AI isn’t failing we’re building it wrong

AI Isn’t Failing, We’re Building It Wrong: Insights from Web Summit Vancouver 2026

At Web Summit Vancouver 2026, a panel featuring Nancy Wang, Jennifer Smith, and Linda Tong, moderated by John Koetsier, explored why AI implementations often underperform. The core message was that AI isn’t inherently flawed; rather, organizations are building it incorrectly. The discussion centered on crucial missing elements in current AI strategies, from context to infrastructure and human integration.

Ms. Smith, Co-Founder and CEO of Scribe, identified a critical gap: the “context layer.” She explained that vital information about specific workflows, edge cases, and dependencies often resides only in human minds, making it inaccessible to AI agents. This lack of organizational legibility prevents even sophisticated foundational models, trained on general skills, from effectively executing company-specific tasks.

Ms. Wang, Partner at Felicis Ventures and CTO at 1Password, highlighted security and identity challenges. She argued against forcing AI agents into traditional identity access management (IAM) patterns. Instead, agents require their own distinct identities and mechanisms to bind intent. This is essential for monitoring their non-deterministic actions and ensuring secure, appropriate access to sensitive systems, a key convergence point for the security industry.

Ms. Tong, CEO of Webflow, emphasized that many companies are merely adding AI features to existing systems, creating “AI-assisted” rather than truly “AI-native” solutions. She stressed the need for a complete infrastructural overhaul, urging organizations to rethink their architecture, data schemas, and modeling from the ground up to fully harness AI’s transformative power.

Ms. Wang further explained the disparity between powerful AI models and practical results. She noted that technological progress outpaces organizational change. To bridge this gap, companies must prioritize robust systems, processes, and change management. AI initiatives should be driven by clear, measurable business goals and a defined return on investment, moving beyond simply adopting AI for its own sake.

The panel also addressed the financial efficiency of AI, particularly “token maxing.” Ms. Tong and Mr. Koetsier advocated for “local maxing”—strategically choosing the right AI model for each task to avoid unnecessary costs. Using an overpowered model for simple operations is inefficient. Ms. Wang added that architectural design, like caching web crawls, and implementing spending caps are vital for reducing token expenditure.

Ms. Smith reiterated that providing more context significantly enhances AI agent performance. However, connecting AI to extensive internal tools and knowledge bases introduces substantial, often overlooked, security and permissioning challenges. Many organizations are currently prioritizing rapid innovation, potentially turning a blind eye to these critical issues that will require resolution.

The discussion highlighted AI’s impact on work roles, with Ms. Tong noting the blurring of traditional boundaries. AI empowers individuals across various functions to become “builders” who rapidly translate ideas into code. This shift fosters immense productivity and enthusiasm, as employees achieve more in a day than previously imagined, though it necessitates adapting to new skill sets and mindsets.

Finally, the panelists considered what AI cannot replicate. Ms. Tong pointed to human creativity, judgment, and taste, noting AI’s current strength in remixing existing ideas rather than generating truly novel ones. Ms. Wang added human agency—the intrinsic drive to define goals and make things happen. These unique human traits, along with the ability to choose between AI and simpler automation, remain indispensable.