HS141: EA vs AI

The episode opens with Jon Burr explaining why EA and AI coexistence is challenging: EA aims to create known, reliable, and deterministic IT environments by defining boundaries and data flows, while AI—particularly agentic AI—operates in emergent, probabilistic, and unpredictable ways, often crossing data and functional boundaries to achieve immediate results without long-term perspective.

Jonna challenges whether EA is actually necessary, prompting Jon to distinguish between EA as a high-level architectural vision versus low-level engineering implementation details. They agree that true enterprise architecture must include infrastructure considerations and should define what systems exist and how they interconnect at a strategic level, without prescribing implementation minutiae. The debate hinges on the level of abstraction appropriate for architecture versus engineering.

When discussing whether AI eliminates the need for EA, Jon argues that relying on AI to maintain boundaries is contradictory given AI's documented inability to respect boundaries. Jonna counters that some organizations might adopt a framework where AI builds and modifies software on-the-fly, with governance enforced through access controls rather than traditional architecture. Jon responds that this reasoning is circular: without understanding what data belongs where and why, you cannot create meaningful governance policies.

They explore whether AI could replace detailed architectural documentation with dynamic, AI-assisted architecture updates—quarterly or monthly reviews where AI suggests changes while humans validate them. Both agree this could work if humans remain in the loop and changes occur on human-friendly business cycles rather than machine speed.

The hosts emphasize that architectural understanding depends on organizational context: UPS-level complexity requires street-level detail maps, while simpler organizations need only interstate-level views. However, both assert that absent significant improvements in AI reliability (such as implementing ACID properties for AI), enterprise architecture is more important than ever for providing guardrails and governance.

They distinguish between different forms of AI, cautioning against using LLMs for mission-critical, deterministic requirements like algorithmic trading, aviation safety, or medical systems. The more you force non-deterministic AI into deterministic applications, the less reliable the results. Finally, Jonna invokes Hilary Putnam's critique of the Turing test to argue that output-based definitions of intelligence ignore semantics and meaning—divorced outputs from reality eventually lead to failures when stakes are highest.