic AI's Polaris constellation completed 1.85 million patient calls by 2025, using a primary clinical model paired with 22+ specialist support models for safety validation. The architecture is the operational answer to hallucination risk: instead of relying on a single LLM's calibration, the system runs each output through specialist gatekeepers before delivery.

The unit economics: a Hippocratic-AI nurse is reported to cost roughly $9 per hour of operation (compute + supervision), against a U.S. RN cost of roughly $50 per hour. The deployment is supplementary rather than substitutive — patients with complex needs continue to receive human care — but the cost differential makes coverage affordable in regions and shifts where human staffing is unaffordable.

7.5 The Hemachandran & Rodriguez treatment

The healthcare chapter in the Hemachandran & Rodriguez (2024) volume emphasises three AI applications that scale particularly well outside high-income settings:

1. Image-based screening for diabetic retinopathy, tuberculosis, and breast cancer — particularly valuable where specialist density is low (Asia-Pacific, sub-Saharan Africa).
2. Clinical decision support in resource-constrained settings, particularly for triage and antibiotic stewardship.
3. Telehealth-plus-AI models pioneered during COVID-19 that have since become permanent infrastructure in India, Indonesia, Vietnam, and the Philippines.

The volume's editors note — and the empirical evidence supports — that healthcare AI's value capture has been disproportionately strongest in screening and operational workflow, and weakest in the diagnostic-decision-and-treatment domains where Watson Health and Babylon foundered.

7.6 The didactic counterexamples

The common thread across these failures is the gap between demo capability and clinical workflow integration — the same lesson MYCIN delivered in the 1970s, restated in fifteen-figure terms. The Olive collapse is particularly instructive: the firm pursued horizontal automation (RPA + ML) across hospital revenue cycle, which sounded plausible to investors but in practice required deep integration with hundreds of legacy hospital systems. The customer adoption was slow, the gross margin was thin, and the unit economics never closed.

Retail combines algorithmic personalisation at unprecedented scale with the era's most cited cautionary tale: Klarna's 2024 AI assistant launch and 2025 reversal.

8.1 The Iansiti-Lakhani anchor: Amazon's AI factory

Iansiti and Lakhani open Competing in the Age of AI with Amazon's digital operating model. The structure is the canonical retail AI factory:

• Data pipeline: browsing histories, purchase behaviour, return patterns, search queries, ad impressions, third-party seller metadata.
• Algorithms: collaborative filtering, content-based ranking, dynamic pricing, demand forecasting, supply-chain optimisation, ad bidding.
• Experimentation platform: Weblab — >30,000 simultaneous experiments at any given time.
• Infrastructure: AWS, which is itself sold to other firms as their AI factory infrastructure.

Amazon's data flywheel — more shoppers → more data → better recommendations → more conversion → more shoppers — illustrates Iansiti and Lakhani's argument that AI factory architecture produces increasing returns. Walmart, by contrast, has spent the last decade rewiring itself around a similar architecture (Walmart Labs, the Bonobos and Jet acquisitions, the JD.com partnership) — partial successes against a much higher run rate of capex.

8.2 Walmart's neural-network demand forecasting

Walmart's neural-network demand forecasting reportedly avoided 30 million unnecessary truck miles. Pactum's autonomous supplier negotiations close 68% of cases at roughly 3% cost savings. These deployments are characteristic of the era: classical-ML forecasting integrated with LLM-mediated negotiation surfaces.

8.3 Amazon's robotic warehouse

Amazon Robotics deployed over 750,000 mobile units by late 2023 (Sequoia, Proteus, Sparrow, Vulcan, Digit pilot), with recordable-incident rates 15% lower at robotics-equipped sites. The Sparrow system is the first AI-vision robot in Amazon's fulfilment centres capable of detecting, selecting, and handling individual products in inventory.

8.4 Visual search and merchandising

Pinterest Lens handles roughly 1.5 billion queries per month and reports +62% conversion versus text-based search. Amazon's seller-side generative AI tools were used by over 900,000 selling partners by mid-2025, with 80–90% accept rates and 40% improvements in listing-quality scores.

8.5 The Klarna case: the canonical reversal

8.6 The Hemachandran & Rodriguez retail chapter

The Hemachandran & Rodriguez volume's retail chapter (Kundu, Mustafa, Hemachandran, and Chola) emphasises three AI applications that emerging-market retailers prioritise: retail chatbots and virtual assistants for customer service in mixed-language markets (Indonesian, Hindi, Tagalog, Malay); customer-segmentation analytics using K-means and DBSCAN clustering on transaction logs; and visual merchandising automation using computer vision to plan store-shelf layouts. The volume's case observation: emerging-market retailers tend to leapfrog the Western analytics-warehouse pattern and go directly to cloud-native, mobile-first AI deployments.

8.7 Stitch Fix: the cautionary footnote

Stitch Fix, the founding case of hybrid-human-plus-ML retail, struggled through 2023–2025: FY2024 revenue fell 16% to $1.34B with active clients down to 2.5 million. The case shows that AI personalisation is necessary but insufficient when assortment and brand are weak.

Manufacturing's AI moment is humanoid robotics — Figure 02 working at BMW Spartanburg from August 2024 — backed by a longer history of generative design, predictive maintenance, and digital twins.

9.1 Humanoid robotics on the factory floor

Figure 02 began commercial deployment at BMW's Spartanburg plant in August 2024 — the first humanoid robot to perform real production work for an automaker, handling sheet-metal parts in body-in-white. Figure 03 launched 9 October 2025 with the explicit goal of building 100,000 robots over four years. Apptronik partnered with Mercedes-Benz and GXO Logistics, raising $403 million in February 2025; Agility Robotics' Digit is in Amazon and GXO warehouses; Tesla Optimus Gen 3 targets early 2026 for limited deployment.

9.2 Generative design and digital twins

Siemens NX Industrial AI automates CAE (computer-aided engineering) with up to 88% noise-prediction error reduction (BMW pilot) and ~15× speedups for non-engineers using design copilots. Autodesk's Project Bernini generates engineering-grade 3D models with manufacturing constraints baked in; the General Motors lightweighting case using earlier Autodesk generative design produced parts 40% lighter and 20% stronger.

Digital-twin platforms (Siemens, Dassault, GE Digital, Microsoft Azure Digital Twins, Nvidia Omniverse) are now standard for new factory commissioning. Unilever's digital twin pilot was credited with a 1–3% productivity improvement across pilot sites — small as a percentage, large as an absolute on a multi-billion-dollar manufacturing footprint.

9.3 Predictive maintenance

The foundational deployment is GE Aviation's predictive maintenance via FlightPulse and Asset Performance Management on the GE9X engine, which records terabytes of sensor data per flight. Across the broader industrial economy, predictive maintenance is the most consistently profitable AI use case in manufacturing — typical results are 10–20% reduction in unplanned downtime and 5–10% reduction in maintenance cost.

9.4 Freeport-McMoRan: the Bagdad concentrator

Already introduced as the Rewired anchor case (§4.9.1), Freeport-McMoRan's Bagdad concentrator throughput improvement of approximately 5% using ML on existing sensor data — without new capital expenditure — is now the most-cited single industrial-AI cost-recovery case study. The lesson: the AI factory's data pipeline can run on the data you already have, often for the cost of a few months of cloud compute.

9.5 The Hemachandran chapter on supply chain and manufacturing

The volume's chapter on AI in supply chain, logistics, and manufacturing (Choudhury and Gorantla) emphasises three Indian-context applications that generalise to other Asia-Pacific economies: demand forecasting in fragmented retail networks (modern trade and traditional kirana / sari-sari retail coexist); cold-chain monitoring with IoT sensor fusion; and last-mile route optimisation in unstructured urban environments where GPS and standard mapping data are unreliable.

A separate chapter — Coskun's Bayesian machine learning approach for evaluating an order-fulfilment reengineering project in downstream oil and gas — is methodologically distinctive. The case applies Bayesian belief networks to project the probability of cost overruns and schedule slippage in supply-chain transformation projects, with explicit prior elicitation from industry experts. The technique is unusual in the broader practitioner literature but increasingly relevant where ML deployment success is itself probabilistic and historical-data-thin.

Three sectors organised together because they share an architectural pattern: AI as the optimisation layer over a previously-human creative or operational decision.

10.1 Marketing — the longest deployment history

Programmatic ad-tech (DSPs, SSPs, RTB) has been ML-driven since the early 2010s. The 2023–2026 development is generative AI for content production at scale.

The Hemachandran & Rodriguez volume's marketing chapter (Manoharan, Durai, Ashtikar, Kumari) emphasises that chatbot-mediated marketing — particularly for SMEs in mixed-language Asian markets — has produced the highest measured ROI of any generative AI application as of 2025. The architecture is simple: an LLM-fronted FAQ, plus warm-handoff to a human, on top of an existing CRM.

10.2 Media — Netflix as the canonical anchor

Netflix's recommendation system, profiled extensively in Chapter 3, is the longest-running AI factory in commercial media. Its architecture produces approximately 80% of viewing originations from algorithmic recommendation rather than search.

Spotify's Discover Weekly, launched in July 2015, was the playbook for music. The 2023–2026 evolution is in generative content: Suno and Udio for music generation, Runway and Sora for video. The legal status of training data for these systems is in active litigation; the technology stack is well-developed; the commercial model is unstable.

10.3 News, media, and the displacement debate

Mata v. Avianca, S.D.N.Y., 22 June 2023 — two attorneys sanctioned $5,000 for filing a brief with six fictitious citations generated by ChatGPT. The case is a touchstone for media-and-AI displacement debates because it demonstrates that the failure mode is not capability but verification.

10.4 Energy — the largest near-term value pool

Energy is where AI-and-physical-infrastructure integration will produce the largest economic gains over the next five years. Three deployments anchor the discussion:

1. DeepMind's data-centre cooling work (2016) reduced Google's data-centre cooling energy by 40%. The deployment generalised: by 2018, the system was running autonomously, and by 2020 the same architecture was being applied across Google's broader data-centre fleet.
2. Tesla Autobidder manages large battery storage assets (e.g., Hornsdale Power Reserve in South Australia) using ML for arbitrage, frequency regulation, and grid services.
3. National Grid ESO and similar TSO deployments are using ML for demand forecasting, congestion management, and reactive-power optimisation — increasingly important as renewable penetration rises.

The Hemachandran & Rodriguez volume's energy chapter (Mathur, Hemachandran, Shanmugarajah) emphasises three deployment patterns: solar-plant performance optimisation, oil & gas reservoir characterisation using deep learning on seismic data, and grid-edge demand response. The emerging-economy emphasis is on the second pattern — particularly important for South-East Asian and Middle Eastern hydrocarbon economies during the energy transition.

10.5 The unifying observation

The three sectors illustrate Iansiti and Lakhani's Rule 2: capabilities are increasingly horizontal and universal. The same recommendation architecture that powers Netflix is in Spotify, Pinterest, TikTok, and YouTube. The same forecasting architecture that powers Walmart's demand model is in National Grid's load forecasting. AI factory components transfer; vertical knowledge increasingly does not constitute the moat.

Three sectors where AI value capture has been comparatively quiet, durable, and large — and where developing-economy deployments matter as much as developed-economy ones.

11.1 Logistics

UPS ORION (On-Road Integrated Optimisation and Navigation) is the canonical pre-deep-learning logistics AI: a route-optimisation system deployed across 55,000 routes in North America by 2016, saving an estimated 100 million miles per year and 10 million gallons of fuel — a $300–400M annual operating saving achieved with classical operations research methods.

The 2024–2026 evolution is multimodal. Maersk uses AI for vessel routing and port-time optimisation; FedEx applies computer vision to package-condition assessment; Flexport's 2024 GenAI platform automates customs-documentation drafting. Amazon Robotics (already covered in Chapter 8) is the warehouse counterpart.

11.2 Agriculture

Agriculture's AI value pool is dominated by precision-application and yield-prediction systems. John Deere See & Spray deploys onboard computer vision to apply herbicide only to weeds — typical reductions of 60–70% in chemical use. Climate FieldView (Bayer), Granular (Corteva), and Indigo Carbon are agronomic-decision-support platforms providing field-level advisory at scale.

11.3 Professional services and the agentic frontier

Legal AI is anchored by Harvey AI, which by March 2026 reached an $11B valuation and approximately $200M ARR, with Allen & Overy (now A&O Shearman) as its first major law-firm customer. The Mata v. Avianca sanction order (S.D.N.Y., 22 June 2023) is the field's defining cautionary tale.

The Hemachandran & Rodriguez volume's legal chapter (Agarwal, Naidu, Swamy) emphasises three professional-services deployments that generalise across emerging-market jurisdictions: contract review and clause classification, regulatory-change monitoring, and vernacular-language legal-aid chatbots.

11.4 The Indian dairy case (Hemachandran & Rodriguez, Ch. 16)

11.5 Cross-sector synthesis

What unites logistics, agriculture, and professional services is that AI's value is captured primarily in operational efficiency rather than product innovation. The sectors are useful for teaching because they show that AI's economic impact is not confined to consumer-facing tech — and that some of the largest aggregate gains are in industries with low margins and modest individual deployment costs.

A consolidated tour of sectors that the Hemachandran & Rodriguez (2024) volume treats in depth — and that are less commonly covered in standard MBA AI texts.

12.1 Insurance

Jafar, Akhtar, and Johl's chapter traces three deployment waves: underwriting (ML risk scoring on alternative data), claims processing (computer vision for vehicle and property damage assessment, NLP for claim narrative analysis), and increasingly fraud detection using graph neural networks.

The Agrawal-Gans-Goldfarb system-solution argument from Chapter 5 — insurance shifting from risk transfer to risk prevention — is most fully realised here. Lemonade and Tractable are AI-native incumbents; Ping An (China) is the rewired-incumbent counterpart, with AI claims-handling speed for car insurance reportedly under three minutes.

12.2 Sports

Agarwal's chapter on AI in sports covers three categories: performance analytics (Hudl, Stats Perform, Catapult Sports), fan engagement, and refereeing decision support (VAR in football, Hawk-Eye in tennis and cricket). The most-cited case study remains Liverpool FC's collaboration with DeepMind on tactical analysis, published in Nature Communications (2021).

12.3 Tourism and hospitality

Nagar, Meghana, and Rout's chapter emphasises dynamic pricing, personalised itinerary recommendation (Mindtrip, Layla, Wonderplan), and destination-level demand forecasting. The case worth knowing for Asia-Pacific students is Marina Bay Sands' AI-driven yield management, which integrates room pricing, F&B reservations, gaming-floor capacity, and entertainment scheduling into a single optimisation surface.

12.4 Public sector

Merugu and Hemachandran's chapter covers four categories: service delivery automation, regulatory technology, infrastructure operations, and public-health surveillance.

The most consequential public-sector AI failures in recent years have been in service-delivery automation:

• Robodebt (Australia, 2016–2019) — automated welfare-recovery system that produced approximately 470,000 unlawful debt notices to social-security recipients. The 2023 Royal Commission found the scheme “crude and cruel”.
• Toeslagenaffaire / SyRI (Netherlands) — algorithmic fraud detection wrongly accused approximately 26,000 families of childcare-benefit fraud. The District Court of The Hague ruled SyRI unlawful in February 2020.

12.5 SMEs and AI adoption

Deepthi B and Bansal's chapter on AI adoption in SMEs is a systematic literature review and bibliometric analysis. Three findings worth memorising: (1) SME AI adoption is bimodal — a small leading cohort adopts at a rate similar to large enterprises; the long tail lags by 5–7 years. (2) The leading cohort overwhelmingly uses SaaS-mediated AI rather than custom development. (3) The most consistent barrier is not cost or talent but data readiness.

12.6 The cross-sector synthesis

The Hemachandran & Rodriguez volume's contribution to this textbook is breadth, not depth. The chapters surveyed here demonstrate that AI is being deployed in essentially every commercial and governmental sector worldwide. The empirical footprint of business AI as of 2026 is much larger than the venture-capital-funded coastal-American narrative suggests, and emerging-economy deployments increasingly drive the most consequential learning about deployment patterns.

By May 2026, the agentic AI question has moved from “is it possible” to “how does it scale through an enterprise without breaking governance.”

13.1 What separates agents from copilots

A copilot suggests; an agent acts. The architectural difference is that agents plan (decompose a goal into sub-goals), perceive (read state from a system), act (call tools or APIs that change state), and iterate (revise plans based on observed results). Copilots are step-level; agents operate over multi-step trajectories with reduced human supervision.

13.2 The methodological substrate

Three papers define the technical substrate: Yao et al., ReAct (ICLR 2023) — interleaved reasoning traces with actions; Schick et al., Toolformer (NeurIPS 2023) — LLMs that can teach themselves to use tools; and Anthropic Model Context Protocol (MCP) (November 2024) — open standard for connecting LLMs to data sources and tools, now widely adopted.

13.3 The 2024–2026 launches

13.4 The McKinsey AI agent factory

The 2026 McKinsey AI Transformation Manifesto introduces the “AI agent factory” — the next-generation extension of Iansiti and Lakhani's AI factory. The pattern adds three new components to the original four: agent orchestration layer, tool registry and access governance, and observability and intervention.

McKinsey's reported deployments of this architecture: 40–70% productivity improvement at one large financial services firm running a greenfield payment system; 50% increases in productivity at LATAM Airlines with smaller engineering teams.

13.5 Where agents are working in 2026

• Customer service. Klarna's reversal notwithstanding, Salesforce, Zendesk, and Sierra are deploying customer-service agents at meaningful scale.
• Software engineering. Cursor, GitHub Copilot Workspace, and Devin operate in this space. Multi-file refactors, test generation, and routine bug fixes are widely automated; design and architecture remain human.
• IT operations. Microsoft Copilot for Security and Google's Sec-PaLM operate as Tier-1 SOC analysts.
• Sales operations. Salesforce Sales Agent, Outreach AI Sales Agent, and Gong's autonomous outreach agents.
• Procurement. Pactum's autonomous supplier negotiations close 68% of cases at ~3% cost savings.

13.6 Where agents are not working — yet

The Gartner forecast that more than 40% of agentic AI projects will be cancelled by end of 2027 reflects three persistent failure modes: reliability ceilings (a 95%-correct agent in a 10-step workflow has a 60% chance of completing without error), authority ambiguity (who is responsible when an agent acts?), and observability immaturity.

13.7 The reasoning-models inflection

OpenAI o1 (September 2024) and DeepSeek-R1 (January 2025) introduced inference-time reasoning — models that “think” before they answer. This category materially extends what agents can do: complex multi-step reasoning, verifiable mathematics, and code that compiles and passes tests on the first attempt all improve substantially with reasoning models behind the agent.

The cost structure shift matters. Reasoning models burn far more inference compute per query than classical LLMs — and the 280× drop in inference cost over 2022–2025 is partly absorbed by this shift. The net economics depend on the task: for high-stakes, low-volume tasks (code, finance, legal), reasoning models pay back; for high-volume, low-stakes tasks (FAQ chatbots), they do not.

The 2024–2026 governance environment is the first in which enterprise AI deployment is materially shaped by binding regulation. This chapter takes up the EU AI Act, the five-pillar operational responsible-AI framework, sovereign AI, and Iansiti and Lakhani's five ethical categories.

14.1 The five operational pillars

Across the major frameworks (EU AI Act, NIST AI RMF, UK Pro-Innovation, ISO 42001, Microsoft AI Principles, Google AI Principles, IBM AI Ethics), five operational pillars consistently appear:

1. Fairness — demographic-bias testing, disparate-impact analysis, ongoing fairness monitoring post-deployment.
2. Transparency — model cards, system cards, decision explainability appropriate to the stakes of the decision.
3. Accountability — clearly assigned responsibility for AI decisions, recourse mechanisms for affected individuals.
4. Privacy and security — data minimisation, differential privacy where applicable, adversarial robustness testing.
5. Human oversight — human-in-the-loop or human-on-the-loop for material decisions; ability to override or appeal.

14.2 The EU AI Act

The EU AI Act came into force on 1 August 2024, with risk-tiered obligations phasing in through 2025–2027. The four-tier structure:

• Unacceptable risk (banned): social scoring, real-time biometric identification in public spaces (with narrow exceptions), emotion recognition in workplaces and schools.
• High risk (heavily regulated): credit scoring, employment-screening AI, education-admissions AI, critical-infrastructure AI, law-enforcement AI.
• Limited risk (transparency obligations): chatbots, deepfakes.
• Minimal risk (light obligations): most consumer applications.

Penalties are calibrated to GDPR-style proportions: up to €35M or 7% of global annual turnover for prohibited-AI violations, up to €15M or 3% for high-risk violations. The Act's extraterritorial reach — it applies to any provider placing AI systems on the EU market — makes it the de-facto global benchmark.

14.3 The Iansiti-Lakhani ethical categories

14.3.1 Digital amplification

Frictionless networks amplify whatever is on them — including misinformation, polarisation, and engagement-optimised emotional content. The 2018 Cambridge Analytica scandal at Facebook, the algorithmic-amplification findings in the 2021 WSJ Facebook Files, and the YouTube radicalisation literature document the empirical pattern.

14.3.2 Bias

Buolamwini and Gebru's 2018 Gender Shades study showed face-classification accuracy disparities of up to 34% between lighter-skinned men and darker-skinned women — a finding that catalysed the entire algorithmic-fairness research programme.

14.3.3 Cybersecurity

AI both expands the attack surface and provides new attack tools (deepfake-mediated social engineering, automated vulnerability discovery, prompt injection). The 2025 Anthropic Antigravity prompt-injection demonstration illustrates the new class of risk.

14.3.4 Control

When algorithms drive the operational critical path, the locus of decision-making authority is unclear. The Robodebt and Toeslagenaffaire cases (Chapter 12) show that the responsibility-attribution problem is the largest source of public-sector AI failure.

14.3.5 Inequality

The seven US hyperscalers control most global AI infrastructure spending; the top quintile of AI-using firms capture 16–30% productivity gains while the bulk of firms capture single digits.

14.4 Sovereign AI

The sovereign-AI movement — domestic AI infrastructure, training compute, and frontier models — has progressed substantially in 2024–2026. Notable national initiatives include the EU AI Continent Action Plan (€200B), India's IndiaAI Mission ($1.25B), the UAE's Falcon series and G42 partnerships, Saudi Arabia's HUMAIN ($23B), Singapore's SEA-LION (Southeast Asian languages), and Japan's Sakana AI.

14.5 Microsoft's six AI principles

The principles are notable not for their content — they are similar to most enterprise AI ethics policies — but for their operational integration: CELA team members participate in development and sales decisions, rather than reviewing finished work. This is the Iansiti-Lakhani lesson: AI ethics works when it is built into the AI factory, not bolted on after the model ships.

The most-debated empirical questions in business AI are about labour displacement, the productivity J-curve, and ROI. This chapter introduces Brynjolfsson's Suitability for Machine Learning rubric, the J-curve evidence, and what we now know about AI ROI as of 2026.

15.1 Brynjolfsson's Suitability for Machine Learning rubric

Erik Brynjolfsson and Tom Mitchell (Science, 2017; AEA Papers and Proceedings, 2018) developed a 23-criterion rubric for scoring whether a task is suitable for machine learning. Applied to all 18,156 tasks in the US Department of Labor's O*NET database, the rubric produces task-level Suitability for Machine Learning (SML) scores that aggregate to the occupation level.

Four findings from the original work that every business student should memorise:

1. ML affects different occupations than earlier automation waves.
2. Most occupations include at least some SML tasks. Almost no job is fully automatable; almost no job is fully untouchable.
3. Few occupations are fully automatable using ML. The pattern is task-level transformation, not occupation-level replacement.
4. Realising ML's potential usually requires redesign of job task content.

15.2 The eight key SML criteria

Within the 23-criterion rubric, eight criteria are most diagnostic:

1. Task involves a well-defined input-output relationship.
2. Task has clearly definable goals and metrics.
3. Large datasets exist or can be created.
4. Feedback is provided rapidly.
5. Task does not require long chains of complex reasoning.
6. No need to explain how the decision was made.
7. No need for high-bandwidth interaction with humans.
8. Tolerance for error / acceptable performance levels can be specified.

15.3 The Brynjolfsson productivity J-curve

The J-curve hypothesis explains why the macro AI productivity slowdown of 2017–2022 is consistent with the AI capability explosion of the same period: conventional productivity statistics undercount intangible investment in the early phase of a general-purpose technology, then overcount its gains in the harvest phase.

The 2024–2026 evidence is starting to show the harvest phase. McKinsey's 2025 work on AI-augmented software development reports 16–30% productivity improvements at top-quintile firms, with 31–45% improvements in software quality.

15.4 The empirical labour-market evidence

The most-cited 2023 study is Brynjolfsson, Li, and Raymond's customer-support deployment, in which generative AI assistance produced a 14% productivity gain on average, with 34% gains for novice and low-skilled workers and minimal gain for experienced workers. The pattern — AI compressing the skill distribution — is now widely replicated.

Subsequent studies:

• Noy and Zhang (Science, 2023) — ChatGPT improved professional writing speed by 40% and quality by 18%.
• Peng et al. (2023) — GitHub Copilot increased programmer task-completion speed by 55.8%.
• Dell'Acqua et al. (2023) — BCG consultants with GPT-4 outperformed those without by 12% on average across 18 tasks, but performed worse on tasks where the AI was unreliable. The “jagged frontier” finding.

15.5 ROI reality as of 2026

The McKinsey 2025 finding: only the top quintile (~5–6%) of firms achieve enterprise-level EBIT impact from AI deployments. For everyone else, AI is a productivity tool that pays back at the project level but does not move the financial needle at the firm level.

The pattern is consistent with the Iansiti-Lakhani-Brynjolfsson synthesis: AI ROI is bottlenecked by complementary intangible investment, not by AI capability.

15.6 The labour reallocation question

Acemoglu and Restrepo's framework of automation-versus-augmentation labour effects is the most cited theoretical lens. Acemoglu's 2024 paper estimates total-factor-productivity gains from AI of approximately 0.66% over a decade — at the low end of widely circulated figures. Goldman Sachs' 2023 estimate of $4.4 trillion in annual generative-AI value sits at the high end.

A working manager needs answers to three diagnostic questions: where is my firm now, which tasks should we prioritise, and what should the next twelve months look like?

16.1 The Gartner AI maturity model

Gartner reports that 34% of low-maturity and 29% of high-maturity organisations cite data quality as a top challenge — even advanced organisations face fundamental data issues. The level-2-to-3 transition is the most common stalling point.

16.2 MIT CISR: four challenges in moving from stage 2 to 3

MIT Center for Information Systems Research's longitudinal work with 721 firms identifies four challenges in the level-2-to-3 jump: (1) moving from individual pilots to integrated production deployments; (2) establishing governance and risk-management frameworks; (3) integrating AI capabilities across the enterprise; (4) measuring and communicating business value beyond technical metrics.

MIT CISR's longitudinal finding: organisations in advanced AI maturity stages perform above industry average financially, while early-stage organisations perform below average.

16.3 The transformation roadmap synthesis

Combining Gartner, MIT CISR, AIMAA, and Rewired produces a working roadmap for the established firm in 2026. The five steps:

1. Diagnose maturity honestly. Use Gartner's 5-level model. Most firms are level 1 or 2, even if they say they are level 3 or 4.
2. Pick 2–5 domains. Per Rewired, large enough to matter, small enough to transform. Apply the Brynjolfsson 8-criterion SML test to candidate tasks within each domain.
3. Build the AI factory. Per Iansiti and Lakhani, the four components.
4. Build the six Rewired capabilities. All six, in concert.
5. Govern from day one. Per the EU AI Act and the operational five-pillar framework.

16.4 The diagnostic table

A consolidated reference of the conceptual frameworks introduced across the book.

17.1 Davenport-Ronanki: the three buckets

The framework's value is in showing that the highest-deployed bucket (process automation) is not the highest-value bucket (cognitive insight) — and that the most-talked-about bucket (cognitive engagement) is also the smallest.

17.2 The Iansiti-Lakhani AI factory (Chapter 3)

Four components: data pipeline, algorithm development, experimentation platform, software infrastructure. Output: predictions, pattern recognition, process automation. The factory creates a virtuous cycle.

17.3 Iansiti-Lakhani's new meta — the five rules (Chapter 5)

1. Change is no longer localised; it is systemic.
2. Capabilities are increasingly horizontal and universal.
3. Traditional industry boundaries are disappearing; recombination is the rule.
4. From constrained operations to frictionless impact.
5. Concentration and inequality will likely get worse.

17.4 Rewired's six capabilities (Chapter 4)

(1) Roadmap. (2) Talent — the 30–70 shifts. (3) Operating model — agile pods. (4) Technology. (5) Data products. (6) Adoption.

17.5 Agrawal-Gans-Goldfarb's three solution layers (Chapter 5)

(1) Point solution. (2) Application solution. (3) System solution.

17.6 The Brynjolfsson-Mitchell SML rubric (Chapter 15)

23 task-level criteria, applied to all O*NET tasks. Eight key criteria distinguish high-SML from low-SML tasks.

17.7 The Brynjolfsson-Rock-Syverson productivity J-curve (Chapter 15)

Conventional productivity statistics undercount intangible investment in the early phase of a general-purpose technology, then overcount its gains in the harvest phase.

17.8 Gartner's five maturity levels (Chapter 16)

Awareness → Active → Operational → Systematic → Transformational. Most firms are at level 1 or 2.

17.9 The five operational responsible-AI pillars (Chapter 14)

Fairness, transparency, accountability, privacy/security, human oversight.

17.10 The one-page diagnostic

Twelve cases for classroom use. Each is presented with a short summary, the framework lens that makes it pedagogically useful, and three discussion questions.

18.1 AlphaFold — from CASP14 to Nobel

Lens: Iansiti-Lakhani Rule 2 (horizontal capabilities). Question 1: Why did DeepMind succeed where decades of structural biologists had not? Question 2: The 2024 Nobel was awarded to Hassabis, Jumper, and Baker. What does this signal about how scientific recognition is shifting? Question 3: Construct the AI factory architecture for Isomorphic Labs.

18.2 Ant Group — the reference case for digital scale

Lens: Iansiti-Lakhani's headline case. Question 1: 10,000 employees, 700M users, three-minute one-click loans. Decompose the operating model. Question 2: What constraints from Chinese regulation now bind Ant Group? Question 3: Could a Southeast Asian or Indian fintech replicate Ant's architecture?

18.3 DBS — the canonical Rewired case

Lens: Rewired's six capabilities, fully implemented. Question 1: 15% to 90% in-source tech in 6 years. What was the binding constraint at each phase? Question 2: The 2-in-a-box platform leadership model is unusual. What does it enable? Question 3: Apply the DBS playbook to a regional bank in your country.

18.4 Khanmigo — AI tutoring at scale

Lens: Davenport-Ronanki cognitive engagement. Question 1: What does Khanmigo demonstrate about consumer-facing LLM products versus enterprise-grade AI tutors? Question 2: Is AI tutoring a Brynjolfsson-style augmentation case or a substitution case? Question 3: Construct the metrics by which Khanmigo's value should be measured.

18.5 Klarna — launch and reversal

Lens: The canonical 2024–2025 cautionary case. Question 1: 2.3M conversations in month one, 700 FTE-equivalent automation. What KPI was Klarna optimising, and what was missing? Question 2: The May 2025 reversal preceded the September 2025 IPO by four months. Read this as a corporate-finance signal. Question 3: Construct a redesigned Klarna service operating model.

18.6 Mata v. Avianca — the verification problem

Lens: Cognitive engagement failures and the verification layer. Question 1: Why did experienced lawyers file a brief with six fictitious citations? Question 2: Construct the Harvey AI architectural response. Question 3: What does the Mata case suggest about the diffusion path for generative AI in regulated industries?

18.7 JPMorgan COiN and the LLM Suite — sustained scale

Lens: Iansiti-Lakhani AI factory in financial services. Question 1: COiN was deployed in 2017. What did the bank invest in over the following eight years to maintain leadership? Question 2: The LLM Suite reaches 230,000 employees. What is the unit cost per employee, and what is the economic value per employee? Question 3: Compare COiN to Watson Health.

18.8 BMW + Figure 02 — humanoid robotics enters production

Lens: Rule 3 (industry boundaries dissolve). Question 1: Why body-in-white first, rather than final assembly? Question 2: Figure 03 targets 100,000 robots over four years. What is the implied unit economics? Question 3: Map this against the Brynjolfsson SML rubric.

18.9 Freeport-McMoRan — value from data already collected

Lens: Rewired's domain-based approach. Question 1: ~5% throughput improvement at Bagdad with no new capex. Why was this not done five years earlier? Question 2: The Bagdad pilot funded a broader transformation. What does this teach about transformation funding? Question 3: Construct an analogous pilot in a sector you know.

18.10 Zillow Offers — a high-profile ML failure

Lens: The limits of point and application solutions. Question 1: Zillow's iBuyer business shut down in November 2021 with $304M in writedowns. The model was technically functional. What broke? Question 2: Frame this as a system-solution failure. Question 3: Compare with Opendoor, which survived. Why?

18.11 DeepSeek — January 2025

Lens: Iansiti-Lakhani Rule 2 and the commoditisation thesis. Question 1: DeepSeek-R1 was released MIT-licensed at training cost reportedly under 1% of OpenAI o1's. What does this do to the moat-economics of frontier AI? Question 2: Why did Nvidia lose $600B on 27 January 2025 — but reach $5T by 29 October? Question 3: Construct the implications for sovereign AI in your country.

18.12 Coca-Cola Create Real Magic — generative AI for participation

Lens: Davenport-Ronanki cognitive engagement, Iansiti-Lakhani Rule 4. Question 1: 120,000 user-generated images. Why did Coca-Cola control the editorial gate rather than fully automate? Question 2: Is this a sustainable model, or a one-off campaign novelty? Question 3: Construct an analogous campaign for a brand in your country.

References are organised by domain. Tags indicate the chapter(s) most relevant to each entry.

Foundational frameworks

1. Iansiti, M. & Lakhani, K. R. (2020). Competing in the Age of AI: Strategy and Leadership When Algorithms and Networks Run the World. Harvard Business Review Press. Ch. 3, 5
2. Lamarre, E., Smaje, K. & Zemmel, R. (2023). Rewired: The McKinsey Guide to Outcompeting in the Age of Digital and AI. Wiley. Ch. 4, 16
3. Hemachandran, K. & Rodriguez, R. V. (eds.) (2024). Artificial Intelligence for Business: An Implementation Guide Containing Practical and Industry-Specific Case Studies. Routledge. Ch. 6–12
4. Agrawal, A., Gans, J. & Goldfarb, A. (2022). Power and Prediction: The Disruptive Economics of Artificial Intelligence. Harvard Business Review Press. Ch. 5
5. Davenport, T. H. & Ronanki, R. (2018). Artificial intelligence for the real world. Harvard Business Review, January–February. Ch. 17

Productivity and the labour market

1. Brynjolfsson, E., Mitchell, T. & Rock, D. (2018). What can machines learn, and what does it mean for occupations and the economy? AEA Papers and Proceedings, 108, 43–47. Ch. 15
2. Brynjolfsson, E. & Mitchell, T. (2017). What can machine learning do? Workforce implications. Science, 358(6370), 1530–1534. Ch. 15
3. Brynjolfsson, E., Rock, D. & Syverson, C. (2021). The productivity J-curve: How intangibles complement general purpose technologies. American Economic Journal: Macroeconomics, 13(1), 333–372. Ch. 15
4. Brynjolfsson, E., Li, D. & Raymond, L. R. (2025). Generative AI at work. Quarterly Journal of Economics, forthcoming (working paper, 2023). Ch. 15
5. Noy, S. & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. Ch. 15
6. Dell'Acqua, F. et al. (2023). Navigating the jagged technological frontier. Harvard Business School Working Paper 24-013. Ch. 15
7. Acemoglu, D. (2024). The simple macroeconomics of AI. NBER Working Paper 32487. Ch. 15

Foundational AI methodology

1. Vaswani, A. et al. (2017). Attention is all you need. NeurIPS. Ch. 2
2. Krizhevsky, A., Sutskever, I. & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. NeurIPS. Ch. 2
3. Devlin, J. et al. (2019). BERT: Pre-training of deep bidirectional transformers. NAACL. Ch. 2
4. Brown, T. et al. (2020). Language models are few-shot learners. NeurIPS. Ch. 2
5. Hoffmann, J. et al. (2022). Training compute-optimal large language models (Chinchilla). arXiv:2203.15556. Ch. 2
6. Yao, S. et al. (2023). ReAct: Synergizing reasoning and acting in language models. ICLR. Ch. 2, 13

Sectoral references

1. Abramson, J. et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630, 493–500. Ch. 7
2. Jumper, J. et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596, 583–589. Ch. 7
3. Linden, G., Smith, B. & York, J. (2003). Amazon.com recommendations: Item-to-item collaborative filtering. IEEE Internet Computing, 7(1), 76–80. Ch. 2, 8
4. Gomez-Uribe, C. A. & Hunt, N. (2015). The Netflix recommender system. ACM TMIS, 6(4). Ch. 3, 10
5. Buolamwini, J. & Gebru, T. (2018). Gender Shades: Intersectional accuracy disparities in commercial gender classification. FAT*. Ch. 14

Maturity and governance

1. Gartner (2024). AI Maturity Model. Ch. 16
2. MIT CISR (2024). Building organizational AI maturity (longitudinal study, 721 firms). Ch. 16
3. European Commission (2024). Artificial Intelligence Act (Regulation (EU) 2024/1689). Ch. 14
4. NIST (2023). AI Risk Management Framework (AI RMF 1.0). Ch. 14
5. ISO/IEC 42001 (2023). Information technology — Artificial intelligence — Management system. Ch. 14

Industry research and trade press

1. McKinsey & Company (2024–2025). The state of AI (annual survey series). Ch. 1, 4, 15
2. McKinsey & Company (2026). The AI transformation manifesto: 12 themes driving growth. Ch. 4, 13, 15
3. McKinsey & Company (2025). Unlocking the value of AI in software development. Ch. 13, 15
4. McKinsey & Company (2023). Rewired in action: Digital and AI transformations. Ch. 4
5. Stanford HAI (2025). AI Index 2025. Ch. 1, 5, 14
6. Menlo Ventures (2025). 2025: The state of generative AI in the enterprise. Ch. 5
7. Foundation Capital (2026). Where AI is headed in 2026. Ch. 5
8. Gartner (2024). Predicts 2025: 30% of generative AI projects abandoned by 2025. Ch. 1, 13
9. Deloitte (2024). State of generative AI in the enterprise, Q4 2024. Ch. 1
10. Bank for International Settlements (2024). Project Aurora. Ch. 6

Cases referenced in Chapter 18

1. Hippocratic AI (2024–2025). Polaris constellation deployment data. Ch. 7, 18
2. Klarna AB (2024). One year of Klarna AI assistant. Press release, 28 February 2024. Ch. 8, 18
3. Klarna AB (2025). CEO interview, May 2025; reversal of full-AI customer service strategy. Ch. 8, 18
4. Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), 2023 WL 4114965 (S.D.N.Y. 22 June 2023). Ch. 10, 11, 18
5. JPMorgan Chase (2017). Annual report; COiN platform disclosure. Ch. 6, 18
6. Royal Commission into the Robodebt Scheme (Australia, 2023). Final report. Ch. 12, 14
7. District Court of The Hague (2020). NJCM v. Netherlands (SyRI). Judgment of 5 February 2020. Ch. 12, 14