Abstract
Responsible AI deployment requires solving two challenges: first, building systems that work without historical performance data, and second, figuring out which decisions AI should make versus which ones need to stay in human hands. My technical project tackled the engineering side of this problem through designing and testing an AI-driven sales prospecting system in a data-scarce environment. My STS paper researched the governance side by
investigating how organizations decide which operational decisions to hand off to AI. Together, these projects show that successful AI deployment requires both engineering discipline to align stakeholder needs and organizational structures to govern what AI is actually allowed to decide.
My technical project asked how to build and deploy an AI-assisted lead generation software for Zbooni, a UAE-based B2B conversational-commerce platform, where no historical conversion data existed to train models. Sales representatives were spending roughly 30% of their working hours on manual prospecting research while converting leads at around 1%. Our interviews with them uncovered conflicting incentives: sales reps optimized for outreach volume while management wanted quality-based targeting and pipeline visibility, but both groups shared the goal of improving acquisition efficiency. Using systems engineering methodology, we applied objective-tree analysis to break these conflicting objectives into five unified system requirements and built a traceability matrix linking each component back to its stakeholder origin. The main contribution is a five-phase validation framework built for zero-baseline environments: it starts from converting stakeholder intuitions into scorable features, and deciding to only turn on machine learning after the system has generated enough real performance data to justify it. In a single run, the system processed 12,720 businesses and produced 2,496 active leads, with 1,398 classified as workable and 19.1% scoring in the highest qualification tier. The framework is domain-independent, meaning that it is transferable to any organization deploying AI qualification without data.
My STS research asked why organizations with similar AI capabilities draw fundamentally different lines around what AI is allowed to decide, with some keeping it advisory and others letting it determine outcomes directly. Using Arnold Pacey’s technology practice framework, I analyzed published case studies, regulatory guidance, and academic research across healthcare, criminal justice, hiring, customer service, and government welfare. Three findings emerged. First, institutional accountability pressures such as liability exposure, regulatory requirements, and reputational risk set the formal boundary of AI delegation. Cases like the Air Canada chatbot ruling and Amazon’s abandoned hiring tool show that organizations can delegate tasks to AI, but cannot delegate responsibility for its outcomes. Second, professional culture shapes how delegation actually works in practice: radiologists in the same hospital responded to identical AI diagnostic tools in entirely different ways, with some treating AI as a second opinion and others deferring to it, based on their professional norms around expertise. Third, governance structures determine whether delegation boundaries hold over time. Delegation works when governance is specific, embedded, and enforceable, and falls apart when it stays symbolic or voluntary, as Australia’s Robodebt scheme illustrates. AI delegation, in short, is a sociotechnical negotiation over authority, accountability, and control, and not a technical configuration problem.
Combined, these projects offer complementary perspectives on deploying AI in organizations. My technical project shows that systems engineering can resolve stakeholder misalignment and get AI qualification off the ground under data scarcity, while my STS research paper shows that even well-designed systems need organizational structures to govern what AI is
permitted to decide. Both projects have limits. The technical results come from a pilot where
outreach and conversion numbers are still pending, and the STS analysis relies on published case studies rather than internal records. The next step would be to study how delegation boundaries change as systems like these build up real performance data, and whether letting AI take on more responsibility produces better organizational buy-in than automating everything at once.
I would like to thank my Capstone advisor, Professor Matthew Burkett, for his guidance throughout the project, and my STS advisor, Professor Caitlin Wylie, for sharpening my thinking on sociotechnical analysis. I would also like to thank my capstone teammates, Madeline Priebe, Carly Elbaum, Zack Sikkink, Ian Girdner, and Dylan Jones, whose collaboration made the LeadFlow system possible. Lastly, I am grateful for the Zbooni team for their partnership and participation in the stakeholder interviews that shaped both the system design and my understanding of AI deployment in practice.
