What is
Artificial Intelligence?
Artificial intelligence is a machine's capability to perform human-like tasks — solving problems, understanding language, recognizing images and sounds. It performs these tasks by processing large amounts of data and applying specific analysis algorithms. The difference from the human brain lies precisely in scale: AI makes connections people would never make, because no human being can process the amount of information these systems can.
The field of study that builds machines able to perform tasks that typically require human intelligence is known as AI. Five families of tasks concentrate that effort:
Pattern recognition
Spotting regularities in massive data that would otherwise go unnoticed.
Learning
Improving performance with experience, without explicit reprogramming.
Problem-solving
Finding solutions through rules, search and inference.
Decision-making
Choosing courses of action based on evidence and forecasts.
Language understanding
Reading, analyzing and producing spoken or written natural language.
Key idea for the lesson: AI is not a single technology — it is a field that gathers many techniques. What unites them is the goal: reproducing, in machines, human-comparable performance on specific cognitive tasks.
The types of AI
The e-book distinguishes four major approaches. They do not compete with one another: they coexist in the same ecosystem and, often, in the same product.
AI has the potential to radically alter many sectors — healthcare, banking, education and transportation — but its development raises ethical concerns, such as the risk of job loss and the call for transparent, trustworthy AI.
The history of AI
on a timeline
The idea of intelligent machines goes back to Greek mythology, but AI as a discipline was born in the mid-twentieth century — and advances in cycles of euphoria and winter.
The first chess software
The first chess-playing programs mark a significant point in the history of AI. In 1956, at the Dartmouth College symposium, John McCarthy uses the term "artificial intelligence" for the first time.
Founding milestoneMachines with human-like cognition
Scientists begin working to build machines with cognitive abilities similar to humans.
The "AI winter"
After accelerated growth, funding and interest in AI research dry up — the period became known as the AI winter.
Disillusionment cycleThe Internet and mass data
The advent of the Internet and the availability of enormous volumes of data enable significant breakthroughs in machine learning and revive the discipline.
Deep learning and autonomous vehicles
Researchers make major strides with an emphasis on deep learning and the creation of autonomous vehicles. AI becomes a rapidly developing field with applications in medicine, finance, education and transportation.
Two hype cycles
The first peak (≈2012) delivers less than the marketing promised; the second (≈2016) has IBM Watson Health as its symbol — billions invested and minimal results, dismantled in 2021. Lesson: impressive demos do not equal deployable solutions.
Anti-hype filterThe generative turn
Generative AI crosses a threshold: models become good enough for non-technical leaders to use them directly, with no engineering team. The cost of running large language models drops by more than 90% between 2022 and 2024, compressing a five-year adoption curve into eighteen months.
Third cycle — real capabilityTechniques and algorithms
An AI algorithm is a set of rules or instructions for completing a task or solving a problem. Choosing the right algorithm requires considering two things: the task and the available data.
Decision trees
Represent decisions and their outcomes in tree form to predict future results from inputs.
Artificial neural networks
Inspired by the structure and behavior of the human brain: interconnected "neurons" learn to recognize patterns and decide.
Support vector machines (SVM)
Seek the hyperplane in high-dimensional space that creates the largest possible gap between categories.
Deep learning
Neural networks with multiple layers process vast volumes of data to improve learning and decision-making.
K-means
An unsupervised approach that groups data points into clusters of high similarity.
Linear regression
Models the linear association between a dependent variable and one or more explanatory factors.
Selection criterion: these are only a few among many algorithms in use. The task and the available data determine the specific algorithm — not the fashion of the moment.
Machine Learning:
how machines learn
The algorithms in this type of AI learn from experience and get better over time without explicit programming — they discover new information and draw conclusions without constant human intervention.
Learning
Supervised
The correct output is known for each input; the model generates predictions and is scored on performance.
Image classificationSpam detectionCredit riskUnsupervised
The model is trained on unlabeled data and finds structure on its own.
Customer segmentationAnomaly detectionSemi-supervised
Combines labeled and unlabeled data — useful when only part of a huge available volume is labeled.
Speech recognitionSentiment analysisReinforcement
Training by trial and error, with rewards and penalties for the model's behavior.
Game playingRobotics and controlActive
Starts with a limited initial dataset and lets the model itself choose which new data to train on.
Medical diagnosisNamed entity recognitionThree ethical alerts from the e-book: (1) bias in training data produces inaccurate results and negative consequences; (2) protecting personal data requires safe, private, GDPR-compliant collection; (3) algorithms can be abused — such as profiling customers and bombarding them with intrusive ads. It falls to businesses to monitor the ethical, responsible application of their algorithms.
The major fields of AI
Six subfields structure the e-book. Browse the tabs — each one summarizes the concept and the typical tasks, just as presented in the material.
Natural Language Processing
The branch of AI that uses algorithms to understand and generate natural-sounding language — from chatbots and virtual assistants to translation.
Part-of-speech tagging
Labeling each word by its function (noun, verb, adjective). It disambiguates meanings: "bank" can be an institution or an action — the grammatical class helps clarify the intended sense.
Named entity recognition
Identifying and classifying people, organizations and locations in text — extracting structured information from unstructured text for retrieval and question-answering systems.
Sentiment analysis
Determining whether a text is positive, negative or neutral. Used in customer service, advertising and social media tracking to understand consumers.
Language translation
Converting text from one language to another. The field spans automatic, human and hybrid translation (machine + human translators).
Text summarization
Producing summaries that capture the essential arguments. Two types: extractive (isolates key sentences from the original) and abstractive (paraphrases, possibly creating new sentences).
Where it shows up
Customer service, marketing, healthcare, language teaching, data mining — with ethical alerts on transparency and bias.
Computer Vision
The subfield that analyzes and understands visual media, allowing machines to detect and interpret the visual world as humans do.
Image processing
Adjusting and analyzing images to extract information: enhancement (brightness, contrast, sharpening), restoration (noise removal) and compression (lossy or lossless).
Feature extraction
Locating edges, corners, textures and shapes that make an image unique and task-relevant — the basis for classification and recognition.
Object recognition
Identifying and categorizing items in images or videos, learning from large collections of tagged photos. Applications: image search, robotics and augmented reality.
Facial recognition
Extracting and assessing distinguishing facial features, trained on massive labeled datasets. Used in security, advertising and social media.
Scene understanding
Deducing the context in which the image was captured — location, time of day, weather. Relevant for robotics, augmented reality and autonomous vehicles.
Where it shows up
Medical imaging, satellite photography and industrial machine vision are among the countless application areas.
Robotics
The study of how robots work and of how they can be built and used effectively — carrying out activities automatically, by instruction or by learning.
Industrial
Welding, painting and packaging on production lines; designed for nonstop precision and efficiency, classified by degrees of freedom (from 1–2 up to 6 or more).
Service
Aid or interact with humans: therapy robots (social and emotional support), telepresence, education and entertainment.
Consumer
Vacuum cleaners, lawn mowers, glass cleaners and toy robots — smaller and cheaper, for domestic use.
Military
Surveillance, bomb disposal and reconnaissance; unmanned ground vehicles (UGVs) that are wheeled, legged, aquatic or airborne.
Research
Labs and universities: manipulation, mobile exploration robots, medical robots and social robots for rehabilitation and education.
Autonomy
From fully autonomous to semi-autonomous and teleoperated — with sensors and algorithms that let them see and react to their surroundings in strikingly human ways.
Expert Systems
Programs that simulate the judgment and decision-making ability of human experts in narrow domains.
Knowledge base
Domain facts, heuristics and rules of thumb, fed and updated by human specialists.
Inference engine
Algorithms and rules of logic that consult the base to process data, draw conclusions and recommend actions.
Decision + justification
The system explains the reasoning behind its choices — a valuable feature for checking the soundness of its conclusions.
Specialization
They solve problems in a narrow domain; they do not reason outside their area.
Rule-based
They apply predefined rules and heuristics developed by human experts.
Human-machine interaction
They depend on human input to work well — relevant data and critiques of their decisions. Used in medicine, finance and engineering.
Neural Networks
A technique inspired by the brain's structure and function: many interconnected "neurons" that adjust connection weights and biases to gain accuracy over time.
Feedforward
The most elementary form: layers of neurons in which data flows in a single direction, from input to output.
Convolutional (CNN)
Optimized for grid-like data such as images; widely used in object recognition and image labeling.
Recurrent (RNN)
Process sequential data (time series, text), "remembering" information from earlier steps.
Generative adversarial (GAN)
Two networks in contest: one generates fictitious data, the other tries to identify whether it is real.
Autoencoders
Learn compressed representations of a dataset and reconstruct it — useful for dimensionality reduction.
Self-organizing maps (SOM)
Classify data into clusters of comparable elements; common in data visualization and outlier detection.
Deep Learning
A machine learning subfield with multilayered neural networks (hence "deep"), responsible for state-of-the-art results in image and audio recognition, NLP and autonomous vehicles.
Learns complex patterns
Understands extremely intricate patterns and correlations in data — state of the art in image and speech.
No manual feature engineering
Learns "features" directly from raw data — visual, textual and audio.
Scales with massive data
Analyzes and learns from gigantic datasets; improves the more data it receives.
Tolerates noise and gaps
Handles imperfect, noisy or incomplete data.
Versatile across tasks
Success in image, speech, NLP and autonomous driving.
Computational cost
Requires lots of labeled data and high-performance resources (GPUs) — training is difficult and resource-intensive.
Architectures cited in the e-book: CNN, RNN, LSTM, GAN, autoencoders, Boltzmann machines, deep belief networks (DBN) and bidirectional RNNs.
AI × Machine Learning × Deep Learning
Related yet distinct fields. Not every AI algorithm is grounded in ML or DL: AI is the vast field that incorporates the other two.
A machine's ability to execute tasks that would ordinarily need human-like intelligence: interpreting language, recognizing patterns, deciding. It includes rule-based systems, decision trees and neural networks.
Teaches computers to "learn" from data without explicit programming; efficiency improves with exposure to more data (supervised and unsupervised).
A subset of ML with multilayered neural networks; it demands lots of labeled data and GPUs, and excels at recognizing the intricacies in data.
Evaluating and improving models
An AI model's performance is not chance: it depends on identifiable factors and improves with specific techniques.
What affects performance
| Factor | Why it matters |
|---|---|
| Data quality | Noisy or incorrect data degrades the model. |
| Architecture | The structure must suit the task. |
| Hyperparameters | Optimal values can be decisive for efficiency. |
| Training data size | More data helps — up to a threshold of diminishing returns. |
| Consistent labels | Inaccurate labels deteriorate learning. |
| Overfitting | A model over-adapted to training data fails on new data. |
How to evaluate and improve
| Technique | What it does |
|---|---|
| Metrics | Accuracy, precision and recall (classification); MSE and MAE (regression). |
| Cross-validation | Splits data into train/test "folds", reducing overfitting. |
| Hyperparameter tuning | Grid or random search for the best values. |
| Regularization | A penalty on the loss function (weight decay/L2, dropout). |
| Data augmentation | Generates new samples through scaling, rotation and noise. |
| Ensembles and transfer learning | Combines models; or starts from a pre-trained model. |
Frameworks don't expire.
Tools do.
The principle that structures the second e-book: the logic for deciding whether and how to invest in AI does not change — what changes constantly are platforms, vendors and prices. From here on, the focus is on decision frameworks, not fashionable tools.
The 4 AI capabilities
every leader must master
Most leaders treat AI as a single technology — something you either "have" or "don't have". In practice, it is a family of four distinct capabilities, with different applications, risks and readiness requirements. Confusing them is how companies buy the wrong tool for the wrong problem.
Predictive AI
Examines historical data and finds patterns: which customer is likely to churn, which machine will fail, which loan tends to default. Many companies already use it without calling it that.
Generative AI
Creates text, images, code and audio. Immediate, real value — as is the risk of hallucination, bias and overconfidence in outputs.
Agentic AI
Executes sequences of actions autonomously, across multiple steps, without a human at each stage. Highest potential; least mature; highest risk.
Decision-support AI
Synthesizes data, surfaces options and recommends — keeping the human in the decision seat. For most businesses, today's highest-return entry point.
Know which capability you are deploying. The business case, governance requirements and change management challenge are completely different for each one.
Where AI creates real value
The Value Chain Map identifies five domains of measurable value. Most companies focus on one or two — and miss the rest.
Operations
The fastest, most quantifiable ROI: predictive maintenance, supply chain optimization, demand forecasting, automated quality control.
Customer experience
Gets the most attention and is frequently overhyped. Personalization and support automation work — only with clean data and deep integration.
Decision-making
A force multiplier for leadership: real-time dashboards, scenario modeling, anomaly detection — evidence instead of gut feeling.
Product development
Underused: accelerates research, synthesizes feedback, generates and tests prototypes, compresses development cycles.
Risk management
The quietest and often most valuable domain: fraud, compliance and contract analysis, at a scale no human team can match.
The budget paradox: most transformation investment flows toward customer experience — but the highest ROI usually lives in operations and risk management.
The 3-filter test for any use case
AI delivers the most value on processes repeated hundreds or thousands of times. A task done twice a year is rarely worth automating.
No data, no AI — regardless of what the vendor promises.
A mislabeled product image is recoverable. A missed fraud signal or a wrong medical recommendation is not.
Use cases that pass all three filters advance; those failing any one are deprioritized until the condition is fixed. The test takes fifteen minutes and saves months of wasted effort.
Start where failure is recoverable. The most successful first pilots live inside the business (invoice processing, contract review, internal knowledge search, demand forecasting), not at the customer boundary — when they fail, the impact remains internal. Customer service is the most marketed, most implemented and most disappointing category in enterprise AI.
Is your company ready?
The 5-dimension scorecard
Before asking "what should we build?", ask "are we ready to build anything at all?". Score each dimension from 1 to 5 — and require written evidence for any score above 3.
Score 1: disconnected silos, no single source of truth. Score 5: automated, standardized, trusted pipelines. Most companies land between 2 and 3 — and overestimate themselves by at least one point.
Cloud, APIs and integration without six-month projects. Legacy monolithic architectures score low.
Openness to experimentation and comfort with uncertainty. A culture that punishes failure kills the pilot before the first result.
It doesn't require data scientists: it requires digital fluency to use tools, interpret outputs critically and write good prompts.
Genuine support shows up in strategic priorities, budget and behavior — not in press releases.
The bottleneck principle: a readiness score is not an average — it is a chain, and chains break at the weakest link. A 4-4-4-4-2 organization is not a 3.6 organization: it is a 2 organization with expensive infrastructure sitting on a broken foundation.
Readiness threshold: a minimum of 3 across all five dimensions. Below that on any single dimension, the investment yields more by closing the gap than by building the system.
Order for fixing gaps: 1st data ("a six-month data quality project is not a delay to the AI initiative — it is the AI initiative"), 2nd leadership alignment, 3rd culture, 4th talent (upskill your own people before hiring) and 5th infrastructure, the easiest to close with budget.
Recency bias
A recent upgrade or a successful small experiment inflates the score across all dimensions. One clean dataset does not make an organization data-mature.
Advocacy bias
The person filling out the assessment is usually the one who wants the initiative to proceed. Optimism is not a measurement methodology: assess as a cross-functional group.
Comparison bias
Comparing yourself to an admired competitor does not answer the question. The scorecard measures absolute readiness, not relative position.
Building the AI strategy
The first strategic question: will AI be a productivity multiplier on top of the existing operation, or a transformation of how the business works? These are not points on the same spectrum — they are different strategies, with different budgets, talent, governance and timelines. The mistake is not picking one: it is not choosing.
⚡ Productivity multiplier
Doing what you already do, faster and cheaper: automating invoices, accelerating contract review, cutting CRM time. Same business model, same engine running at lower cost. Measurable ROI within months, low organizational change.
🚀 Business transformation
Changing how value is created and delivered: from product to service, previously impossible platforms, data as new revenue. It takes years, demands deep change, carries real risk — and the possibility of a durable competitive advantage.
Build × Buy × Partner
Once the direction is set, choose the execution path — as a function of software development maturity and uniqueness of need.
Buy packaged solutions or build components
High maturity + common need: leverage the market; build only specific pieces.
Buy components or build internally
High maturity + unique need: if the capability is your differentiator, build and own it.
Buy packaged solutions
Low maturity + common need: the right path for the majority of business applications.
Build with a partner
Low maturity + unique need: a domain specialist who knows the regulation, edge cases and failure modes.
Four filters: company size (smaller ones almost always buy), technical capability (if you can't maintain it, don't build it), budget (building always costs more than the initial estimate) and strategic priority (if it's core to your differentiation, own it; if not, buy it).
A single thesis — and a portfolio
Cost reduction
Operational efficiency; success measured in dollars saved and hours reclaimed. The easiest to justify to a CFO.
Revenue acceleration
Finding, converting and retaining customers better; requires clean CRM and marketing-sales integration.
Customer experience
Faster, more personal, more relevant interactions; requires the most mature data infrastructure.
Competitive moat
A capability that is hard to replicate (proprietary data, network effects); highest risk and reward, measured in years.
Mixing theses is the most common strategic mistake. Choose one; align the portfolio to it; revisit annually. Structure the portfolio as 1 transformational bet (protected from quarterly pressure) + 3 to 5 operational pilots that deliver value within 90 days and build the data and governance muscle the big bet will eventually require.
The business case
that survives the CFO
Most AI ROI calculations are wrong — not because the gains are fabricated, but because the cost side is systematically incomplete. A complete case has six components; most include two.
1 · 36-month TCO
Total cost of ownership over three years, with all six components — not the first-year software cost.
2 · Conservative break-even
60% adoption at 90 days as the base case; show what happens at 40%.
3 · The cost of doing nothing
Errors, manual hours and missed opportunities of the status quo — make it expensive on paper, because it already is in practice.
Payback reality: vendor projections are marketing material. The realistic timeline is 12 to 18 months from go-live — add 3 to 6 months of implementation and the span from approval to break-even reaches 15–24 months. Killing the pilot at month six because "it hasn't paid off yet" means paying the full cost and collecting none of the benefit.
When to abandon the business case: the required data doesn't exist and can't be created within the timeline; break-even exceeds 36 months under conservative assumptions; leadership alignment is conditional on results that only appear after the investment; or a simpler, cheaper alternative solves the same problem with lower risk. In those cases, propose a 90-day proof of concept instead of a 24-month program that demands faith.
Source: Business Explained (2026), ch. 5The pilot playbook
and the 90-day plan
Most AI pilots don't fail at the finish line — they fail at the starting line, through scoping errors, success criteria errors and team composition errors. Five questions separate a real pilot from corporate theatre:
Name the process, the people affected and the current cost in time, money or error rate. If it doesn't fit in two sentences, there is no scope.
Exact, accessible, sufficiently clean sources. Access problems discovered mid-pilot are program killers.
"Better" is not a criterion. "Reducing invoice processing from 4 days to 1" is. Measurable, attributable and observable within the window.
The question that is almost never asked. Pre-commit to the kill switch: performance failure (no signal by day 45), data failure or adoption failure.
Name the person — and involve them from the start, not only at the results presentation.
Why write the kill switch first? Without pre-agreed stop conditions, the decision to end is made under sunk cost, stakeholder pressure and career risk — conditions in which the rational individual decision is always to continue, even when the right call is to stop. The Klarna case (below) shows the price of skipping this step.
The 90-day action plan
Ninety days is the right window: long enough to gather meaningful data, short enough to keep attention and political protection, and aligned with the quarterly cycle. If the work doesn't fit in 90 days, reduce the scope — don't stretch the timeline.
- Run the readiness scorecard (any dimension < 3 requires a gap-closure plan in parallel)
- Map the value chain across the 5 domains
- Apply the 3-filter test to the candidates
- Choose one pilot — not three, not a shortlist
- Answer the 5 scoping questions in writing
- Define success and failure before execution
- Name a business owner, a technical lead and an end user
- Pilot live by day 60 — with real users and real data
- Continuous measurement from day 61; a weak signal by day 75 triggers the kill switch conversation
- Pre-registered criteria — no mid-course changes
- Comparison against the baseline, not against zero
- Day 90: performance, adoption, unexpected findings and a recommendation — proceed, pivot or stop
Real win
Meets the criteria, adoption at the projected rate, no governance issues — plan the scale-up with the same rigor as the pilot.
Partial win
Meets some criteria or works in a subset — a scoped pivot and a deliberate decision about whether to scale.
Learning to be killed
Misses the criteria or reveals a wrong assumption — stop, document and apply the learning. A killed pilot is not a failure: it is the system working.
The human layer:
people before technology
The technology works; the people don't adopt it. This is the most common outcome of enterprise AI rollouts — and it is almost never discussed honestly. Organizations spend 90% of the budget on the technical layer and 10% on the human layer, then are surprised when the system sits unused six months after go-live.
Enthusiasts
Adopt immediately and become informal champions. Valuable — but a minority, and they would adopt anyway.
Skeptics
Resist openly. Managed well, they become quality control; managed poorly, organized resistance.
The silent middle
Decides everything. They watch whether the system makes work easier or harder and whether leadership is serious. The first signal they receive is the one they act on — and the decision is almost irreversible. Target this group first.
The honest conversation about jobs
🔄 Roles that change
The majority. The work doesn't vanish — the skill profile shifts: the analyst who spent 60% of their time pulling data now interprets it and acts on it. This requires upskilling, not restructuring.
🫥 Roles that disappear
Real, and they must be named explicitly: high-volume, low-judgment repetitive tasks. Pretending otherwise destroys credibility — and the workforce never forgets.
🌱 Roles that emerge
Prompt engineers, AI output reviewers, workflow designers, governance leads. Identify and fill them from within before going to market.
Permission not to automate: when judgment, relationship or expert craft is the actual product, AI degrades the outcome — complex negotiations, client crisis management, unprecedented problems and domains of expertise built over years. Knowing where AI doesn't belong is as important as knowing where it does.
Change management when the technology changes weekly: train principles, not clicks; build internal communities of practice; normalize experimentation with psychological safety; and rotate internal champions across teams — peer credibility transfers faster than any formal training.
From scale to governance
A pilot succeeds, leadership approves the next phase... and nothing happens. This is the pilot-to-production gap, where most AI value disappears. Four root causes:
Infrastructure mismatch
What worked for 50 users breaks at 5,000; the infrastructure required for scale was never part of the pilot's design.
Ownership vacuum
The project team disbands after the demonstration; production requires a permanent owner accountable for performance and continuous improvement.
Integration debt
The pilot ran in isolation; production demands connections to the ERP, CRM and data warehouse — complexity deliberately excluded at the start.
Budget structure mismatch
Pilots are CapEx with an end date; production is ongoing OpEx — a conversation many organizations never have.
The 4 dimensions of governance
🔐 Data privacy
What the AI may access, how it is stored and what happens when the data is no longer needed.
⚖️ Bias mitigation
Systems inherit the biases of their training data; testing before deployment is, in several jurisdictions, a legal requirement.
📜 Regulatory awareness
GDPR, the EU AI Act and US state-level frameworks — knowing which rules apply, where, and how they are changing.
📋 AI policy
Translates external requirements into specific, realistic, maintained internal rules — neither vague nor overly restrictive.
The hallucination problem is the most underestimated operational risk in enterprise AI: the wrong output arrives fluent and confident — a hallucinated legal citation looks exactly like a real one. Install a validation discipline before go-live, classifying outputs by consequence level: low consequence, light review; high consequence, mandatory expert review.
Agents with oversight: agents are the right pattern when the task is sequential, well defined and with recoverable errors (onboarding, procurement, internal tickets). Wrong when genuine judgment in novel situations is needed or the error is irreversible. Three controls: human checkpoints at consequential decisions, audit trails for every action and technically enforced scope boundaries — not merely documented ones.
Who owns AI? Ownership at two levels: the platform (infrastructure, security, data, vendors) sits centrally, with the CTO/CDO; the application sits with the business unit that derives the value. Centralize infrastructure and standards; decentralize development and business decisions. And think in compounding capabilities — skills, infrastructure and processes each initiative leaves for the next — rather than isolated projects.
Source: Business Explained (2026), chs. 8 and 9Real cases:
what worked and what didn't
Nine case studies from the e-book, each tied to a framework. Click to expand — note that every case also records what did not work.
Worked Computer vision and ML in the seed pipeline compressed a cycle that used to require years of field trials, with productivity gains cited in the 2022–2023 annual reports.
Didn't work The hard part was trust: experienced breeders ignored recommendations that contradicted their field intuition — often rightly. A human validation layer had to be created giving scientists final say, and it took longer to design than the algorithm itself.
Lesson: your highest-ROI use case is probably not a chatbot — it is a high-frequency, data-rich operational or R&D process.
Worked Generative AI accelerated creative production (variations, campaign localization) and predictive AI improved demand forecasting, with AI cited as a margin contributor in 2023.
Didn't work Scale created governance complexity: dozens of markets, inconsistent standards and regional teams building their own workflows outside the central program.
Lesson: a portfolio only works if the center holds — establish standards, oversight and escalation paths before expanding across functions.
Worked Predictive readmission models allowed earlier intervention with high-risk patients, with ROI documented through avoided penalties and shorter stays.
Didn't work Physicians who received risk scores without action guidance simply ignored them: valid predictions, zero behavioral change. The redesign embedded the scores into the clinical workflow with recommended actions per alert level.
Lesson: a system that produces correct outputs but changes no behavior has delivered zero value. Design the workflow change before building the model.
Worked In 2024, the AI assistant was doing the work of 700 agents, with millions of conversations and falling cost per interaction.
Didn't work The pilot won on the primary metric (cost) and failed on the ones never defined: quality consistency, edge cases and satisfaction in complex situations. By 2025 the company was quietly rehiring humans.
Lesson: what was missing was a customer-satisfaction floor, defined in advance, below which the system would be scaled back regardless of efficiency gains. Write the kill switch first.
Worked The immediate cost reduction was real: AI-generated content lowered unit costs and increased volume.
Didn't work Freelance creators' contracts ended with minimal notice triggered swift public backlash; users questioned quality and cultural nuance. The message received was not "AI helps us do more" — it was "AI replaces people without warning".
Lesson: the efficiency calculation must include reputational cost, morale cost and the signal sent to those who remain.
Worked Genuine capability on narrow, well-defined tasks with quality structured data. The technology wasn't fraudulent — the ambition was.
Didn't work Real clinical data is messy and contextual; recommendations conflicted with established guidelines. MD Anderson terminated the contract after ≈US$62 million with no clinical deployment; IBM sold the division at a loss in 2021, after billions invested.
Lesson: when a system asks experts to trust its outputs over their own judgment, the burden of proof is on the system — not on the experts.
Worked ChatGPT Enterprise for 3,000+ employees with custom GPTs per function (research, regulatory, commercial): one platform, consistent governance, experimentation without central IT for each new use case.
Didn't work Quality varied across functions; the support needed by less technical teams was underestimated, leaving a long tail of GPTs built but never integrated into daily work.
Lesson: giving everyone the ability to build tools is not the same as the capability to build good ones. Budget enablement at the same level as the platform.
Worked Anti-fraud AI significantly reduced fraud in high-risk markets with little credit history — where rule-based systems consistently failed; logistics optimization improved delivery reliability across Latin America's complex last mile.
Didn't work Building in-house capability in markets with a limited AI talent pipeline created hiring and retention challenges, competing against dollar-denominated salaries.
Lesson: AI works in resource-constrained environments — but the binding constraint is not budget; it is talent. Audit the pipeline before deciding to build.
Worked An industrial components maker with ≈150 employees deployed demand forecasting with off-the-shelf tools integrated into its ERP: under US$40k in cost and payback within eight months from inventory reduction alone.
Didn't work The production scheduling module underperformed because data on real constraints (cycle times, operator variability, maintenance) was not recorded in the ERP — the model received incomplete data and produced unusable recommendations.
Lesson: off-the-shelf tools deliver enterprise-grade results at SMB scale — but only on data that actually exists. Audit the required fields before signing the contract.
AI for small and
mid-sized businesses
Most AI advice assumes resources SMBs don't have. The reframe: for an SMB, AI is not a transformation program — it is a productivity tool. The question is not "how do we build an AI strategy?", but "which decisions do we repeat weekly, which processes consume disproportionate time and which customer interactions could be faster?"
Routine inquiries
Pricing, availability, scheduling: AI absorbs the volume; humans handle the exceptions.
Recurring content
Weekly emails, posts and proposals: AI writes the first draft; a human edits and approves.
Metrics and reporting
Weekly numbers, pipeline, inventory: AI aggregates and formats, freeing the owner to decide.
Decision synthesis
Competitive research, feedback patterns, market signals: AI summarizes; the owner decides.
A working capability under US$500/month: identify the 2–3 highest-frequency manual tasks; pick one general-purpose tool and learn it deeply before adding others; document the prompts that work (individual discovery becomes organizational capability); and set a review protocol for every output that reaches a customer or informs a significant decision. Typical access cost: US$20–200/month.
General-purpose assistant
Writing, research, summarization and analysis.
AI-enhanced email & calendar
Drafts responses and reduces scheduling friction.
Content production
Integrated into the channels where the business publishes.
Customer-facing layer
For the highest-volume repetitive inquiry type.
Documents & knowledge
AI search: institutional knowledge findable without depending on whoever holds it.
The SMB's competitive advantage: the company evaluates a tool on Monday, pilots it on Wednesday and decides on Friday — while enterprise adoption takes six to twelve months across security, legal and IT. Speed of experimentation, willingness to iterate and short approval cycles compound over time. These five categories cover 80% of the AI value available to most SMBs.
Ethics, transparency
and the limits of AI
The two e-books converge: the more pervasive AI becomes, the more the questions stop being technical and become questions of responsibility.
⚖️ Responsibility and bias
As systems decide on their own, who answers for the consequences? Systems trained on biased data produce unfair, discriminatory decisions in employment, credit and beyond.
🔍 The transparency paradox
We need to understand how AI decides, yet many algorithms are too complex for human comprehension. Paths: natural-language explanations (NLP), explainable AI (XAI), interpretable machine learning (trees, linear models, rules) and regulation requiring justified decisions.
🔐 Privacy
AI systems rely on vast volumes of personal data; collection and use demand security, consent and compliance (GDPR and, on the regulatory front, the EU AI Act and US state frameworks).
🤔 Frontier debates
Can machines be creative? Can they be conscious? Rights for robots? The e-book presents both sides: advocates point to works indistinguishable from human ones; skeptics note that AI recombines what it learned without leaving the parameters of its own code.
Risks of advanced AI: malicious use (surveillance, manipulation of public opinion, autonomous weapons, cyberattacks), job displacement with potential inequality and social tension, bias amplification and the speculative "singularity" scenario. The material's answer is not panic but cautious development, responsible research, proper rules and oversight.
Who controls the technology? Governments argue for regulation to protect citizens; the private sector argues for freedom to innovate, with tech giants at the development frontier. The likely outcome combines state regulation and market forces — governments, organizations and individuals collaborating so AI serves the public interest.
Suggested lesson plan
A teaching sequence to guide the class through this material in two sessions:
- Opening: what AI is and the five families of tasks (§01)
- Types of AI and the timeline — discuss the hype cycles (§02–03)
- Techniques, algorithms and the five types of machine learning (§04–05)
- Applied fields in the tabs: NLP, vision, robotics, expert systems (§06)
- Conceptual close: the AI × ML × DL diagram and model evaluation (§07–08)
- The 4 capabilities and the value map — exercise: classify the class's cases (§09–10)
- Readiness scorecard: each group scores a company they know (§11)
- Strategy, business case and the 90-day pilot with kill switch (§12–14)
- People, scale and governance (§15–16)
- Seminar: each group presents a real case and its lesson (§17)
- Final debate: ethics, transparency and regulation (§19)
Guiding question for the whole journey: "does this capability work on messy, real-world data without specialists nearby?" — the filter that separates real capability from hype, from the 1950s to today.
References
All content on this page was produced from the two attached e-books, licensed to Diocélio Goulart. References in APA style (7th ed.):
Business Explained. (2023). Artificial intelligence explained. Business Explained. https://www.business-explained.com
Business Explained. (2026). Artificial intelligence in business explained: A business leader's guide to AI-powered growth. Business Explained. https://www.business-explained.com
Methodological note: the case studies (Bayer, Coca-Cola, Cleveland Clinic, Klarna, Duolingo, IBM Watson Health, Moderna, Mercado Libre) and the figures cited (e.g., a drop of more than 90% in LLM costs between 2022 and 2024) reproduce what the source e-books state; for formal academic use, verification in the primary sources of each case is recommended. The Andrew Ng epigraph appears as the opening quotation of the 2023 e-book.