Key Takeaways
Rebuild workflows around AI instead of sprinkling AI onto old processes
Change the operating model. The book contrasts a “+AI” approach, which adds AI to existing processes, with “AI+,” which starts by reconsidering how work should be done. Break workflows into smaller tasks, assign routine work to AI where appropriate, and design human responsibilities around the new process. The authors compare the moment to the early public internet: access to a new tool matters less than what organizations build with it.
Start with the foundations. Their updated AI Ladder begins with an information architecture: the systems and practices used to collect, organize, protect, and govern data. From there, organizations can add AI to applications, automate workflows, and redesign processes. The aim is not to remove people by default, but to stop treating old procedures as fixed simply because they already exist.
The distinction is useful because it shifts attention from purchasing models to changing the work those models support. It also resembles process redesign in operations research: automating a poor process can make its errors faster and harder to notice. The AI-first framing is strongest when tied to a clear outcome and careful task analysis. It should not become a slogan for removing human judgment from work where context, accountability, or care matter. The practical test is whether the redesigned process improves service, quality, or capacity, not whether it contains more AI.
Make proprietary data your advantage, not a free input to someone else's model
Create value, not just access. The authors distinguish an AI User, who relies on embedded software or another provider’s model, from an AI Value Creator, who uses a platform to adapt models with organizational data and governance. They estimate that at most about 1% of enterprise data is represented in commonplace LLMs, leaving specialized knowledge largely outside them. L’Oréal, for example, is developing a private model to help its researchers use formulation and materials knowledge.
Choose how data enters. Retrieval-augmented generation, or RAG, supplies relevant documents to a model when it answers a question. Fine-tuning changes model parameters using new examples. InstructLab offers a collaborative way to add skills and knowledge using curated examples and generated training data. These methods serve different needs. Sensitive or frequently changing information may be better retrieved under access controls than embedded in model weights.
The central insight is that data ownership and model access are different forms of power. A company can use a public model without surrendering its most valuable information, but only if vendor terms, access controls, and data flows are understood. Proprietary data is not automatically a moat: it can be stale, biased, or poorly documented. Its advantage comes from relevance, quality, and the ability to apply it responsibly. The strongest data strategy therefore combines competitive use with stewardship, rather than treating secrecy alone as a source of value.
Start with internal automation, then reinvest the savings in growth
Classify the business case. The book asks leaders to distinguish spending to save money, called renovation, from spending to make money, called innovation. It then sorts AI’s role into automation, optimization, or prediction. A potato processor facing a crop disease used AI-assisted inspection and washing to reduce reliance on scarce manual labor, then directed savings toward export relationships and new markets.
Earn experience safely. For organizations starting out, the authors recommend low-risk internal automation before public-facing or high-stakes applications. An employee vacation-booking assistant is one example: it can simplify a cumbersome HR process, while mistakes are contained and correctable. The larger strategy is to shift routine work earlier in the process, then use the released time and budget for more ambitious changes. The authors call that movement from shifting left to shifting right.
This sequence links risk management to investment strategy: an internal pilot can build capability while producing a visible operational benefit. It also counters the tendency to judge AI only by headcount reduction. Savings matter most when they create room for better service, new products, or more skilled work. Still, savings should be measured against the full cost of implementation, including data preparation, supervision, and maintenance. A pilot that is safe but never scales can become another technology demonstration rather than a business improvement.
Find horizontal AI patterns before tailoring them to your industry
Reuse the underlying capability. Many AI tasks recur across industries: understanding language, classifying images, summarizing documents, and predicting patterns. The authors call these horizontal use cases because the same technical pattern can support different businesses. Their leading examples are customer care, software development, and digital labor, meaning software that performs meaningful parts of a work process alongside people.
Move from trials to value. Their AI Value Creation Curve begins with experimentation, proceeds to using organizational data, and reaches a Value Tipping Point where automation and digital labor can drive broader change. Examples across sectors include early detection of diabetic foot ulcers, AI assistance with clinical trial recruitment, and precision spraying that targets crop disease. Synthetic data, artificially generated information used to supplement or replace real examples, can also help with privacy or scarce training data, though it requires bias checks.
A horizontal lens helps organizations borrow tested patterns instead of inventing every application from scratch. It is similar to identifying a reusable mechanism in engineering before adapting it to local constraints. Yet the same technical task can carry very different consequences by setting: a mistaken image classification in manufacturing is not equivalent to a mistaken medical assessment. Reuse should apply to components and learning, not to assumptions about acceptable error, oversight, or impact. Industry knowledge remains essential for determining whether a generic capability actually solves the problem.
Treat trust as product infrastructure, not a compliance checkbox
Design for accountability. The book’s governance levers are fairness, robustness, explainability, and lineage. Fairness means checking whether systems treat groups unjustly. Robustness concerns resistance to manipulation and failure. Explainability makes decisions more understandable, while lineage records how data and models were developed and maintained. These safeguards matter because models can use outdated information, invent plausible answers, or inherit bias from training data.
Make risk visible. A legal team in the book relied on fabricated case citations generated by an LLM and faced sanctions after failing to verify them. The authors also describe a hiring model that can screen candidates unfairly and a public-facing airline chatbot that gave incorrect policy information. They argue for governing the full model lifecycle, from proposal and testing to deployment and continuous monitoring, and for scaling oversight to the potential harm of each use case.
Trust is not merely a feeling users have; it is an operational property built from evidence, controls, and recourse. This connects AI governance to safety engineering, where reliability depends on anticipating failure modes rather than assuming they will not occur. The book rightly emphasizes proportional oversight, but its call for explainability needs a practical interpretation: not every model can reveal a simple internal rationale. Useful accountability can also come from traceable sources, documented limits, independent testing, and a clear path to challenge consequential decisions.
Upskill the whole workforce, not just the technically curious few
Treat learning as operating capacity. The authors argue that AI skills must extend beyond data scientists to managers, business teams, and technical specialists. A durable program inventories skills, defines observable levels, provides role-specific learning, and makes practice and coaching routine. Because technical knowledge becomes outdated quickly, one-off training is not enough. The book recommends hiring for curiosity and digital fluency, while broadening recruiting beyond familiar degrees and institutions.
Make learning practical. In IBM’s Skills Challenge, employees had to complete training and then build AI prototypes. In 2024, about 160,000 employees participated, forming roughly 30,000 teams and submitting more than 12,000 prototypes. One project cut routine questions to an engineering support team from 116 hours of work each week to less than two minutes. The authors also point to ATM history: automating teller transactions changed the job, while teller employment and banking services expanded.
The book’s strongest workforce claim is that skills development is part of the technology system, not an optional benefit added after deployment. Practice, peer learning, and visible demonstrations make abstract familiarity more usable. The ATM comparison is suggestive rather than proof that AI will preserve employment: technologies differ, and transition costs fall unevenly across workers. A responsible skills strategy therefore needs more than optimism about new roles. It also needs worker participation, time to learn, credible pathways into changing work, and honesty about which tasks are likely to disappear.
Replace the one-model bet with efficient teams of specialized models
Match model to task. The book argues that no single model will suit every business need. Smaller language models can be cheaper to run and, when trained on high-quality or domain-specific data, can perform strongly on focused tasks. The authors describe model distillation, where a larger model generates examples that teach a smaller one, and model routing, where a system directs each request to a model expected to handle it well.
Use specialization deliberately. A Mixture of Experts model contains multiple parameter groups, activating only some for a given request. This can reduce inference costs, though the benefit depends on workload and system design. In a cited routing study, a library of models reached about 72% benchmark accuracy, compared with 68% for a single large model. A library restricted to models of 13 billion parameters or fewer reached 70%, showing a possible efficiency trade-off rather than a universal replacement for larger models.
The move from a single model to a portfolio resembles portfolio design in finance and ensemble methods in statistics: performance depends on matching tools to tasks, not finding one universal winner. Smaller models can improve cost, speed, and deployment flexibility, but benchmark scores do not guarantee reliability in a particular workplace. Distillation can also transfer weaknesses, safety gaps, or licensing constraints along with capability. The useful managerial question is therefore not simply which model is best, but which model is adequate for each task under acceptable cost, risk, and monitoring requirements.
Give agents bounded goals, tools, and human-supervised authority
Agents act toward outcomes. An AI agent uses a model to plan tasks, call tools such as web searches or databases, and revise its actions toward a goal. Unlike a chatbot that answers one prompt at a time, an agent may coordinate multiple steps or specialized models. The book’s examples include drafting a blog and related social posts, supporting clinical trial coordination, and helping analyze customer support tickets.
Autonomy needs limits. Agents can repeat actions in costly loops, expose sensitive information, or take steps their users did not intend. The authors recommend activity logs, runtime limits, interruption mechanisms, and human feedback. High-impact actions, such as sending mass communications or moving money, should require approval. Access should be restricted to the information and tools needed for the task, especially when multiple agents share work across sensitive systems.
Agents shift the design problem from answer quality to delegated authority. This makes them resemble junior employees: useful when goals are clear, permissions are bounded, and someone can inspect the work. The analogy also highlights a limitation. Agents do not share human accountability or judgment, even if their output looks deliberate. Logs and approval gates help, but they cannot compensate for vague objectives or excessive access. A sound deployment begins by defining what the agent may do, what it must ask first, and how a person can stop or reverse its actions.
Build LLM applications like software, with structure, tests, and runtimes
Move beyond giant prompts. The book’s closing proposal, generative computing, treats models as a new computing component alongside classical and quantum computing. Instead of passing sprawling text instructions to a model, developers could use structured inputs, explicit memory, safety checks, and a runtime that manages model calls and outputs. The authors argue that today’s prompt-heavy applications often become brittle: small wording changes can produce unexpected results, while parsing model responses creates extra complexity.
Program capabilities. The proposed alternative is to define capabilities in reusable code and generate training examples for them, then incorporate those capabilities into models without erasing existing ones. The approach connects to InstructLab’s use of recipes and generated examples. The authors also see inference-time reasoning and specialized hardware as part of this direction, since models may spend more computing effort evaluating possible responses. The concept is presented as an emerging architecture, not a settled standard.
This proposal reframes generative AI as an engineering discipline rather than a conversation trick. Structured interfaces, explicit error handling, and reusable components could make applications easier to test and maintain, much as software engineering matured around modular design. The challenge is that probabilistic behavior does not become fully predictable just because it is placed inside a runtime. Better structure can constrain and observe a system, but it cannot guarantee truth. The most valuable vision here is complementarity: use conventional code for exact control and models where language, ambiguity, or flexible pattern recognition are useful.
Analysis
The book’s central argument is organizational rather than narrowly technical: durable advantage comes from joining models to proprietary data, redesigned work, governance, and broad employee capability. Its business logic follows a sequence. Start with low-risk internal work, learn what the system can and cannot do, make data usable under appropriate controls, then expand into new workflows and services. This is more durable than betting on a particular model release, especially in a field where model costs, capabilities, and vendors change rapidly.
The most important tension is between speed and stewardship. The authors encourage urgency because productivity, labor supply, and competitive conditions create pressure to act. At the same time, they warn that AI can fabricate facts, reproduce unfairness, expose data, and expand security risks. Those ideas are not contradictory. They imply that responsible deployment must be designed into the pace of experimentation: low-risk trials, clear ownership, measurable outcomes, and rapid stopping mechanisms. The book is strongest when it translates that principle into operational choices, such as internal pilots, access controls, model monitoring, and human approval for consequential actions.
Its repeated emphasis on openness, enterprise platforms, and data ownership also reflects a particular business perspective. Open ecosystems can broaden innovation and reduce dependence on a single provider, but openness alone does not guarantee trustworthy data, durable governance, or fair outcomes. Nor does proprietary data automatically confer a meaningful advantage. The advantage arises when organizations can find, validate, protect, and apply relevant information better than competitors can.
The book’s broadest contribution is a practical change in unit of analysis: stop asking whether AI can perform a job, and examine which components of a workflow can be automated, augmented, or redesigned. That approach makes the promise more concrete and the risks more local. It also prevents a false choice between technological enthusiasm and caution. Organizations can move quickly while still distinguishing low-stakes convenience from decisions that shape health, employment, finances, or rights.
People Also Read
Glossary
AI Value Creator
Business that builds differentiated AIAn organization that uses an AI platform to combine models, its own data, and governance to create solutions tailored to its business. Rather than relying only on another provider’s general model, it can customize models and control how data is used, aiming to build and retain value specific to its operations.
AI+
AI-first workflow redesignThe book’s term for designing business processes with AI as a starting point, rather than adding AI onto existing procedures. The approach breaks work into smaller tasks, identifies where AI can contribute, and redesigns the human and automated parts of a workflow around desired business outcomes.
Information architecture
Systems for governed data useThe organizational and technical foundation for collecting, organizing, integrating, transforming, storing, protecting, and governing data. In the book’s framework, an information architecture helps make enterprise data accessible and trustworthy enough to support AI applications and explain their outputs.
Retrieval-augmented generation
Grounding answers in retrieved informationA method in which a system retrieves relevant information from an external source, such as a database or document collection, and adds it to a model’s prompt. The model can then use that context to produce an answer. The model’s parameters are not changed, and the external information can be updated without retraining the model.
InstructLab
Collaborative model customizationAn open source approach for adding skills and knowledge to language models through structured contributions. Contributors provide example-based recipes, which can generate synthetic training data for model alignment. The process is designed to make model customization more accessible and collaborative while reducing the risk that new training erases general capabilities.
Small language model
Compact task-capable language modelA language model smaller than many frontier models, often described in the book as having fewer than 13 billion parameters. Smaller models can cost less to run and may perform well on focused tasks when trained with high-quality or domain-specific data. Their usefulness depends on the task, training, and evaluation.
Model routing
Assigning requests to suitable modelsA system that selects among available models for each request, based on predicted task suitability. A router may send simple or specialized work to smaller, cheaper models and reserve larger models for tasks that require them. The goal is to balance accuracy, cost, speed, and resource use.
Mixture of Experts
Selective activation of model componentsA model architecture with distinct groups of parameters called experts. A routing mechanism activates a subset of these groups for a given input rather than using all parameters equally. This can improve inference efficiency, though actual savings depend on factors such as memory use, batching, and which experts are selected.
Digital labor
Software performing workflow tasksSoftware-based labor that can independently complete meaningful parts of complex processes, often using AI to interpret requests, sequence tasks, and interact with other systems. Digital labor is intended to work alongside human employees, handling routine work so people can focus on tasks requiring more judgment or expertise.
Generative computing
Structured computing with generative modelsThe book’s proposed computing paradigm in which language models operate as components within structured software systems. Rather than relying on unstructured prompts alone, applications would use runtimes, explicit inputs, memory, safety checks, and reusable model capabilities. The idea is presented as a developing direction that complements conventional computing.
Download PDF
Download EPUB
.epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.