Executives & Directors: the cheat sheet

The map: one idea per chapter

M00 · Course Introduction

1. Welcome and How This Program Works — Level 1 builds AI decision-makers, not engineers. Ten questions in order, one route for every role, and a method that turns reading into action.
2. Why Executives Need AI Literacy — Specialists can say what AI can do; only leaders decide what the organization does with it. Literacy is judgment, questions and accountability.
3. From AI Awareness to AI Leadership — Awareness changes what you know; leadership changes what you do. Four behaviors mark the shift: sponsor, ask for evidence, redesign work, own decisions.
4. Your AI Leadership Starting Point — Know where you start. Rate what you can explain and show, not what you feel, then choose one gap and one real problem to carry.

M01 · The AI Revolution

1. AI Is Changing Everything — Almost everyone has AI; few have changed the business. Advantage comes from redesigning the work around it, as it did with electricity.
2. The Four Industrial Revolutions — AI is new; the pattern is not. General-purpose technologies pay off only after organizations build the complements - processes, skills and new ways of working.
3. Why AI, Why Now? — AI is old research; readiness is new. Web-scale data, specialized chips, foundation models and easy access converged, for everyone at once.
4. The AI Adoption Curve — AI adoption is a curve of depth, not only spread. Your position depends on where the AI lives, not on how many people have access.
5. AI Leaders vs AI Followers — The best models sit close together, so behavior decides who leads: choose a real problem, own it, change the work and prove the result.
6. The AI Competitive Advantage — A rival can buy your AI model in weeks. Advantage lives in the layers around it, which are slow to copy and slower together.
7. The Speed of AI Adoption — Generative AI spreads at employee speed because a first try needs nothing new. Advantage goes to organizations whose own learning keeps up.
8. AI Is Already Inside Your Organization — AI is already inside the organization - embedded, official and informal. See it first, give it owners, then buy.
9. Generative AI Changes the Game — Generative AI did not invent AI. It made language the interface, so almost anyone can use it - widening opportunity and responsibility.
10. From Copilots to AI Agents — AI is moving from answers to action. Copilots draft while people act; agents act themselves, so permissions must match the action.
11. AI and the Future of Work — The unit of change is the task, not the job. Leaders choose automate or augment task by task, redesign roles and tell people the truth.
12. The AI Talent Gap — The AI talent gap is a missing capability across four layers, not a shortage of experts. You cannot hire your way out of it.
13. The AI Transformation Challenge — A working model is the easy part. Value needs six components holding together, and the weakest one sets the ceiling.
14. Where Should We Start With AI? — Where you start decides what your organization learns about AI. Start where the work matters, feedback is fast and mistakes can be undone.
15. Measuring AI Business Value — Activity is not value. AI is worth what the business measurably changes because of it, and only measurement shows which.
16. The Economics of AI — The model price is one line of the bill. Judge what an outcome costs, and give one owner the whole bill.

M02 · Understanding AI & Generative AI

1. What Exactly Is Artificial Intelligence? — AI is a label for a family of capabilities. Judge any system by what it does, how reliably, and what happens when it is wrong.
2. AI vs Automation — Automation follows rules people wrote. AI follows patterns learned from data. They fail differently, so design the simplest combination, with people owning exceptions.
3. AI vs Machine Learning — AI is the field; machine learning is one way to build it. Ask which parts learn, how they learn, and where their answer key comes from.
4. The Major Types of AI — AI labels answer four questions: what it does, how it is built, how general it is and how much it does alone.
5. How Machine Learning Learns — A model learns patterns by reducing its error on known examples. They pay only if they hold on unseen cases and keep holding.
6. Deep Learning — The Executive Mental Model — Deep learning is machine learning with many-layered neural networks that learn their own features from raw data. Depth is capacity, not a guarantee.
7. Training vs Inference — Training creates capability; inference delivers it on every request, without changing the model, and carries much of the running cost.
8. Data, Models and Compute — AI capability is data, model and compute working like factors in a product. The weakest corner sets the ceiling.
9. Foundation Models — A foundation model is trained broadly once, then adapted to many tasks. The model is the bottom layer; value and accountability sit above it.
10. Large Language Models — An LLM predicts the next token from context. It supplies the language; your systems of record supply the facts.
11. Tokens, Context and Embeddings — Tokens are what a model reads and bills; context is what it sees now; embeddings find the right content by meaning.
12. Multimodal AI — Multimodal AI works across text, images, audio and video. Add a modality when it holds information that changes the decision.
13. How Generative AI Works — Generative AI composes new output from learned patterns, drawing on chance at every step. It gives the likely answer, not the checked one.
14. Prompting and Context Engineering — Executive Mental Model — Prompting tells the model what to do; context engineering decides what it has to work with. Ask what the model could see.
15. RAG and Enterprise Knowledge — Executive Mental Model — Retrieve, don't retrain: your systems own the facts and hand the model only what is relevant, current and permitted, at question time.
16. From AI Assistants to AI Agents — An assistant answers a request. An agent pursues a goal through a loop of steps, tools and checks, so judge it on the whole job.

M03 · Business Value of AI

1. What Does AI Value Actually Mean? — AI value is the measurable change in business results that AI causes, compared with what would have happened without it, net of cost.
2. Where AI Creates Revenue — AI creates revenue only when it changes behavior. Count the incremental part, measured against what would have happened anyway.
3. Where AI Reduces Cost — AI reduces cost only when a named budget line falls, net of what AI costs, with quality held. Hours saved are capacity, not cash.
4. AI and Customer Experience — Better customer experience becomes business value only when customer behavior changes. Aim AI at the effort customers waste, wherever in the journey it costs most.
5. AI and Workforce Productivity — A faster task creates capacity, not productivity. It pays only when the whole job improves and management decides what the time is for.
6. From AI Capability to Business Outcome — A capability becomes valuable only when it causes a measurable change in a business outcome. Every arrow between the two is an assumption.
7. Baselines, Metrics and Measurement — AI value is the observed outcome minus a credible estimate of what would have happened without AI.
8. Productivity vs Realized Capacity — Time saved is potential value. It becomes realized capacity only when management decides where it goes and measures the result.
9. Leading vs Lagging AI Metrics — Leading metrics steer; lagging metrics judge. Link them in one chain, check that leaders really lead, and give guardrails early warnings.
10. Building the AI Business Case — A business case is an argument that an option beats business as usual on expected value, with its range, its breaking points and its evidence shown.
11. AI ROI and Value Realization — ROI is calculated twice, as a forecast and as a result. Benefit leaks while cost stays, so realization must be owned and judged forward.
12. Total Business Impact — Total impact maps every effect across six dimensions, then counts each economic benefit once, where the money lands and against the plan.
13. AI Value Scorecard — An AI value scorecard is one page that ends in a decision - realized against expected value, confidence on every figure, rules agreed in advance.
14. Module 03 Synthesis — Choosing Value Over Hype — Hype skips links between capability and cash. Walk every claim through definition, scope, adoption, capture and cost, then let evidence choose the verb.

M04 · AI Strategy

1. What Is an Enterprise AI Strategy? — An AI strategy is a few hard choices - a diagnosis, a guiding policy, refusals and shared capability - that trace back to the business strategy.
2. AI Strategy vs AI Adoption — Strategy is choices; adoption is behavior. Usage only counts when it changes an important workflow that serves a strategic priority.
3. Start With Business Strategy — Start AI strategy with how the business wins and what constrains it, then ask whether AI materially changes that constraint.
4. Defining AI Ambition — AI ambition is the choice of how much AI should change the business. The right level fits strategy, readiness, budget and risk appetite.
5. Finding Strategic AI Opportunities — Walk down from the strategy to where value leaks, size the pool roughly, then ask whether AI beats the best alternative.
6. Build vs Buy vs Partner — Buying is the default. Name the few capabilities to own, decide layer by layer, and keep a way out.
7. Centralized vs Federated AI — Ask where each AI capability belongs, not who owns AI. Centralize for scale and reuse; federate for context and ownership.
8. AI Platform Strategy — An AI platform exists to create leverage. Build the smallest shared platform that repeated demand justifies, and judge it by the effort it removes.
9. Data Strategy for AI — Having data is not enough. Start from the AI outcome you chose, find the data gap that blocks it, and fix that first.
10. AI Talent and Capability Strategy — Talent can be hired; capability has to be built. Blueprint the capabilities, set role targets, measure the gap, then decide the sourcing.
11. AI Operating Model — Strategy sets direction. The operating model - owners, rights, funding and routines - makes execution repeatable for every use case.
12. AI Decision Rights and Accountability — Deploying AI moves decision rights. Choose where they land - many responsible, one accountable owner, authority scaled to stakes, the right to stop assigned.
13. AI and Competitive Advantage — When rivals can buy the same AI, customers keep the gains. Advantage needs a gap in cost or value and a guard that keeps it open.
14. Proprietary Data, AI Moats and Differentiation — Unique data is a head start, not a moat. It defends a lead only if more of it keeps improving the product for customers.
15. Module 04 Synthesis — The Executive AI Strategy — An AI strategy is one chain of choices, from business problem to advantage, where every choice fits the others.

M05 · Enterprise AI Use Cases

1. The Enterprise AI Use-Case Landscape — Enterprise AI is five patterns repeated in every function. Read it by workflow and outcome, not by tool or department.
2. AI Use-Case Discovery and Design — AI projects fail on the problem, not the model. Map the work, give each task to a rule, AI or a person, then screen it.
3. AI in Finance — Finance AI should draft, forecast and detect at speed, while the authority to approve and pay is designed deliberately, never granted by default.
4. AI in Human Resources — Use AI freely for HR service work; when it shapes someone's job, pay or opportunity, a named person decides and outcomes are tested.
5. AI in Sales and Marketing — AI gives sellers time back and makes messages almost free. The value is better commercial decisions, approved claims and fair use of signals.
6. AI in Customer Service — Answering is not resolving. Service AI succeeds when the problem is fixed with less effort, within limits, and it knows when to hand over.
7. AI in Operations and Supply Chain — Operational AI ends in a physical move. A signal creates value only when someone may act on it in time.
8. AI in Software Engineering — AI makes writing code much faster. Value arrives only when verification, review and release keep pace, so measure delivery and fund the checks.
9. AI in Product Development — Ideas and prototypes are now nearly free. The scarce resource is evidence of what customers do, and people still choose what to build.
10. AI in Cybersecurity — Attackers use AI to go faster. Use it to decide faster, within written limits, and judge it by the risk it removes.
11. AI in Knowledge Management — AI makes knowledge cheaper to capture, combine and deliver. It cannot decide which knowledge is right - that is leadership work.
12. AI in Document and Data Processing — Document AI pays when trapped information becomes checked data a process uses, routed by error cost and traceable to its page.
13. AI in Prediction, Forecasting and Optimization — AI makes predictions cheap and ranged. Value comes when leaders plan for the range, price both errors and test causes before acting.
14. AI Agents and Intelligent Workflows — Score an agent on three outcomes - verified, handed off, silently wrong. Price the silent errors, write its mandate, design its handoffs.

M06 · AI Risks

1. The AI Risk Landscape — AI risk is the business impact when an AI-enabled system is wrong. Accuracy belongs to the model; risk belongs to the use.
2. Accuracy, Hallucination and Reliability — AI will be fluent and sometimes wrong. Reliability is designed around the model - measured on real work, grounded, checked and able to decline.
3. Bias and Fairness — Accurate is not the same as fair. Check what the system predicts, test results by group, choose the fairness objective and keep monitoring.
4. Privacy and Confidential Data — AI opens new paths for sensitive information. Design privacy into the whole data path instead of assuming it from an approved tool.
5. Security and AI Attacks — Language is now an attack surface. We cannot make AI impossible to fool, so we limit what a fooled system can reach and do.
6. Intellectual Property and Copyright — AI-generated does not mean we own it - or that anyone can. Rights depend on inputs, provider terms, law and human contribution.
7. Model and Third-Party Risk — AI dependency is business dependency. Critical AI providers need an owner, proportionate controls and a tested way out.
8. Operational and Workforce Risk — AI can make an organization faster and more fragile at once. Keep the human capability to check the AI and to run without it.
9. Agentic AI and Autonomous Actions — When AI can act, risk moves from what it says to what it does. Scope its authority and grant more only on evidence.
10. Human Oversight and AI Incidents — A human in the loop is not control. Oversight needs authority, information, time and tools, and incidents need a rehearsed response.
11. Module 06 Synthesis — Understanding AI Risk — Do not ask whether AI is safe. Ask what the controls leave behind, and whether a named owner accepts it.

M07 · AI Governance

1. What Is AI Governance? — AI governance is the system that decides who may make which AI decisions, under what rules and controls, with what accountability.
2. Why Responsible AI Matters — Responsible AI earns informed trust, and trust drives adoption. Valuable AI that people do not trust, or trust blindly, will not scale.
3. AI Governance Principles — Seven stable principles guide AI decisions whatever the model. Control depth follows the risk, and every principle needs an owner and a control.
4. AI Policy vs AI Governance — Policy states what must be true about AI. Governance is the system that makes it true and keeps it true.
5. AI Roles, Ownership and Accountability — Accountability for AI is designed or it defaults. Each line gets a distinct job; each material system gets one named owner who can stop it.
6. The AI Governance Operating Model — Centralize what must be consistent, delegate what needs business context, and let risk decide how far each decision travels.
7. AI Governance Committee and AI Governance Office — The committee decides what is material, the office makes governance run, and the business owns the outcome. Both need written authority.
8. AI Inventory and Risk Classification — The inventory shows what AI exists. Classifying each use, legal tier first, decides how closely it is governed.
9. AI Lifecycle Governance — An approval is a snapshot of a moving system. Govern the AI you run, from discovery to retirement, with one named owner.
10. AI Evaluation and Approval Gates — Evaluation is evidence about the conditions tested. Approval is an owned decision to proceed, under conditions, with triggers that reopen it.
11. AI Policies, Standards, Monitoring and Audit — A policy says what must be true. Standards make it testable, controls make it happen, monitoring shows it, and independent audit proves it.
12. Module 07 Synthesis — Governing AI at Scale — At scale, governance fails at the seams between its parts, not inside them. Test the joins, and accept only evidence.

M08 · AI Economics

1. The Economics of AI — AI is metered. Judge it by what each unit of value costs, across all cost layers, and whether that holds at ten times the use.
2. Understanding AI Total Cost of Ownership — Total cost of ownership is build plus run plus operate plus change over the system's whole life. The vendor quote is one line.
3. Model, Compute and Infrastructure Costs — Price is the numerator. Divide by the useful share, success rate or utilization, to get the cost that matters.
4. Data, Integration and Operational Costs — The model is priced per request. Data, integration and operations are priced by sources, connections, change and exceptions, and usually cost more.
5. Experimentation vs Production Economics — An experiment buys information, priced by the decision it informs. Production buys reliable delivery, priced by full cost of ownership.
6. AI Unit Economics and Economics at Scale — Scale multiplies what each unit leaves behind. Know contribution per unit, break-even volume and the assumption that moves both.
7. Cost Optimization and AI FinOps — AI FinOps is a loop that makes AI spend visible, owned and improvable. Aim for value per unit of spend, not the smallest bill.
8. Build vs Buy Economics and Investment Decisions — Compare the decision, not the quote. One TCO boundary for both options, then delay, people, risk and exit, then the flip point.
9. Module 08 Synthesis — Making AI Economically Sustainable — Sustainable AI means each outcome pays, the whole life is paid at real volume, and written triggers keep both true.

M09 · Executive AI Roadmap

1. Assessing Enterprise AI Readiness — Do not ask whether you are AI-ready. Ask whether you are ready for a specific ambition, what it needs, and which gap holds everything back.
2. The AI Maturity Model — Maturity is not pilots or spend. It is how reliably you run and repeat governed, measured AI value, at the level your strategy requires.
3. Prioritizing the AI Portfolio — Prioritizing AI is capital allocation: choose the combination your people can finish, stage the money behind evidence, and stop what fails its test.
4. Quick Wins, Strategic Bets and Transformation Initiatives — Quick wins buy value now, strategic bets buy options, transformation buys structural change. Give each its own rules and decide the mix deliberately.
5. From AI Pilot to Production to Scale — Four proofs: a pilot shows it is feasible, validation that it is valuable, production that it is operable, scale that it is repeatable.
6. Building the 90-Day AI Plan — A 90-day AI plan is a contract: few owned outcomes, long poles started in week one, a not-yet list, and a production decision on Day 90.
7. Building the Multi-Year AI Roadmap — A multi-year AI roadmap sequences decisions under falling certainty: a quarterly contract, a one-year commitment, a direction beyond, held up by named assumptions.
8. AI Transformation Governance and Executive Sponsorship — Transformation governance settles cross-functional trade-offs on evidence and by a date. A sponsor needs mandate, money and time, not just a title.
9. AI Change Management, KPIs and Operating Rhythm — Transformation is done when work changes and results move. Change makes it possible, a scorecard shows it, a fixed rhythm decides.
10. Module 09 Synthesis — From Strategy to Execution — Run the AI roadmap as one loop: each part hands the next something usable, and evidence from execution changes the plan.

M10 · Future of AI

1. Where AI Is Going Next — AI already reasons, perceives and acts. Watch four dials - reliability, cost, autonomy, physical reach - and act when one crosses your written threshold.
2. The Evolution of AI Models — Top AI capability is crowded and changes hands in months. Match each workload to a model tier, test releases on your own cases, keep models swappable.
3. Multimodal and Agentic AI — Seeing and doing are becoming one AI system. The leadership choices are the door it uses, what it may watch, and where a person signs.
4. AI + Robotics — A wrong movement is not a wrong answer. Start from the physical workflow, keep safety outside the model, and judge cost per successful task.
5. Autonomous Workflows and AI-Native Organizations — AI-native means redesigning the flow of work as if AI were there from the start, not adding an assistant to every step.
6. AI-Native Products and Business Models — Sell the hole, not the drill. Price the unit that tracks your cost and the customer's result; defend with what rivals cannot rent.
7. AI, Regulation, Geopolitics and Global Competition — Rules say what you may do, market by market, and they move. Geopolitics says what you can get. Map both for every critical use.
8. Preparing for AI Uncertainty and Strategic Change — Prepare, do not predict. Test every AI move across several futures, start the no-regret moves now, and match decision speed to reversibility.
9. Module 10 Synthesis — The Future AI Leader — Prepare the organization, not the forecast. Watch thresholds, test on your own work, commit at the speed of undo, and review on a date.

Module cards

Frameworks to draw from memory

Terms

10-20-70 rule

BCG's description of how AI leaders spend resources: about 10% on algorithms, 20% on technology and data, 70% on people and processes.

90-day AI plan

A contract for one quarter: a few measurable outcomes, one owner each, the dependencies they need and three decision dates.

Abstention

The system declines, asks a clarifying question or escalates when evidence is missing, unreadable or unclear.

Agent loop

Plan, act, observe, decide - repeated until the goal is met or a stopping condition ends it.

Agent mandate

The written delegation for an agent - its outcome, systems, actions it may take alone, handoff rules and business owner.

Agentic AI

An AI system that pursues a goal in a loop - planning, calling tools and acting - without a new human instruction at each step.

AI adoption curve

The path from owning AI tools to redesigned work, as AI moves from licenses to people, into processes and into how the business competes.

AI advantage stack

Six layers from foundation models to customer experience; the higher the layer, the harder it is to copy.

AI agent

A system that pursues a goal by choosing and taking a sequence of actions through tools, adjusting to each result.

AI awareness

Knowing that AI matters to your industry and organization, without yet knowing what it can reliably do or what to change.

AI follower

An organization with the same AI that starts with tools, spreads effort across disconnected pilots and counts activity instead of outcomes.

AI governance

The system of decision rights, accountability, policies, controls and oversight that keeps AI use responsible and aligned with business objectives.

AI inventory

A living record of every AI use, with its owner, purpose, data, provider, users, actions, legal and internal tier, and controls.

AI leader

An organization whose habits - problem choice, ownership, redesign and proof - turn widely available AI into measured changes in how work is done.

AI leadership

Changed behavior: sponsoring AI work, asking for evidence, redesigning processes and owning the decisions AI touches.

AI literacy

Understanding what AI can and cannot reliably do well enough to reason about its use, risks and requirements.

AI portfolio

All AI experiments, products, automations and shared capabilities the enterprise funds, managed together against one budget and one pool of people.

AI risk

The potential business impact when an AI-enabled system produces, amplifies or acts on an incorrect, unsafe, unauthorized or inappropriate outcome.

AI system (OECD and EU)

A machine-based system that infers from its input how to generate outputs such as predictions, content, recommendations or decisions.

AI transformation

Changing how an organization works so that an AI capability produces repeatable business results, not just a successful demonstration.

AI translator

A person with enough domain knowledge and AI literacy to turn a business problem into sound AI work, or to say no.

AI value

The measurable business benefit AI causes by changing work, compared with what would have happened without it, net of cost.

AI washing

Claiming that a product or service uses AI, or uses it more capably, than it actually does.

AI-enabled professional

Someone who uses AI well, safely and critically in their own role without being a technical specialist.

Allocation rule

One written method for charging shared platform costs to AI systems, applied to every case so nothing is hidden or counted twice.

Anchor problem

One real business problem, not an AI project, that a leader carries through every module as a learning reference.

Artificial intelligence

The capability of a computer system to perform tasks that normally require human thinking, such as recognizing, predicting, generating or deciding within limits.

Attention

The Transformer mechanism that lets each token weigh the earlier tokens in its context, so context steers the prediction.

Augmentation

AI raises the speed or quality of a person's work while that person stays in control and owns the outcome.

Automate versus augment

Whether AI replaces a task or strengthens the person doing it. The same exposure can shrink hiring or raise performance.

Automation

A machine performs a repetitive task end to end, usually to cut cost, time or errors.

Benefit owner

The named business leader accountable for a benefit line; finance validates the money.

Benefit variance

The gap between planned and realized benefit, split by cause: adoption, benefit per use or conversion.

Blast radius

Everything a manipulated AI system could access, change, send or trigger before someone stops it.

Bottom-up adoption pressure

Employees adopt a technology before the organization approves it, so leaders must see and bound use, not only introduce it.

Bounded experiment

A visible trial with a few written rules, a measure and a date to keep, change or stop it.

Break-even volume

Fixed cost divided by contribution per unit: the volume at which total value covers total cost.

Brussels effect

Firms adopting EU rules worldwide because one global standard is cheaper than several (Anu Bradford, 2020).

Build

Developing a significant AI capability inside the organization, owning its behavior, data and roadmap.

Business owner

The leader with authority over the process who is accountable for the AI outcome, as distinct from the team that builds the AI.

Business productivity

Valuable output produced relative to the resources used: people's time, technology, capital and outside services.

Buy

Acquiring a mature capability from the market; it still needs decisions on data, security, integration and governance.

Calibration

How well a system's stated confidence matches how often it is actually right.

Capability

What the technology can do - generate, classify, predict, retrieve or optimize - independent of any business workflow.

Capability dials

The four things still changing fast: reliability, cost per task, autonomy and reach into the physical world.

Capability profile

What a particular system does reliably, task by task - strong at some tasks, weak or blind at others.

Capability radar

A list sorting each tracked capability into Adopt, Trial, Watch or Park, moved only by evidence.

Capability versus reliability

What a system can do on its best day versus how dependably it does it; fluent output can still be wrong.

Capability-to-value chain

Capability, adoption, behavior change, outcome, value. Each link is necessary; none is sufficient on its own.

Capacity creation

Time and attention AI frees. It becomes value only when management assigns it to output, quality, growth or lower cost.

Catastrophic forgetting

A neural network losing skills it had when it is trained on new material, which is why every retrained version needs testing.

Cheat sheet

One page per level that gathers every chapter's card: one idea, key points, terms, frameworks, red flags and interview questions.

Competitive advantage

Producing at lower cost than rivals, or delivering more perceived value, or a mix of the two (Rumelt).

Complements

The processes, skills, data and organizational changes that a general-purpose technology needs before it pays off.

Constraint layer

The capability layer whose weakness currently limits what the organization can achieve with AI.

Contribution per unit

Value of one unit minus the variable cost of delivering it, including expected review and failures; what remains to cover fixed cost.

Controlled retirement

Switching an AI system off deliberately - access removed, records archived, data handled, contracts ended, users told.

Convergence

Several technologies maturing at roughly the same time, so that each one makes the others useful.

Copilot

An AI assistant that drafts, suggests or summarizes while a person stays in control and takes the action.

Copy test

Asking what would still be hard to copy if a competitor got exactly your AI model tomorrow.

Cost of a first try

What a person must buy, install, learn or ask permission for before trying a technology; for generative AI it is close to zero.

Counterfactual

What would have happened without the AI initiative; estimated with a baseline, comparison team or staggered rollout.

Data poisoning

Planting tainted material in the data a model learns from so it misbehaves later, often long after the plant.

Decision rights

Clear answers to who may propose, build, approve, deploy, change and stop an AI system, and who is accountable for its outcome.

Deployer

Under the EU AI Act, an organization that uses an AI system under its own authority. It has duties separate from the provider's.

Differentiation test

Ask whether we would still have an advantage if rivals had this capability tomorrow. If yes, it is not our edge.

Digital Omnibus on AI

The 2026 EU regulation amending the AI Act; it moved high-risk dates to December 2027 and August 2028.

Diseconomies of scale

Cost per unit rising with volume, through harder cases, rising review shares, step costs or lower value per unit.

Embedded AI

AI built into software bought for another purpose - forecasting, ranking, routing or fraud scoring - often without the AI label.

Enabler

A shared capability, such as clean master data, that several initiatives need and that has little standalone return.

Enterprise AI readiness

The organization's ability to turn a specific AI ambition into repeatable, governed, economically sustainable work.

Enterprise AI strategy

A coordinated set of choices about where AI matters, what the organization builds and funds, and what it will not do.

Escalation of commitment

The tendency to invest more in a failing course of action one is personally responsible for.

Evidence question

A request that names a claim, a baseline, a threshold, a guardrail and the decision the result will settle.

Excessive agency

OWASP's name for a system with more functionality, permissions or autonomy than its job needs.

Executive AI literacy

Enough understanding of AI to judge its strengths and limits, ask the questions that expose value and risk, and own the decision.

Executive sponsor

A senior leader who owns an AI initiative's business outcome, can fund it into production and has the authority to stop it.

Exit strategy

A plan for leaving a vendor or partner - data portability, migration effort, alternative suppliers and contract terms.

Explain test

Rate your confidence, explain the topic in three plain steps, point to a decision where you used it, then re-rate.

Feedback loop

The time between doing the work and knowing whether it worked. A first win needs one measured in weeks or months, not years.

Fine-tuning

Further training a model on examples to change its behavior or specialization; not a way to keep facts current.

First win

A first AI project chosen to deliver real value and to teach the organization: hard enough to matter, realistic enough to finish.

Fixed workflow

Steps set in advance by code, with AI used inside single steps; cheaper and easier to test than an agent.

Fixed-cost absorption

Fixed cost spread over more units as volume grows; low adoption leaves the same cost on fewer units.

Flywheel

A loop where use creates feedback, feedback improves the offer and a better offer brings more use.

Foundation model

A general model trained once on broad data at scale and adapted to many tasks, such as drafting, summarizing and translating.

General-purpose technology

A technology that is pervasive, keeps improving and spawns complementary innovation, such as steam, electricity or computers.

Generative AI

General-purpose AI that drafts, summarizes and transforms content in response to requests in ordinary language.

Grounded answer

An answer built from retrieved evidence, with sources a person can check against each claim.

Groundedness

Whether an answer is supported by the source the system was supposed to use, such as current company policy.

Hallucination

Fluent model output that is false or unsupported, such as an invented fact, figure or reference.

Handoff

A case the agent stops on and passes to a named person, with the request, what it checked and why it stopped.

If-then plan

A commitment of the form 'when X happens, I will do Y', which makes follow-through more likely than a general intention.

Illusion of explanatory depth

The tendency to feel you understand how something works far better than you can actually explain it.

Imperfect imitability

Barney's term for resources rivals cannot easily copy, because of history, causal ambiguity or social complexity.

Indirect prompt injection

Hostile instructions planted in a document, email or web page that the AI later reads; the attacker never talks to it.

Inference

Using a trained model to turn an input into an output. The parameters stay unchanged.

Informal AI

AI tools employees bring themselves - personal accounts, extensions, departmental subscriptions - outside approval. Often called shadow AI.

Installation and deployment

Perez's two periods of a revolution: speculative build-out first, broad productive use later, often after a crash.

Isolating mechanism

Whatever stops rivals from closing an advantage, such as contracts, relationships, reputation, scale or tacit know-how.

Jagged frontier

The uneven boundary of AI capability: similar-looking tasks can fall inside it, where AI helps, or outside it, where AI hurts.

Language as the interface

Reaching AI by stating the outcome you want in ordinary words, instead of learning code, query languages or specialist screens.

Large language model

A large neural network trained on vast text and code to predict the next token from context; the base of most AI assistants.

Leader's loop

Watch thresholds, test on your own work, commit at the speed a decision can be undone, and review on a fixed date.

Least privilege

Running each program or agent with only the privileges its task requires, and nothing because it might be useful later.

Legal tier

The category the EU AI Act assigns to a use - prohibited, high-risk, transparency or minimal - regardless of internal scoring.

Lethal trifecta

Private data, untrusted content and external communication in one system; together they let an attacker steal data.

Lifecycle governance

The policies, decisions, controls, reviews and accountability applied to an AI system across its whole life, from discovery to retirement.

Long pole

A dependency on the critical path, often in another function, such as a security review or data agreement, that sets the earliest finish.

Material change

A change that alters an AI system's risk, impact, data, users, autonomy or decision consequences, and so requires reassessment.

Metered cost

Cost that moves with each request, page of context, answer and human check, unlike a license that sits still.

Minimum required readiness

Building the capabilities the current stage and the next stage need, rather than every foundation before starting.

Mixed diagnosis

An honest assessment that names where an organization leads and where it follows, function by function, with one strength and one gap.

Model serving

Running a trained model in production: accepting requests, scaling with demand and returning answers fast enough and cheaply enough.

Net AI value

Total benefit minus total cost of an AI capability, with both sides tested at expected scale.

Net present value

Future net cash flows discounted at the required rate of return, minus the investment.

Not-yet list

The written list of deferred work, each item with its reason and the quarter in which it will be reconsidered.

Official AI

AI tools the organization licensed as AI, with a contract, a named owner and a usage policy.

Over-trust

Approving AI output because it sounds fluent and confident, without checks matched to the cost of an error.

Oversight duty

A director's duty to make a good-faith effort to have reporting on mission-critical risks and to act on red flags.

Parity

An investment every competitor can make. It protects the business but does not set it apart.

Partner

Combining our domain knowledge and data with a specialist's skills to create a capability neither could easily make alone.

Pattern mix

The combination of create, understand, predict, decide and act that a problem needs, as opposed to the one a team already owns.

Payback

The time it takes to recover the investment; it ignores the time value of money and flows after payback.

Permission-aware retrieval

Search that applies the asking user's access rights before any content reaches the model.

Planning fallacy

The tendency to underestimate how long one's own work will take, even when asked for a worst-case estimate.

Post-training

Training after pretraining, using examples and human ratings, that teaches a model to follow instructions helpfully and safely.

Posture

A steady way of operating that keeps an organization able to benefit whichever way AI moves, instead of betting on one forecast.

Practice testing

Recalling material instead of rereading it. It improves retention after a week, even though rereading feels more reassuring.

Pre-mortem

Before committing, imagine the initiative has already failed and list why. Prospective hindsight surfaces more reasons than asking what might go wrong.

Preemption

Federal law overriding state law. In US AI policy it is being sought by executive action, which does not by itself repeal state laws.

Process carries the AI

The workflow itself depends on AI, so results no longer rely on individuals remembering to use a tool.

Productivity paradox

Wide adoption of a technology with little measured productivity gain, as Solow observed for computers in 1987.

Prompt injection

Untrusted text that steers an AI system to act against its owner's intent, typed directly or hidden in content it reads.

Proportionate governance

Controls that rise with the impact of an AI use, so low-risk experiments move fast and high-impact uses get stronger review.

Quality-adjusted productivity

Output change times quality change. Thirty percent more output at 20 percent lower quality is only about a 4 percent gain.

Readiness record

For each dimension, the current score, the evidence, the level the ambition requires, and the action that closes the gap.

Realized ROI

Measured benefit minus actual full cost, divided by actual full cost; known only after the money is spent.

Reasoning model

A language model trained, largely with reinforcement learning, to work through intermediate steps before answering; stronger on multi-step problems, slower and costlier.

Rebound effect

When a resource gets cheaper to use, total use can grow so much that total spending rises. Also called the Jevons paradox.

Recoverable error

A mistake that is seen and fixed before it counts, such as a draft a specialist checks before it is sent.

Relevant range

The band of activity within which a fixed or step cost stays flat. Forecasts are valid only inside it.

Repeat-run reliability

How often an agent succeeds every time the same task is run again, not just once.

Retrieval-augmented generation (RAG)

Retrieving relevant, permitted information and giving it to a model as context before it generates an answer.

Review date

The date on which a commitment is decided again: keep, change or stop. A planned stop is a result, not a failure.

Risk classification

Assigning each AI use a tier that decides who reviews it, who approves it and how closely it is monitored.

Risk multipliers

Reversibility, reach, detectability and speed - the four questions that make the same error trivial or serious.

Role track

A curated list of 10 to 14 Level 1 chapters for one role, such as CEO, CFO, COO, risk and legal, or CTO and CIO.

Role-specific AI literacy

Enough understanding to judge AI output and redesign the work in your own role. It does not mean learning to code.

Rules versus access

Rules govern what you may do in a market; access governs whether you can obtain chips, compute, models or markets at all.

Silent error

A task the agent completed wrongly without anyone noticing; found only later, or by sampling completed work.

Six components

Technology, data, people, process, governance and leadership - the parts that must hold together before AI creates business value.

Socio-technical

NIST's term for AI risk arising from technology together with how, where and by whom it is used.

Spaced practice

Reviewing material in short sessions spread over time, such as a two-minute card a week later, rather than in one sitting.

Spread and depth

Spread counts who has access or uses AI. Depth asks whether the work itself now depends on it.

State

The agent's running record of where the job stands, so it neither repeats nor skips steps.

Step cost

A cost that stays flat within a range of activity, then jumps to a new level, such as one more reviewer or capacity block.

Stop rule

The result, written before a test starts, that would make the team halt or redesign the use case.

Strategic refusal

An explicit decision about what the organization will not do with AI, so scarce talent, data and money go to the priorities.

Strategy kernel

Rumelt's three parts of a good strategy: a diagnosis of the critical obstacle, a guiding policy and coherent actions.

System of record

Where the business keeps the official version of something, such as the CRM, billing, ledger or ticketing system.

Task decomposition

Breaking a role into the tasks it actually contains, so AI's impact is judged task by task rather than by job title.

Task productivity

How much faster or better one task is done with AI. An input to business results, not proof of them.

Task-level suitability

Deciding for each task in a workflow whether a fixed rule, AI, a person or removal is the right answer.

Test-time compute

Extra computation a model spends while answering, such as reasoning step by step; it raises cost and latency per answer.

Three clocks

Technology (a capability appears), employee (people use it) and enterprise (the organization sees, bounds and scales it). Leaders close the last gap.

Three task states

Human-led (the person does it), AI-assisted (the person uses AI and stays accountable), AI-automated (AI does it; the person handles exceptions).

Threshold

The quality, cost, speed and risk line a capability must cross for one named workflow before you act.

Time horizon

METR's measure: the length of task, in human expert time, that a model completes at a given success rate.

Total cost of ownership

Everything it costs to build, run, operate and change an AI system over its whole life, including migration and retirement.

Trace-back test

Checking that an AI initiative links up through a priority and the AI strategy to the business strategy before it is funded as a bet.

Traditional AI

Task-specific systems that predict, classify or optimize, usually built into a business application rather than used directly.

Training

Repeatedly adjusting a model's parameters on example data until its outputs improve; this creates the model's capability.

Training at scale

Teaching a model on vast data with large clusters of specialized chips; at the frontier, its cost keeps rising.

Transformation

Redesigning the process, roles, controls and measures around what AI makes possible, not just adding a tool.

Unit of economics

The business unit a case is measured in - per case, document, customer or shipment - for both cost and value.

Use case

A specific application of an AI capability to improve a business activity, decision or workflow, with a user and a measurable outcome.

Use-by-market map

A table of each AI use against each market served, showing the duties and dates that apply in each.

Use-case chain

Role, current problem, AI capability, new workflow, measurable outcome. A proposal missing a link is not yet a use case.

Use-case design brief

One page naming the problem and measure, four roles, workflow map, task split, data and systems, and the smallest test with a stop rule.

Value hypothesis

A testable claim: if this capability, for this workflow, then this outcome improves by this much, while a constraint holds.

Weakest link

The dimension whose gap limits the whole ambition, however strong the others are; after Kremer's O-ring theory.

Web-scale data

The public internet's text, code and images used as training material; it does not include your company's private information.

Workflow integration

AI wired into the systems, hand-offs and sign-offs where work happens, rather than used in a separate chat tab.

Interview questions

Answer out loud first, then flip the card.

Red flags to avoid

• Treating AI as a tool purchase and calling the rollout a transformation while roles, measures and controls stay the same.
• Claiming leadership because the company has licenses, a platform or many pilots - access and activity instead of behavior and outcomes.
• Calling generative AI the whole of AI, or handing it to everyone because it is easy, with no rule for checking what it produces.
• Saying the pilot failed because the model was wrong, or that the AI team owns the transformation.
• Treating 'AI-powered' as an answer, or arguing about whether a system is 'really AI' instead of asking what it reliably does.
• Saying the model is already trained so using it is nearly free, or that it learns from every conversation.
• Asking the model for facts, figures or references and trusting the fluent answer, or making it the system of record.
• Proposing to train the model on every company document so it knows the business, with permissions to be sorted out later.
• Calling a chatbot an agent, or judging an agent on its accuracy at one step instead of the whole job.
• Reporting logins, prompts or pilot accuracy as proof that an AI initiative created value.
• Treating hours saved as cash saved, or announcing that a 20 percent productivity gain means 20 percent fewer staff.
• Treating the approved ROI as achieved, or deciding whether to continue by looking at the money already spent.
• Presenting a count of AI projects, a vendor roadmap or a slogan such as "AI-first" as the AI strategy.
• One word for the whole AI effort - "AI is strategic, so we build it all" or "a vendor sells it, so we buy it all".
• Presenting an AI tool every rival can license as a strategic advantage, or treating being first as if it were being ahead.
• Starting from a tool or a department list, then reporting the number of AI pilots as progress.
• Funding an AI proposal from its demo before anyone has mapped the workflow or written a baseline and a stop rule.
• Scaling the agent with the highest completion rate without asking what its misses are or what they cost.
• Approving or rejecting an AI system on its accuracy score alone, without asking what happens when it is wrong.
• Approving a system on one overall accuracy figure without asking who wrote the test and how it scores on each slice of real work.
• Saying the model provider handles security, or that a better prompt filter will solve prompt injection.
• Giving an agent its user's full access, or a high rung on day one, because the demo worked.
• Answering "we have an AI policy" or "Legal handles it" when asked who decides what AI is acceptable.
• Classifying the model or vendor instead of the use, or saying a prohibited practice can be approved with stronger controls.
• Calling a model swap, new data, new users or new permissions "implementation details" and leaving the original approval in place.
• Approving an AI case on the model price and its falling trend, with no unit, no review rate and no ten-times test.
• Presenting the vendor quote as the annual cost of the AI system, with no model-change budget and no named step costs.
• Approving a scale-up because cost per request fell and usage grew, with no contribution per unit, no break-even volume and no scenario range.
• Announcing a readiness percentage without being able to say ready for what, or what evidence supports it.
• Funding every sponsored idea at once, ranked by projected ROI, with no capacity limit and no written stop rules.
• "Our first 90 days will launch two use cases, test three bets, stand up governance and build the platform."
• Presenting reasoning, multimodal AI or agents as future steps, or setting strategy by predicted dates instead of written thresholds.
• "We'll wait until the AI rules settle" or "the federal order means state laws no longer apply".
• Answering "what is our AI future?" with a prediction about which model or vendor will win.
• Treating the program as reading to finish, or as a tool tutorial, instead of practice on real decisions.
• Saying AI is the technology team's job, or approving an AI proposal because the demo was impressive.
• Pointing to your own daily AI use, or to the number of tools deployed, as proof that you are leading on AI.
• Rating yourself by how confident you feel, or starting with your strongest area because progress there feels quick.
• Treating market excitement or spending as evidence of value, or assuming the gains will follow adoption automatically.
• Saying AI arrived with one product, that we must build our own model, or that we should wait for the next model before learning.
• Reporting licenses, pilots and active users as proof of transformation, or giving the whole enterprise one score.
• Naming the vendor, the model or "our data" as the advantage without saying what a competitor could not copy.
• Equating a fast license rollout with transformation, or proposing to ban AI until the policy is ready.
• Saying the AI journey starts when we pick a platform, as if nothing were already running or the vendor owned the outcome.
• Calling a renamed chatbot an agent, or limiting an agent by instruction instead of by enforced permission.
• Answering "How many jobs will AI replace?" with a global number, or promising that no job will change.
• Answering the talent question with a hiring number, as if one central AI team were the whole answer.
• Starting wherever is easiest or most exciting, announcing it as a transformation, or giving every department its own pilot.