
Making AI Work for People
Description
Alles über E-Books | Antworten auf Fragen rund um E-Books, Kopierschutz und Dateiformate finden Sie in unserem Info- & Hilfebereich.
Design AI applications that inspire trust, solve real problems, and put people first
Making AI Work for People: A Framework for Designing and Building Impactful AI-Powered Applications by Asmaa Ibrahim offers software engineers, product managers, and app designers a comprehensive framework for creating AI-powered applications that truly serve humanity. Ibrahim, Product Lead at Aleph Alpha, experienced ML team leader, AI-driven transformation expert, and leader of the team that created Gemini in Firebase, addresses the critical challenge facing today's tech professionals: how to harness AI's transformative potential while guaranteeing applications remain ethical, accessible, and aligned with human values.
This book explores cutting-edge innovation and provides the PRESS framework for designing human-first solutions that enhance - rather than replace - human capabilities. The PRESS framework is a systematic approach that guarantees AI-powered products are designed with five essential principles in mind: People-first design, Responsible in the core, Explainability and interpretability, Safety via benchmarking, and Sustainability and long-term impact. The author offers real-world case studies, proven development methodologies, and insights from successful AI implementations across the telecom, retail and other sectors, showing you how to navigate the complex landscape of AI integration while remaining laser-focused on user needs, trust, and ethics. The book covers everything from identifying opportunities for AI integration and measuring success with human-centered KPIs to scaling applications from MVP to production-grade systems.
Making AI Work for People also includes:
- A comprehensive PRESS framework that guides every stage of AI product development from conception to deployment
- Practical strategies for human-AI collaboration that amplify rather than replace human strengths and capabilities in real-world applications
- A proven evaluation flywheel to guarantee that AI applications meet real user needs
- Real-world case studies from telecom and retail demonstrating successful human-first AI implementations
- Essential guidance for scaling AI applications while managing costs, infrastructure challenges, and long-term maintenance requirements
This book is essential for software engineers integrating AI into product development, product managers leading AI initiatives, and app designers creating human-AI interfaces who want to build applications that inspire trust and deliver genuine value through the PRESS framework. It's also a must-read for executives and managers of technical teams seeking crucial insights into systematic, responsible AI development practices that drive measurable business outcomes.
More details
Other editions
Additional editions

Content
CHAPTER 1
The Trillion-Dollar Challenge
The AI revolution was supposed to transform business. Instead, we're looking at a massive failure epidemic.
The RAND Corporation (2024) supports this claim: the primary reason AI projects fail is due to a misunderstanding or miscommunication of the problem that AI needs to address. This highlights leadership and organizational challenges, rather than issues with the technology itself. Other studies back this up, some showing failure rates as high as 87%. MIT Sloan Management Review and BCG's 2019 "Winning With AI" study surveyed over 2,500 executives and found that "seven out of 10 companies surveyed report minimal or no impact from AI so far" (Ransbotham et al., 2019). It's gotten worse since then. Compare that to regular IT projects, where maybe 25-50% fail, and you'll see we've got a real problem here.
This scenario exemplifies a common pattern across industries. I've seen this pattern repeated across major telecom operators. The operator is serving millions of customers and dropped millions on AI-powered network optimization. The pilots looked amazing, with near-perfect accuracy and executives popping champagne. Six months later? Different story. The system was running 70% slower than promised. The system couldn't communicate with 40% of their legacy infrastructure. They burnt millions on emergency maintenance. And the kicker: their engineers ignored the AI's recommendations 85% of the time. I call this company "EuroConnect" and you'll see their journey throughout this book as they learned from failure and eventually succeeded. As illustrated in Figure 1.1, their timeline shows a predictable pattern of initial success followed by cascading failures.
Figure 1.1: EuroConnect's AI implementation timeline showing the progression from pilot success to multiple system failures over 18 months.
The retail sector faces similar challenges. A company developed an AI inventory tracker that excelled during testing. Put it in the real world, and it couldn't do the one job it was built for: tracking out-of-stock items. Years later, after throwing good money after bad, they're still trying to make it work.
This chapter breaks down why AI-powered application development fails so spectacularly, what actually works, and how you can avoid joining the 80% club. After seeing and observing the development of hundreds of projects across telecom, retail, and beyond, here's what I've learned: the technology usually works well. The most common causes of failure are organizational and human factors-misunderstanding the problem, poor change management, and lack of user buy-in. This is actually good news, because those problems? We can fix those.
Stats About AI-Powered System Development in Enterprise
Let's talk numbers. They paint a sobering picture of where we really stand with enterprise AI adoption.
The Investment-Reality Gap
From 2021 to 2025, we've watched AI go from "the future" to "a failure." It's a weird situation. Industry research consistently shows that approximately 80% of AI projects fail to deliver their expected value-with RAND Corporation (2024), Gartner, and other major analysts all documenting similar failure rates. But the ones that work are making serious money, averaging returns of 3.7 times the investment.
The failure rate keeps climbing. S&P Global Market Intelligence (2025) just reported that 42% of businesses expect to kill most of their AI projects in 2025. In 2024, that number was only 17%. That's more than double in 12 months. Gartner says that less than half of AI projects even make it to production, and those that do take about eight months to get there (CIO Dive, 2025). Informatica predicts that 30% of GenAI projects will die right after the pilot phase (Informatica, 2025). And BCG Global (2024) found something really depressing: only 4% of companies actually make real money from AI.
Where the Money Goes
The financial side of this is mind-blowing. Companies regularly blow two or three times their AI budget. European businesses have it worse; their budgets are 70% smaller than U.S. companies to begin with, plus their energy costs jack up data center expenses by half.
When AI works, though, it really works. Microsoft and IDC surveyed 4,000 business leaders and found winners reporting 3.7 times returns on average. Note that these are self-reported figures from organizations, which may include reporting bias. The best performers had 10 times returns (IDC, 2024). But here's where it gets interesting. BCG surveyed finance executives about AI returns specifically in finance functions, and less than half could even measure their ROI. The ones who could? They were seeing around 10% returns. Not exactly the revolution everyone's selling-and that's just in finance departments, where you'd expect the numbers people to have this figured out.
While organizations target 20% returns, research shows that a third achieve less than 5%. Another quarter gets between 5% and 10%. McKinsey & Company (2025) reports that only 1% of companies have really figured out AI. No wonder the results are so bad.
Reality Check for ROI
Even though many AI projects fail, companies that do it right get great results. Microsoft and IDC's study of 4,000 leaders found that the average return on investment (ROI) was 3.7 times, and the best performers got 10.3 times their money back. But the BCG survey of more than 280 finance executives found a lower median ROI of 10%, and only 45% of them could even say how much they made.
Different AI, Different Story
The type of AI matters. Figure 1.2 explores these three identified categories of AI.
- Traditional machine learning (ML): Represents the foundation of AI implementation, operating through mathematical models that learn patterns from historical data to make predictions or classifications. These models deliver predictable, deterministic outputs where the same input always produces the same result. They excel at well-defined tasks with structured data, offering clear decision boundaries and explainable logic. While these models lack the flexibility of newer approaches, they provide reliability, consistency, and proven ROI for specific business problems. Their straightforward nature makes them ideal starting points for organizations beginning their AI journey (Hastie et al., 2009; Murphy, 2012).
- Generative AI (GenAI): Fundamentally transforms how we interact with machines by creating new content rather than simply analyzing existing data. Unlike traditional ML's predictable outputs, GenAI operates probabilistically, producing varied but contextually appropriate responses to similar inputs. This technology enables natural language interaction, creative problem-solving, and complex reasoning capabilities that feel remarkably human-like. The flexibility comes with unique challenges: variable costs based on usage complexity, the potential for generating plausible but incorrect information, and the need for sophisticated prompt design to achieve desired outcomes (Vaswani et al., 2017; Brown et al., 2020; OpenAI, 2023).
- Specialized models: These include domain-specific models (like medical imaging AI and autonomous vehicle perception) and specific architectures (like computer vision models and speech recognition systems). Bridging the gap between traditional ML's reliability and GenAI's flexibility, they are engineered for specific industry applications or complex tasks. These systems combine multiple AI techniques optimized for particular domains, delivering superior performance within their intended scope. What distinguishes them is their deep integration of domain knowledge into the model architecture itself. They require significant expertise to develop and maintain but offer unmatched accuracy and efficiency for their specific use cases. Organizations choose specialized models when performance in a particular area matters more than general-purpose adaptability (LeCun et al., 2015; Goodfellow et al., 2016).
Figure 1.2: AI technology popular types with the key considerations.
Generative AI has seen a more rapid adoption than any other technology in recent history. Two-thirds (67%) of those surveyed reported that their companies are currently using generative AI, and over a third of this group (38%) reported that their companies have been working with AI for less than a year (O'Reilly Report, 2023). "The adoption of generative AI is certainly explosive, but if we ignore the risks and hazards of hasty adoption, it is definitely possible we can slide into another AI winter," said Mike Loukides, vice president of content strategy at O'Reilly and the author of the report.
Computer vision systems, as a specialized AI category, are more likely to work well in factories, where vendors claim they can be significantly more accurate than manual inspections for quality control, although independent verification of specific accuracy improvements varies by implementation (VisionPlatform, 2024). The market for computer vision systems was worth $11.22 billion in 2021 and is growing at a rate of 7.0% per year (Softweb Solutions, 2022).
Remember that, while generative AI captures headlines, traditional machine learning models remain foundational to production AI systems, powering critical applications...
System requirements
File format: ePUB
Copy protection: Adobe-DRM (Digital Rights Management)
System requirements:
- Computer (Windows; MacOS X; Linux): Install the free reader Adobe Digital Editions prior to download (see eBook Help).
- Tablet/smartphone (Android; iOS): Install the free app Adobe Digital Editions or the app PocketBook before downloading (see eBook Help).
- E-reader: Bookeen, Kobo, Pocketbook, Sony, Tolino and many more (not Kindle).
The file format ePub works well for novels and non-fiction books – i.e., „flowing” text without complex layout. On an e-reader or smartphone, line and page breaks automatically adjust to fit the small displays.
This eBook uses Adobe-DRM, a „hard” copy protection. If the necessary requirements are not met, unfortunately you will not be able to open the eBook. You will therefore need to prepare your reading hardware before downloading.
Please note: We strongly recommend that you authorise using your personal Adobe ID after installation of any reading software.
For more information, see our ebook Help page.