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
Real World AI

Real World AI

A Practical Guide for Responsible Machine Learning
Turn machine learning experiments into useful products without losing sight of who bears the risks.
by Alyssa Simpson Rochwerger 2021 145 pages
4.01
88 ratings
Listen
Immersive
V2.3
Amazon Kindle Audible
Try Full Access for 3 Days
Unlock listening & more!
Continue

Key Takeaways

Audit what your model learns, not just its accuracy score

Models inherit human choices. Machine learning finds patterns in examples people collect and label, then applies those patterns to new inputs. Training teaches the model; inference is its use on fresh data. The system’s behavior depends on the examples, categories, and goals people chose, so it is never simply neutral. Most business applications use supervised learning, where training examples include the expected answer.

Accuracy is not enough. IBM’s image-labeling system had been tuned for a stronger F1 score, a measure of classification performance, yet tagged a wheelchair user as “loser.” The team had not inspected all the associations in its enormous dataset. Finding and removing offensive tags required a manual review and a full retraining. Omitting sensitive attributes does not guarantee fairness either, because other inputs can act as proxies.

Analysis

A model can perform well against a test set while failing the people who encounter it. This is a measurement problem and an accountability problem: teams need to ask who defines acceptable errors, who bears their cost, and whether outcomes differ across groups. The lesson also complicates the popular idea that removing sensitive data makes a system fair. Without those attributes, a team may lose the ability to detect unequal effects. Auditing requires both technical tests and informed human judgment, especially when the model labels people or influences access to important services.

Measure the value customers and the business actually receive

A proxy can mislead. At eBay, a search model was tuned to increase purchases per visit. That measure improved, but revenue fell because the results often put inexpensive accessories, such as phone cases, above the phones shoppers wanted. The team called this “accessory pollution.” It had optimized a behavior that looked like success without checking whether it created the value the business needed.

Change the target. The team added price-related signals and shifted its measure to gross merchandise value per visit. The broader lesson is to connect model performance to the real outcome, not a convenient substitute. A medical model that always predicts “no cancer” could appear 99 percent accurate when the condition affects fewer than 1 percent of people, while detecting no cases at all.

Analysis

This is a practical example of Goodhart’s law: when a measure becomes a target, people and systems can optimize it in ways that distort the underlying goal. The same danger appears in education, customer support, and public services when easily counted activity stands in for quality. A sound metric is not merely a number that correlates with success; it should track the consequences stakeholders actually value. Even then, one metric rarely captures every trade-off, so teams should pair outcome measures with checks for harms such as false negatives, poor user experience, or unequal treatment.

Start with a real business need, not an AI-shaped publicity stunt

Technology needs a job. A fashion brand’s “cognitive dress” used AI to change its lights based on the mood of online posts. It drew attention, but the AI did little to improve the business. By contrast, a fabric company used visual search to find designs in a vast archive of old drawings. Designers could offer clients relevant existing patterns instead of starting each request from scratch.

Define the purpose first. AI is a family of technologies, not a business objective. A model may classify, predict, or find patterns, but the organization must specify what practical decision its output supports. Before funding the work, identify the user’s problem, the intended business result, and why a machine-learning approach is more suitable than a simpler solution.

Analysis

The contrast is a useful antidote to technology theater. A compelling demonstration can generate attention without establishing that the system solves a recurring problem or earns its upkeep. This distinction resembles the difference between invention and innovation: novelty matters only when it becomes useful in context. The fabric archive example is especially strong because AI unlocks dormant organizational knowledge rather than merely automating a visible task. Still, the best decision may be not to use AI. A simpler search tool or redesigned workflow can be cheaper, easier to explain, and more dependable.

Choose a narrow task with data, value, and room to expand

Find the manageable middle. The authors call a promising first use case a “Goldilocks problem”: a task that is valuable, narrow enough to solve, and supported by suitable data. Autodesk began by identifying password-reset requests, not by automating every customer-service question. Historical tickets made examples available, and the request was relatively clear-cut. Automation reduced the average resolution time from about a day and a half to 10 to 15 minutes. The company later expanded the system to 60 use cases, cutting average resolution time to 5.4 minutes.

Make the first win legible. Favor repeatable tasks where people generally agree on the answer, the current performance can be measured, and an early mistake is tolerable. A good first project demonstrates business value and creates a base for expansion without requiring a perfect model from day one.

Analysis

A narrow initial scope limits technical uncertainty and makes organizational learning cheaper. It also improves the quality of feedback: a team can tell whether a password request was routed correctly more easily than whether a broad, subjective category was assigned well. This resembles staged experimentation in product design, where a small release tests assumptions before a major investment. The caution is that easy wins can become a permanent comfort zone. Leaders should use early success to build skills and infrastructure, while ensuring that the first use case is genuinely connected to a meaningful business need.

Treat data quality and continuity as part of the product

Data does the heavy lifting. A self-driving system must recognize more than the obvious pedestrian: it may need to interpret a wheelchair user, a reflection, or a person pictured on a sign. If annotators receive unclear instructions, they label such cases inconsistently, and the model inherits that confusion. Data must be relevant, representative, accurately labeled, secure, and available for as long as the system is used.

Build a repeatable pipeline. Walmart’s shelf-scanning project needed a sequence of human labels, from dividing aisle images into shelving units to transcribing product labels. New store layouts and redesigned packaging then exposed gaps in the original examples. Data provenance, meaning where data came from and how it was changed, matters too. Without it, teams can mistake a partial or transformed dataset for the full reality they intend to model.

Analysis

The emphasis on data shifts attention from algorithmic glamour to the infrastructure of knowledge: collection, documentation, labeling rules, and maintenance. This is similar to measurement in the sciences. An instrument cannot produce reliable findings if observations are inconsistent or the sample excludes important conditions. Yet completeness is an aspiration, not a realistic guarantee; rare events and changing contexts make exhaustive coverage impossible. Strong practice therefore combines careful data design with explicit limits, human escalation paths, and monitoring that reveals when real-world inputs fall outside the model’s experience.

Build AI with the people who own its consequences

Models are only one component. At Figure Eight, product, engineering, and machine-learning groups operated in silos. Their work did not connect reliably, and researchers sometimes built models that addressed no valuable business need. The company organized cross-functional teams around business problems and aligned part of their incentives with shared company results. With largely the same staff and resources, annual product enhancements rose from three or four to 33.

Keep expertise connected. A working team may need a product lead, data scientists, machine-learning engineers, designers, infrastructure specialists, and people who understand the data and its users. The structure can vary: a central AI team can serve a smaller organization, while a matrix or embedded model may suit a larger, more mature one. The essential point is clear ownership and frequent collaboration.

Analysis

This organizational lesson reflects a basic principle of systems engineering: performance depends on interfaces between components, not just the quality of each component in isolation. A technically excellent model can fail at handoff, integration, or adoption. Shared goals help counter local incentives that reward model accuracy, uptime, or shipping speed even when those measures conflict. Cross-functional work also makes ethical review more concrete because affected stakeholders can challenge assumptions before they become code. The structure does have costs, including coordination and role ambiguity, so leaders need clear decision rights rather than collaboration for its own sake.

Use pilots to prove impact and rehearse production

A pilot is a controlled test. It should have a defined goal, limited scope and duration, measurable criteria, and little risk to core operations. Only about 20 percent of AI pilots make it into production, according to the figures cited in the book. One car-sales company spent millions trying to detect dents and rust in varied photographs, but models reached only 60 to 70 percent accuracy and the business value remained unclear.

Test the path to scale. OmniEarth helped water districts identify properties likely to be overusing water by analyzing satellite images. It started in one county, used imagery that could also support broader deployment, and connected its findings to billing data so employees could contact the right households. A pilot should reveal whether the whole workflow works, not merely whether a model can make a prediction.

Analysis

A pilot is most informative when it tests the riskiest assumptions, including data access, workflow integration, operating cost, and user response. A model demonstration can answer only a narrow technical question; it cannot establish that the organization can act on the result. This is closely related to experimental design: success criteria and comparison points need to be set before results arrive, or teams can reinterpret ambiguous outcomes as wins. A carefully bounded pilot also creates room to stop. Abandoning a weak idea early is evidence of disciplined learning, not necessarily failure.

Expect production to surprise you, then monitor and adapt

Real users change the conditions. IBM’s image-tagging demo was designed for a small audience, but a Reddit post sent thousands of visitors to it each minute. The demo failed under the unexpected load and briefly brought down a data center, although the underlying API remained functional. Production planning must account for availability, response speed, and the ability to handle more demand than a pilot encounters.

Keep a recovery plan. Models can also face unfamiliar inputs, new legal constraints, changing data, or deliberate attacks. The book describes spammers probing filters and malicious users manipulating systems that learn from feedback. Teams should monitor performance, limit harmful access, and know how to pause or roll back a model. A successful launch is not the end of the work; it begins ongoing operation.

Analysis

The demo story shows how a seemingly minor interface can become a system-level risk through dependencies that were invisible during development. Safety engineering addresses this with layered defenses: testing, monitoring, access controls, and rehearsed recovery procedures. Machine-learning systems need the same operational discipline as other critical software, with the added difficulty that their behavior can change as data changes. No test suite can anticipate every user or attack. The practical goal is not perfect foresight but fast detection, limited blast radius, and accountable decisions about when human review should override automation.

Make AI a company capability, with governance beyond the data team

Scale requires shared habits. Organizations that lead with AI help every department learn to spot suitable problems, manage usable data, collaborate across functions, and fund worthwhile work. Governance provides rules for data quality, security, ethical use, and choosing among models that optimize competing goals. It should include business priorities, not just technical approval.

Keep humans in the loop. YouTube used machine learning to flag problematic videos while human reviewers removed content and generated feedback for the systems. The authors report that this combination made moderation five times faster and more accurate than human review alone. Because standards and content change, the book recommends regular performance checks and model updates, with frequency suited to the use case. An organization should track not only how many models it deploys, but whether they are used, monitored, and maintained.

Analysis

Organizational maturity is more than the number of models in service. It is the capacity to assign responsibility, question outcomes, and revise systems without treating every correction as a crisis. Governance is most effective when it is integrated into ordinary decisions about data access and product goals, rather than isolated in a compliance office. Human review can catch edge cases and create useful feedback, but it is not an unlimited resource and can expose reviewers to harmful material. The scale of moderation therefore calls for careful division of labor, clear escalation criteria, and support for the people doing the review.

Build what differentiates you; buy what merely enables you

Start with your core value. An organization should decide whether to build or buy each part of an AI system by considering its strategic importance, available expertise, cost, quality, security, and time to market. A fashion retailer may gain little by creating its own data infrastructure, since that is not what makes its customers choose the brand. Buying standard tools can free the team to focus on fashion and customer experience.

Treat sensitive functions differently. Facial recognition in a phone’s photo library may be a minor convenience, while facial recognition used to unlock the device directly affects a brand promise of security. A company may reasonably accept an outside component for the first use and want tighter control over the second. Build and buy are not all-or-nothing choices; organizations can combine components from vendors, open-source tools, and internal systems.

Analysis

This is a strategic make-or-buy decision, not a referendum on whether internal engineering is inherently superior. The key question is where control creates durable value and where it simply consumes scarce attention. Security can favor either choice: a vendor may have stronger expertise, while outsourcing can increase dependency and reduce visibility. The decision should therefore include exit options, contractual rights, data portability, and the consequences of vendor failure. A component that is ordinary today can become strategically important if it shapes customer trust or creates a unique source of learning.

Analysis

The book’s strongest argument is that real-world AI is less a model-building exercise than a coordinated management discipline. The authors repeatedly connect technical performance to choices made outside the algorithm: which problem deserves investment, what counts as an acceptable error, who labels the examples, how outputs enter a workflow, and who notices when conditions change. Their stories make that claim memorable because failures often arise from a mismatch between a locally successful metric and the system’s wider purpose. The eBay search model increased purchases while hurting revenue; the IBM image model improved a score while learning offensive associations.

This focus on organizational design is valuable because AI distributes decision-making across data suppliers, product owners, engineers, executives, and users. Treating the system as a neutral tool obscures those responsibilities. The book also offers a sensible route from experimentation to deployment: choose a contained problem, establish a business baseline, test the complete workflow, and plan for maintenance before calling the pilot a success.

The framework is strongest for business applications where goals and workflows can be specified. Its language of measurable value becomes harder to apply when benefits are public goods, harms fall on people with little power, or short-term business returns conflict with fairness. In those settings, “the right metric” cannot be settled by management alone. Affected communities and independent review may need a role in defining acceptable outcomes. Similarly, high-quality data does not automatically make a decision legitimate; a precise prediction can still serve an unjust policy.

The practical conclusion is not that every company needs more AI. It is that organizations considering AI need stronger habits of problem definition, measurement, data stewardship, and accountability. The model is one part of a sociotechnical system. The system succeeds only when its predictions are useful, its limits are visible, and its human owners remain answerable for what it does.

Summary in 30 Seconds
Machine learning reflects the choices behind its data, labels, and objectives, so accuracy alone cannot establish fairness or usefulness. Tie model measures to customer and business outcomes, and begin with a contained task supported by reliable data. Only about 20 percent of AI pilots reach production, making workflow testing and scale planning essential. Deployed models need monitoring, rollback plans, clear ownership, and human oversight as conditions change. Cross-functional teams and governance keep technical work connected to its consequences.
Contains spoilers

Last updated:

Report Issue

Review Summary

4.01 out of 5
Average of 88 ratings from Goodreads and Amazon.

Real World AI receives praise for its practical insights on implementing AI in business. Readers appreciate its accessible language, real-world examples, and focus on ethical considerations. The book is commended for bridging technical and business perspectives, offering valuable advice on project management, data handling, and avoiding common pitfalls. While some find it simplistic, many consider it an excellent starting point for non-technical readers and business professionals. The book's coverage of AI ethics and responsible data use is particularly highlighted. Overall, it's recommended for those looking to understand AI implementation in business contexts.

Your rating:
4.51
173 ratings
Want to read the full book?

Glossary

Goldilocks problem

valuable, manageable first use

A first AI use case that is valuable to the business, narrow enough to solve, and supported by suitable data. It should have clear criteria for success, manageable risks, and a path to broader use. A simple, repeatable task with a measurable outcome is often a stronger starting point than a large, ambiguous challenge.

MLOps

machine-learning lifecycle operations

Practices and systems for managing machine-learning models across their working life. MLOps includes preparing models for deployment, putting them into production, monitoring their behavior, and refreshing them when performance changes. It connects model development with the operational work needed to keep a system reliable and useful.

transfer learning

adapt a model with specific data

A method that starts with a model already trained on a related task and adjusts it using data specific to a new use case. For example, a general speech-recognition model can be refined with examples from a particular call center. This can reduce the effort needed to build a specialized system, but the new data must reflect the people and conditions it will serve.

model drift

changing data weakens performance

A decline in a model’s usefulness as real-world conditions or incoming data change from what it encountered during training. Shifts in language, products, user behavior, or operating environments can make earlier examples less representative. Monitoring and updated data help teams identify drift and decide when to retrain or revise a system.

data provenance

data origin and transformation history

The record of where data came from, how it was collected, and what changes were made before it reached a model-building team. Provenance helps teams determine whether a dataset is complete, representative, and appropriate for its intended use. Without it, a model may learn from a partial or altered view of the real process.

data governance

rules for responsible data use

The policies, responsibilities, and oversight that guide how an organization collects, transforms, accesses, and uses data and models. Effective governance covers data quality, security, documentation, ethical concerns, and decisions among competing goals. It should fit the organization’s size while receiving visible executive support and clear ownership.

supervised learning

learn from labeled examples

A machine-learning approach that learns a mapping from inputs to expected outputs using example pairs. For instance, a system can learn to classify support requests when earlier requests have been labeled by people with the correct category. The resulting model uses those learned patterns to predict answers for new, unlabeled inputs.

accessory pollution

irrelevant accessories outrank products

A search-result problem described in the book’s eBay example. A model optimized for purchases per session placed inexpensive accessories, such as phone cases, above the more valuable phones shoppers were seeking. The results increased a narrow engagement measure but harmed the buying experience and revenue, illustrating the risks of optimizing a proxy metric.

FAQ

What's "Real World AI: A Practical Guide for Responsible Machine Learning" about?

  • Overview: "Real World AI" by Alyssa Simpson Rochwerger is a comprehensive guide on implementing responsible machine learning in real-world business scenarios. It emphasizes ethical AI practices and the importance of data quality.
  • Purpose: The book aims to demystify AI for business leaders and decision-makers, providing actionable insights into developing and deploying AI strategies effectively.
  • Structure: It covers a range of topics from AI basics, strategy development, data management, to scaling AI solutions, making it a practical resource for organizations at various stages of AI adoption.

Why should I read "Real World AI"?

  • Practical Insights: The book offers real-world examples and case studies, making complex AI concepts accessible and applicable to business contexts.
  • Ethical Focus: It emphasizes the importance of responsible AI, addressing potential biases and ethical considerations in AI deployment.
  • Comprehensive Guide: Whether you're a beginner or have some experience with AI, the book provides a roadmap for integrating AI into business processes effectively.

What are the key takeaways of "Real World AI"?

  • Data Quality is Crucial: The book stresses that high-quality, annotated data is more important than the model itself for successful AI implementation.
  • Cross-Functional Teams: Successful AI projects require collaboration across various business functions, not just data scientists.
  • Ethical AI Practices: It highlights the need for ethical considerations and bias mitigation in AI systems to ensure fair and responsible outcomes.

What are the best quotes from "Real World AI" and what do they mean?

  • "Responsible AI isn’t just good for business; it’s good for the world." This quote underscores the dual benefit of ethical AI practices, enhancing both business success and societal well-being.
  • "Success, much of the time, is all about what you choose to measure." It highlights the importance of selecting the right metrics to guide AI projects towards meaningful business outcomes.
  • "AI represents a massive shift in technology, as revolutionary as electricity or the internet." This emphasizes the transformative potential of AI in reshaping industries and business practices.

How does "Real World AI" define responsible AI?

  • Ethical Considerations: Responsible AI involves designing systems that are fair, transparent, and free from harmful biases.
  • Continuous Monitoring: It requires ongoing evaluation and adjustment of AI models to ensure they align with ethical standards and business goals.
  • Stakeholder Involvement: Engaging diverse teams in the AI development process helps identify and mitigate potential biases and ethical issues.

What is the significance of data in "Real World AI"?

  • Foundation of AI: The book emphasizes that data is the backbone of any AI system, and its quality directly impacts the model's performance.
  • Data Annotation: Properly labeled and annotated data is crucial for training accurate and reliable AI models.
  • Ongoing Data Management: Continuous data collection and management are necessary to adapt to changes and maintain model accuracy over time.

How does "Real World AI" suggest building a cross-functional AI team?

  • Diverse Expertise: The book recommends assembling teams with varied skills, including data scientists, product managers, and domain experts.
  • Collaborative Approach: Encourages regular communication and collaboration across different functions to align AI projects with business objectives.
  • Incentive Alignment: Aligning team incentives with overall business goals ensures that all members work towards common objectives.

What strategies does "Real World AI" propose for developing an AI strategy?

  • Identify Business Goals: Start by clearly defining the business problems AI is intended to solve.
  • Pilot Projects: Begin with small, manageable projects to demonstrate AI's value and build momentum.
  • Scalability Planning: Design pilots with scalability in mind to ensure they can be expanded into full-scale solutions.

How does "Real World AI" address AI biases?

  • Bias Identification: The book stresses the importance of identifying potential biases in training data and model outputs.
  • Diverse Data Sources: Using diverse and representative data helps mitigate biases and improve model fairness.
  • Regular Audits: Conducting regular audits of AI systems ensures they remain aligned with ethical standards and business goals.

What role does continuous learning play in "Real World AI"?

  • Adaptation to Change: Continuous learning allows AI systems to adapt to new data and changing environments, maintaining their relevance and accuracy.
  • Model Retraining: Regularly updating models with new data helps prevent performance degradation over time.
  • Feedback Loops: Implementing feedback mechanisms ensures that AI systems learn from real-world interactions and improve over time.

How does "Real World AI" suggest handling AI security concerns?

  • Data Protection: Emphasizes the importance of securing data to prevent unauthorized access and breaches.
  • Adversarial Defense: Implementing measures to protect AI systems from adversarial attacks and manipulation.
  • Compliance and Governance: Ensuring AI systems comply with legal and ethical standards to maintain trust and integrity.

What are the challenges of scaling AI according to "Real World AI"?

  • Infrastructure Needs: Scaling AI requires robust infrastructure to handle increased data and computational demands.
  • Cost Management: Balancing the costs of scaling with the expected business benefits is crucial for sustainable growth.
  • Cultural Shift: Organizations must foster a culture that embraces AI and supports its integration across various business functions.

About the Author

Alyssa Simpson Rochwerger and Wilson Pang are the authors of "Real World AI". Alyssa Simpson Rochwerger is a technology leader with expertise in AI and product management. She has held leadership positions at various companies, focusing on implementing AI solutions in real-world business scenarios. Wilson Pang is also an experienced professional in the field of AI and data science, having worked on large-scale AI projects for major tech companies. Together, they bring practical knowledge and insights from their extensive experience in the AI industry, which is reflected in their book's approach to AI implementation and ethical considerations in business contexts.

Download PDF

To save this Real World AI summary for later, download the free PDF. You can print it out, or read offline at your convenience.
Download PDF
File size: 0.38 MB     Pages: 20

Download EPUB

To read this Real World AI 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: 1.40 MB     Pages: 26
Want to read the full book?
Follow
Listen
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 25,
cancel anytime before.
Consume 2.8× More Books
2.8× more books Listening Reading
Our users love us
1M+ 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...