I can’t tell you how many development AI projects I’ve seen crash and burn. It’s not because the technology wasn’t impressive or the intentions weren’t noble. Because the team fundamentally misunderstood why generative AI works in a real-world development context.
The latest evidence comes from Dalberg Data Insights’ People-Centered AI Playbook, which systematically deconstructs the Silicon Valley mindset that is heavily imported into development contexts.
After working with NGOs, social enterprises and governments in the areas of health, agriculture, education and financial inclusion, their message remains consistent. Organizations need practical support to move from theory to action and don’t want to reinvent the wheel.
Sign up now for more insights into digital development
Problems with the technology-first mindset
The development sector is falling into the same trap that plagued the early ICT4D initiatives. This means assuming that if you import a methodology from a high-resource context, it will somehow work in an environment with very different constraints. The 18 AI applications we just highlighted show what’s possible, but they don’t explain why so many similar efforts have stumbled.
Dalberg’s framework begins with a fundamental premise. This means that before considering technology, teams need to root their ambitions in real user needs, organizational realities, and workflow challenges. Their six-step approach (discover, define, design, develop, pilot, scale) intentionally brings forward human research, which most teams treat as an afterthought.
Consider the “discovery” phase. This phase can take weeks of user interviews, workflow mapping, and organizational assessments before a single line of code is written. This is a fundamental rejection of the “build first, find users later” mentality that dominates mainstream AI development.
Three key insights that challenge conventional wisdom
This playbook is more than just a framework. This directly challenges how we think about AI deployment in low-resource environments. The central argument is provocative. Most AI projects fail because teams cut out human work to make the technology sustainable.
1. Responding to AI is about human resources systems
Most AI readiness assessments focus on technology infrastructure such as bandwidth, devices, and data pipelines. Dalberg turns this on its head by emphasizing what they say is “.”preparation of people“The level of willingness, skill, and willingness of targeted users, staff, and partners to adopt and sustain AI solutions.
This playbook refers to diagnostic tools that assess strategy, data maturity, ethical considerations, and organizational culture as key determinants of success. While Microsoft’s AI Readiness Assessment and GSMA’s AI Ethics framework are mentioned, Dalberg’s own DART assessment is specifically built with social impact in low-resource settings in mind.
This human-first approach explains why government-led initiatives that integrate existing infrastructure consistently outperform standalone digital solutions, as observed in our analysis of AI governance challenges.
2. Problem definition trumps solution innovation
The most contrarian element of the playbook is the definition phase, which systematically tests whether AI is the right tool for the identified challenges. These include decision-making frameworks that ask whether the task is high-volume, repetitive, or pattern-based, and whether it cannot be effectively solved with simple tools.
This represents a fundamental philosophical change. Rather than starting with AI capabilities and looking for applications, teams start with specific workflow challenges and test whether AI provides measurable performance improvements compared to alternatives such as workflow redesign, basic digital tools, training, or policy changes.
The framework includes clear guidance to flag challenges with non-AI approaches and avoid building AI for its own sake. In resource-constrained situations, deploying AI without a clear fit can waste time, introduce risk, and make systems more vulnerable.
3. Scaling means building robust systems
The final insight questions the mindset for successful AI implementation. Dalberg’s Scale phase isn’t about user acquisition. Scale is about institutionalization, continuous development, and adaptation to context.
Their framework recognizes that what works in one setting may not work in another, and as the solution expands across geographies and user groups, teams must reconsider assumptions, languages, and data flows. This adaptive approach stands in stark contrast to platform thinking that assumes universal applicability.
The handbook emphasizes that scaling requires a move from activity tracking to impact assessment using proportionate and reliable methods to examine performance, equity, and cost-effectiveness. This evidence-based, scalable approach explains why USAID’s AI implementation guidance emphasizes continuous learning and iterative development.
Cross-cutting enablers: Where the real work happens
Perhaps most importantly, the playbook identifies three cross-cutting enablers that drive all six phases: talent, equity and inclusion, and data governance. These are basic design requirements rather than additional considerations.
- people dimension recognizes that success depends on building trust, aligning leadership, and equipping teams with the skills and confidence to use AI responsibly. This human-centered approach requires people to remain at the center, engaged, trained and supported for adoption.
- Equity and inclusion framework Teams should examine who is included in the data, who is participating in testing, and who faces barriers such as connectivity limitations, literacy, language, and access to devices. This systematic attention to inclusivity helps prevent unintended harm and enables AI to deliver value across a variety of needs and contexts.
- data governance Covering data quality, access, privacy, security, and compliance throughout all phases to ensure AI systems are ethical, reliable, and context-appropriate.
Implementation reality check
This playbook acknowledges what practitioners already know: Few teams have all the skills in-house.
Their pragmatic approach suggests partnerships with universities, regional technical assistance groups, or global networks for specialized support, while outsourcing short-term tasks such as labeling data or training in-house teams in core functions.
This collaborative model is consistent with our observation that successful AI initiatives require a multidisciplinary approach that combines technical expertise and domain knowledge. As Stanford University’s Institute for Human-Centered AI demonstrates, bringing together computer scientists, ethicists, social scientists, and domain experts creates more robust and sustainable solutions.
The framework also provides practical templates for user persona development, problem statement framework, use case definition, and feasibility assessment. These tools transform abstract methodologies into practical workflows that your team can immediately implement.
What this means for our development practices
This playbook is important because it provides a methodologically rigorous alternative to both AI evangelism and AI skepticism.
We will not deny the potential of AI or accept its introduction uncritically. Instead, it provides a systematic approach to determining when, where, and how AI can create meaningful value in a development context.
This framework’s emphasis on repetition and evidence-based decision-making reflects what we know about successful technology adoption in resource-limited settings. This means solutions need to be designed for local conditions, validated through real-world testing, and adapted based on user feedback.
Most importantly, this handbook positions AI as one tool among many, rather than an inherent good. By requiring teams to justify AI solutions against alternatives and measure their impact against defined value drivers, you foster responsible innovation that addresses user needs rather than technical possibilities.
For development organizations considering AI initiatives, this framework provides both a roadmap and a reality check.
The future of developing AI will not be determined by algorithmic advances or funding announcements. It depends on our willingness to do the human-centered work that makes technology truly useful. This playbook shows you how.
