Carnegie Mellon Software Engineering Institute
The race to deploy generative and agentic AI has generated extraordinary momentum—and an equally extraordinary amount of confusion. Organizations are counting on breakthrough gains, yet many are instead facing mounting technical debt, inconsistent outcomes, and AI initiatives that quietly stall or fail. Despite decades of progress in software engineering, AI adoption is moving faster than most organizations can reliably engineer, govern, or sustain any of their AI initiatives. This raises a critical question: Is today’s rush toward generative AI repeating the maturity challenges that once hampered early software engineering—and can disciplined, capability-driven approaches guide us out of the chaos again?
To tackle this challenge, the Carnegie Mellon Software Engineering Institute and Accenture have partnered to develop a new AI Adoption Maturity Model. Drawing on executive interviews, survey data, background studies, and pilot engagements, we identified the core engineering and organizational practices required for predictable, repeatable, and scalable AI adoption—capabilities essential to delivering AI-enabled systems and workflows that achieve measurable value.
Today’s challenge is no longer simply embracing AI; it is keeping pace with rapid technological change while strengthening the engineering practices and tools needed to build reliable, trustworthy, and maintainable AI-enabled systems and workflows—whether created by humans, AI, or anything in between. This keynote shares insights from the development of the AI Adoption Maturity Model, outlining a path to balance continuous evolution of practice with the foundational discipline needed to achieve trustworthy, enterprise-level AI impact.
Dr. Ipek Ozkaya is the Technical Director of the Engineering Intelligent Software Systems group at the Software Engineering Institute (SEI) at Carnegie Mellon University. Her work is at the intersection of technical debt management, AI-augmented software engineering, AI-enabled systems development, and large-scale modernization. She has spent much of her career advancing how we understand and reduce technical debt in systems, including co-authoring a practitioner book titled Managing Technical Debt: Reducing Friction in Software Development. Her current work focuses on shaping the techniques, practices, and tooling that enable AI-native software engineering and enterprise-wide AI adoption. At the SEI, she puts her research into practice by helping government and industry organizations to solve their software and AI challenges.
PUC-Rio, Brazil
AI engineering for discovery is rarely limited by the capabilities of individual models. In practice, the dominant challenges arise at the system level: integrating heterogeneous data and simulators, managing evolving requirements and operational constraints, validating results under uncertainty, and ensuring that complex workflows remain observable, maintainable, and trustworthy over time. This keynote offers a retrospective of my prior work developing technologies and application frameworks for subsurface characterization, decision-making under uncertainty, and materials discovery. These three domains differ in data modalities and scientific processes, yet converge on similar end-to-end engineering difficulties.
Building on these lessons, we argue that the adoption of foundation models and agentic workflows creates new opportunities for discovery applications, but also introduce new risk surfaces. To deliver real impact, we need architectures that go beyond “agents as automation” and treat trust as a first-class system property. We introduce the concept of a governable agentic knowledge fabric, which combines: (i) knowledge representations and provenance-aware evidence tracking, (ii) orchestrated agents and tool ecosystems spanning data, simulation, and reasoning, and (iii) governance mechanisms that support accountability, auditability, robustness, and safe evolution across the AI lifecycle.
The talk outlines key building blocks and design tactics for such fabrics, including modular pipeline architectures, evaluation harnesses aligned with stakeholder goals, uncertainty-aware validation strategies, human-in-the-loop patterns for expert sense-making, and operational feedback loops to detect drift and manage change. We conclude by connecting these ideas to our current work at the PUC–Behring Institute for AI, framing a knowledge-centric research agenda for engineering reusable, dependable AI-enabled systems that can scale discovery workflows across disciplines and industries.
Renato Cerqueira is Director of the PUC-Behring Institute for Artificial Intelligence at PUC-Rio, whose mission is to advance and disseminate AI technologies that amplify human potential in creating, sharing, and applying knowledge to tackle urgent socio-economic challenges. From 2011 to 2025, he was a Senior Research Manager at IBM Research, where he helped establish the company’s Brazil Research Lab and led a group focused on AI for decision support and knowledge management, with applications in Oil & Gas, Sustainable Materials, Mining, Agriculture, and Finance. From 2002 to 2011, he served as faculty member of PUC-Rio’s Department of Informatics, returning in 2025 as Collaborator Professor in its Postgraduate Program. He holds a B.Sc. in Computer Engineering, an M.Sc. and Ph.D. in Informatics from PUC-Rio, and a postdoctoral fellowship at the University of Illinois at Urbana-Champaign. His current interests include Epistemic AI, Governable & Trustworthy AI, engineering of AI-based systems, multimodal knowledge representation, neuro-symbolic AI, and AI infrastructures for knowledge governance, integration, and curation (knowledge fabric).