Start free trial
Trending New Genres Authors
EnglishEnglish
EspañolSpanish
简体中文Chinese
繁體中文Chinese (Traditional)
FrançaisFrench
DeutschGerman
日本語Japanese
PortuguêsPortuguese
ItalianoItalian
한국어Korean
РусскийRussian
NederlandsDutch
العربيةArabic
PolskiPolish
हिन्दीHindi
Tiếng ViệtVietnamese
SvenskaSwedish
ΕλληνικάGreek
TürkçeTurkish
ไทยThai
ČeštinaCzech
RomânăRomanian
MagyarHungarian
УкраїнськаUkrainian
IndonesiaIndonesian
DanskDanish
SuomiFinnish
БългарскиBulgarian
עבריתHebrew
NorskNorwegian
HrvatskiCroatian
CatalàCatalan
SlovenčinaSlovak
LietuviųLithuanian
SlovenščinaSlovenian
СрпскиSerbian
EestiEstonian
LatviešuLatvian
فارسیPersian
മലയാളംMalayalam
தமிழ்Tamil
اردوUrdu
Searching...
SoBrief
Applied Artificial Intelligence

Applied Artificial Intelligence

A Handbook for Business Leaders
Choose AI projects that solve real bottlenecks, survive deployment, and earn trust.
by Mariya Yao 2018 227 pages
3.32
93 ratings
Listen
Immersive
V2.3
Amazon Kindle Audible
Try Full Access for 3 Days
Unlock listening & more!
Continue

Key Takeaways

Treat today's AI as a specialist, not a synthetic executive

Narrow skill, not general mind. The book distinguishes the narrow systems available in 2018 from artificial general intelligence, or AGI: a hypothetical machine able to learn from limited experience and transfer knowledge across different fields. A program that beats champions at Go does not thereby know how to drive or write a song. Its expertise does not automatically travel.

The authors' Machine Intelligence Continuum is a map of capabilities, from fixed-rule automation and prediction to learning, generating content, and interpreting human signals. These levels overlap, but the map helps leaders ask what a system actually does instead of accepting the label AI. Most business systems solve bounded tasks. Even impressive deep-learning tools need lots of examples, specialist knowledge, and ongoing maintenance.

Analysis

The distinction is useful because the word intelligence invites people to infer flexibility that a system may not possess. A reliable way to assess a product is to ask what happens when its inputs, users, or operating conditions differ from the examples it was built around. This is similar to testing a specialist employee outside their area of expertise. The book's capability ladder is a practical antidote to hype, though real systems often combine rule-based software, statistical prediction, and learned components. A single label can obscure that mixture.

Judge AI by the human problem it unlocks, not its technical novelty

Useful beyond corporate margins. In India, FarmGuide used satellite images to estimate individual farms' crop yields. Better estimates could help lenders and insurers assess risk more fairly, giving vulnerable farmers an alternative to predatory loans. In Liberia, UNICEF's U-Report let young people report community problems by SMS, a practical choice where smartphones and reliable internet were not guaranteed. A poll of 13,000 users found that 86 percent reported teachers exchanging grades for sex, helping bring public attention and a government response.

Medical applications offer another measure of value. The book describes a diagnostic system that helped reduce benign breast surgeries by 30 percent. In each case, the important innovation was not simply an advanced model. It was matching data and technology to a problem people urgently needed solved.

Analysis

These cases make a strong point about innovation: the most consequential design choice may be access, not algorithm. A text-based service can be more useful than a sophisticated app if its intended users have limited connectivity. The examples also suggest a test for proposals: identify who gains access, safety, or bargaining power, not only what the organization saves. At the same time, a model's prediction does not itself repair predatory lending or abuse. Technology can reveal a problem and help coordinate a response, but institutions still have to act.

Treat every dataset as a designed proxy, not a neutral mirror

Evidence has a history. Ground truth means observable, verifiable information used as a benchmark for judging a model. But data is produced through human choices: what counts as a customer, which events get recorded, who is included, and how errors are corrected. If the benchmark is wrong, a model can appear accurate while repeating a mistaken account of reality.

The book lists failures at every stage: unclear collection goals, inconsistent definitions, faulty measurement, processing mistakes, missing populations, and unrepresentative samples. One court risk tool made a wrong recommendation after jail time was entered incorrectly. Digital behavior also has limits as evidence: a click reveals an action, not the person's motive. Before modeling, leaders need to trace how records were created and ask whose experience is absent.

Analysis

This is a foundational warning for anyone using data to make decisions about people. A measurement is not the thing measured: recorded attendance, for example, may reflect access to transport as much as commitment. That distinction connects AI practice to measurement theory, where researchers examine whether an instrument captures the concept it claims to measure. The book's emphasis on data provenance is especially valuable because model testing cannot rescue a defective target. Yet complete neutrality is unattainable. The practical aim is to make assumptions visible, test their consequences, and correct them when affected groups show the measure is failing.

Fix organizational foundations before buying an AI promise

Infrastructure precedes intelligence. A company with disconnected databases, incompatible software, and no dependable way to share information may not be ready for enterprise AI. Central data systems and consistent internal interfaces reduce duplicated work and make records easier to access securely. Without them, a new model can add another isolated tool to an already tangled system.

Culture matters just as much. The book calls a powerful executive's preference for personal judgment over evidence a HiPPO, meaning the highest-paid person's opinion. Such leaders can block useful analysis or redirect projects without testing the evidence. A credible executive champion needs authority, budget, technical understanding, and willingness to experiment. A cross-functional AI SWAT Team, typically five to fifteen people, brings business units and front-line perspectives into project selection and oversight.

Analysis

This insight treats AI adoption as organizational change rather than a software purchase. Central infrastructure creates coordination, while cross-functional teams help surface knowledge that is invisible in a database, such as why staff use an informal workaround. The HiPPO label is memorable, but disagreement with an executive is not automatically irrational; data can also be incomplete or misinterpreted. The healthier standard is to make decisions contestable: state assumptions, compare them with evidence, and let domain experts challenge both. That practice makes analytical culture more than a slogan.

Choose a measurable business bottleneck before choosing an algorithm

Start with the decision. The book's AI Strategy Framework asks leaders to examine a project's strategic fit, potential size, required investment, possible return, risk, timeline, and stakeholder support. A task is a stronger candidate when it happens often, consumes time, has a measurable outcome, and draws on relevant data. If people cannot agree on what a correct decision looks like, a model may simply automate disagreement.

Pick a useful success measure, not a vanity number. Facebook tracked whether new members connected with ten friends in a week; Slack looked at teams exchanging 2,000 messages. These measures attempted to capture meaningful engagement, not mere registration. A practical business case should also count integration, management, and maintenance costs, then compare likely outcomes with a realistic baseline.

Analysis

The sequence protects organizations from a common category error: treating a technology as a goal rather than a means. It also makes metrics a design decision. A measure shapes what teams optimize, so a narrowly chosen target can improve one number while damaging the user experience or shifting costs elsewhere. This is a version of Goodhart's law, the risk that a measure loses value when it becomes a target. A sound metric therefore needs a counterbalance, such as pairing speed with error rates or revenue with customer retention.

Run small trials, then budget for the model's whole life

A model is a process. A typical project defines a business goal, examines and prepares data, trains a model, tests it on separate data, deploys it, monitors performance, and repeats the cycle. Keeping test data separate matters because a model can memorize examples it has already seen without performing well on new ones. Even a good model can lose accuracy when customer behavior, products, or regulations change.

The book recommends short, iterative development cycles rather than one large bet. PwC teams described four-week experiments that let clients test an idea in a limited way before expanding investment. The work does not end at launch: data pipelines, retraining, security, and monitoring all require resources. Google researchers warned that shortcuts around this surrounding infrastructure create technical debt, meaning costly future rework and maintenance.

Analysis

The lifecycle view is a useful corrective to demonstrations that showcase a model but leave production realities offstage. It resembles the difference between a laboratory result and a dependable service: the latter needs monitoring, ownership, and recovery plans. Separate testing also guards against overfitting, where a system performs impressively on familiar examples but poorly in the world outside them. Small pilots reduce exposure, but they should test operational conditions as well as predictive performance. A successful prototype that cannot be integrated or maintained is evidence about the prototype, not proof of business value.

Buy, assemble, or build according to strategic control and capability

Talent is scarce and costly. The book cites a 2017 estimate of fewer than 10,000 people worldwide qualified for state-of-the-art AI research and engineering. Most organizations cannot hire every specialist needed, so third-party products and cloud services may be sensible ways to test an opportunity. But outsourcing can create dependence on a vendor, expose valuable data, or leave a company unable to shape a capability central to its business.

Compare the full cost of building with the full cost of buying, including integration, security, support, and future updates. Check whether a vendor has relevant data and industry experience, whether its measures align with yours, and whether it can explain limitations. A practical route can move from buying a focused tool, to combining it with internal systems, to building custom components if evidence justifies the investment.

Analysis

This is not a simple argument for technological independence. It is a choice about where an organization needs control and where it can benefit from shared infrastructure. The same logic appears in make-or-buy decisions across manufacturing: firms retain capabilities that distinguish them, while purchasing standardized inputs when suppliers can do them better. The danger is treating a vendor's AI label as proof of learning or fit. A transparent evaluation should ask what the system does, how it fails, who can access the data, and what happens if the provider changes its product or disappears.

Automate repetitive steps so expertise can move to harder work

Apply AI where patterns repeat. Across business functions, the clearest opportunities are tasks that are frequent, structured, and tedious. Finance teams can use software to flag expense anomalies; legal teams can use language tools to find clauses for lawyer review; HR systems can help match candidates to roles; and sales models can prioritize promising leads. Recommendation systems personalize product choices, while customer-service bots handle routine questions or help human agents find answers.

The book describes Deloitte reducing contract-review work from four or five months to about a week with language classification. It also notes that recruiting administration can consume much of a recruiter's day. These systems work best as part of a human workflow, where people handle ambiguous cases, verify important outputs, and bring context the model does not possess.

Analysis

The strongest examples are not machines replacing an entire profession, but machines changing the distribution of attention within it. This resembles decision support in medicine: automated pattern recognition can help a specialist focus on suspicious cases, while clinical judgment remains essential for context and accountability. That division is not guaranteed by the technology itself. If staff are overloaded or discouraged from questioning automated recommendations, assistance can become automation bias, the tendency to defer to a system even when it is wrong. Workflow design must make review meaningful, not ceremonial.

Build fairness, transparency, and security into the system's design

Errors have unequal costs. The book describes online systems that showed fewer high-paying job advertisements to women and a delivery service that excluded some predominantly Black neighborhoods. Removing race or gender from a dataset does not eliminate discrimination if other variables act as proxies. Homogeneous teams can miss harms that are obvious to people affected by them.

The risks include deliberate misuse as well as accidental damage. The cited threat analysis warns that AI could make cyberattacks, deception, surveillance, or targeted physical attacks easier to scale. Ethical design therefore needs diverse perspectives, clear responsibility for decisions, understandable operations, and ways to detect and correct harm. In high-stakes settings, leaders should examine who bears false positives and false negatives, not just report a single accuracy score.

Analysis

Fairness is not a final polish applied after a model is built. Choices about labels, training populations, thresholds, and appeal processes already distribute benefits and burdens. This is why procedural safeguards matter alongside technical tests: affected people need a route to challenge an outcome, and an accountable organization must be able to investigate it. The book's call for diversity strengthens the design process, although representation alone cannot guarantee justice. Teams also need authority, relevant evidence, and incentives to act when evaluation reveals unequal effects.

Retrain workers before automation turns productivity gains into a social bill

Tasks change before jobs vanish. The book cites research estimating that 45 percent of work tasks could be automated, while only 5 percent of jobs had been fully displaced by automation. The distinction matters: software may take over parts of a role while leaving judgment, communication, and exception handling to people. Accenture initially expected automation to eliminate 17,000 roles in some operations, but employees moved into higher-value advisory work and staffing grew.

That transition does not happen automatically. The authors report a striking gap: 74 percent of surveyed executives expected to automate, but only 3 percent planned to increase worker training. They argue for ongoing education, practical retraining, and regular conversations with employees about changing workflows. AI strategy should count workforce capability and disruption among its outcomes, not treat them as external costs.

Analysis

This section turns workforce policy into an operating requirement. Retraining can preserve institutional knowledge while helping employees move toward work that depends on contextual judgment and human relationships. But the Accenture example should not be mistaken for a universal forecast: whether new roles appear, who qualifies for them, and who captures productivity gains depend on business choices and labor institutions. Responsible leaders should track transitions by worker group, not just total headcount. A rising productivity measure can conceal concentrated losses among people with the least power to adapt.

Analysis

The book's central contribution is not a prediction about superintelligence. It is a management discipline for bringing imperfect machine systems into ordinary organizations without confusing novelty with value. Its argument proceeds from capability literacy to organizational readiness, problem selection, data quality, iterative deployment, and ethical responsibility. That sequence matters: it makes AI adoption a chain of dependencies, where a model cannot compensate for a bad objective, an unrepresentative dataset, or a workflow that nobody trusts.

The strongest conceptual move is treating data and organizational design as part of the technology itself. A prediction system reflects choices about what to measure, whose behavior counts, and which errors matter. In that sense, AI is sociotechnical: its results emerge from software interacting with institutions, incentives, and human routines. The book's cases show why technical performance and social usefulness are not equivalent. An accurate classifier can still reinforce exclusion if the target reflects unequal access or if no one is accountable for its use.

The practical frameworks also have a productive tension. They encourage executives to define measurable goals, yet metrics can narrow attention and invite gaming. That tension is not a flaw so much as a governance problem. Teams need a primary measure for decision-making, plus safeguards that reveal costs displaced onto customers, employees, or excluded populations. Similarly, pilot projects reduce financial risk but can overstate readiness if they avoid messy integration and real-world variation.

Because the book dates to 2018, some product examples and numerical claims should be read as period-specific illustrations, not current benchmarks. Its more durable advice is procedural: verify what the system actually does, establish a credible baseline, test on data it has not seen, monitor drift, and give affected people a voice. The book is best read as a guide to responsible organizational learning, not a catalog of timeless AI products.

Summary in 30 Seconds
Business AI works best on frequent, structured tasks, with people checking ambiguous or high-stakes results. Start with a measurable problem and a realistic baseline, then test a model on data it has not seen. Four-week trials can limit early risk, but deployment brings ongoing costs for integration, monitoring, security, and retraining. An estimated 45 percent of tasks could be automated, while only 5 percent of jobs were fully displaced; yet 74 percent of executives expected automation and only 3 percent planned more worker training. Data choices and error costs shape who benefits and who is harmed.
Contains spoilers

Last updated:

Report Issue
Want to read the full book?

Glossary

Artificial General Intelligence (AGI)

Transferable, general machine intelligence

Artificial General Intelligence refers to a hypothetical machine with human-level or greater intelligence that can learn from limited experience, abstract concepts, and transfer knowledge across different domains. The book contrasts AGI with today's narrow systems, which can perform impressive specialized tasks but do not automatically apply their capabilities to unrelated problems.

Machine Intelligence Continuum

Map of machine capabilities

The Machine Intelligence Continuum is the book's framework for describing machine systems by the complexity of their capabilities. It ranges from fixed-rule automation and data-based prediction through learning, content creation, and emotional interpretation, toward hypothetical abstraction and self-evolution. Its levels are not rigid categories, and the framework helps business leaders distinguish what a system can actually do from broad marketing claims.

Ground truth

Verified benchmark evidence

Ground truth is observable, verifiable information treated as an objective reference for assessing a machine-learning system. It is used to determine whether predictions are correct. In practice, it can be difficult to establish because records may be incomplete, subjective, inferred, or collected from a biased sample. A flawed benchmark can make a model seem successful while producing misleading or harmful results.

HiPPO

Highest-paid person's opinion

HiPPO is the book's shorthand for a decision-making culture in which the highest-ranking or highest-paid person's preference overrides evidence and analysis. A HiPPO may promote a favored strategy or reject a data-backed proposal based on authority or intuition. The term describes an organizational obstacle to analytical decision-making, not a claim that senior leaders' experience is always mistaken.

AI SWAT Team

Cross-functional AI project group

An AI SWAT Team is a small, multidisciplinary group that supports an organization's AI champion. It brings together people from different departments, job levels, and areas of expertise to identify, prioritize, execute, and communicate automation opportunities. The book recommends including front-line and nontechnical perspectives so that teams can identify workflow needs, implementation risks, and effects that technical specialists might overlook.

True North metric

Primary measure of meaningful progress

A True North metric is a clear, actionable measure used to keep a major project focused on a meaningful business outcome. It should be understandable across the organization and serve as a credible proxy for success, rather than a vanity statistic. The book's examples include measuring whether new users become meaningfully active, rather than counting sign-ups alone.

Technical debt

Cost of future technical rework

Technical debt is the future cost created by shortcuts or design decisions that make software harder to maintain, extend, or repair. In machine-learning systems, debt can arise not only from code but also from unreliable or outdated data, complex models, and fragile pipelines. As models change or operating conditions shift, neglected maintenance can compound into expensive failures and rework.

Machine Learning as a Service (MLaaS)

Hosted machine-learning capabilities

Machine Learning as a Service refers to platforms or vendor offerings that provide tools or infrastructure for developing, training, deploying, or operating machine-learning systems. Organizations can use MLaaS to access capabilities without building every component themselves. The book also describes internal MLaaS platforms that standardize workflows, store experiment records, and help teams move models into production.

Download PDF

To save this Applied Artificial Intelligence summary for later, download the free PDF. You can print it out, or read offline at your convenience.
Download PDF
File size: 0.29 MB     Pages: 15

Download EPUB

To read this Applied Artificial Intelligence summary on your e-reader device or app, download the free EPUB. The .epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.
Download EPUB
File size: 1.36 MB     Pages: 20
Want to read the full book?
Follow
Listen2 mins
Home
Swipe
Library
Get App
Try Full Access for 3 Days
Listen, bookmark, and more
Compare Features Free Pro
📖 Read Summaries
Read unlimited summaries. Free users get 3 per month
🎧 Listen to Summaries
Listen to unlimited summaries in 40 languages
—
❤️ Unlimited Bookmarks
Free users are limited to 4
—
📜 Unlimited History
Free users are limited to 4
—
📥 Unlimited Downloads
Free users are limited to 1
—
Risk-Free Timeline
Today: Get Instant Access
Listen to full summaries of 26,000+ books. That's 12,000+ hours of audio!
Day 2: Trial Reminder
We'll send you a notification that your trial is ending soon.
Day 3: Your subscription begins
You'll be charged on Nov 24,
cancel anytime before.
Consume 2.8× More Books
2.8× more books Listening Reading
Our users love us
600,000+ readers
Trustpilot Rating
TrustPilot
4.6 Excellent
This site is a total game-changer. I've been flying through book summaries like never before. Highly, highly recommend.
— Dave G
Worth my money and time, and really well made. I've never seen this quality of summaries on other websites. Very helpful!
— Em
Highly recommended!! Fantastic service. Perfect for those that want a little more than a teaser but not all the intricate details of a full audio book.
— Greg M
Save 62%
Yearly
$119.88 $44.99/year/yr
$3.75/mo
Monthly
$9.99/mo
Start a 3-Day Free Trial
3 days free, then $44.99/year. Cancel anytime.
Unlock a world of fiction & nonfiction books
26,000+ books for the price of 2 books
Read any book in 10 minutes
Discover new books like Tinder
Request any book if it's not summarized
Read more books than anyone you know
#1 app for book lovers
Lifelike & immersive summaries
30-day money-back guarantee
Download summaries in EPUBs or PDFs
Cancel anytime in a few clicks

Why one summary?

Scanner
Find a barcode to scan

⭐
⭐
⭐
We have a special gift for you
Open
38% off
DISCOUNT FOR YOU
⭐
⭐
⭐
$79.99
$49.99/year
Only $4.16 per month
Continue
2 taps to start, super easy to cancel
Settings
General
Widget
Loading...