Descriptions:
Raj Navakoti, a staff software engineer at IKEA’s Delivery Services organization (a domain spanning over 100 engineers across six product teams), delivers a workshop at AI Engineer on a problem he argues is underexamined: why enterprise AI agents fail to move the needle on actual delivery metrics despite strong general capabilities.
Navakoti opens with a McKinsey data point — 88% of companies report using AI, yet only 6% see meaningful value creation — and traces the gap to a context problem. Drawing on the film Memento as an analogy, he argues that today’s agents are strong on reasoning and code generation but weak on institutional knowledge: the domain-specific context that defines what ‘done’ means in a real enterprise. His proposed framework, which he calls demand-driven context, inverts the typical approach: rather than engineers pre-loading agents with context, the agent itself surfaces what it needs, generating probes to test knowledge coverage before acting.
A live demo shows a platform operations agent ingesting 20 historical incident records and cross-referencing them against a flat-file knowledge base (a stand-in for Confluence). The agent auto-generates test probes, runs them against the knowledge base, and returns a scored breakdown of which incident types are fully documented, partially covered, or entirely undocumented — surfacing tribal knowledge gaps at scale. For engineering leaders trying to close the gap between AI capability and sprint delivery, this session offers a concrete, enterprise-grounded starting point.
📺 Source: AI Engineer · Published May 05, 2026
🏷️ Format: Deep Dive
