Technology Overview
The AI technology stack for commerce — foundational models, agentic AI, LLMs, privacy & security, investment costs, and what comes next — based on Book Part 5 of the AI Best Practices for Commerce reference.
Agentic AI
From assistants that suggest to agents that act
Risks and Challenges
The flexibility that makes agents powerful also makes them unpredictable. As deployment expanded through 2024, patterns emerged around failure modes and risks.
Unexpected actions occurred when agents optimized for stated goals using strategies designers hadn’t anticipated. An agent tasked with a metric could find creative ways to improve that metric that violated implicit assumptions about appropriate behavior. The solution wasn’t obvious, making rules more explicit often led to brittle systems, while keeping goals abstract risked continued surprises.
Stuck loops paralyzed systems when agents couldn’t determine the next action. Edge cases that violated an agent’s assumptions could create circular reasoning: The agent tried to look up information, which pointed to another lookup, which pointed back to the original, consuming resources until automatic limits intervened.
Unsafe tool execution happened when agents had access to powerful tools but insufficient judgment about when to use them. An agent given access to systems and instructed to achieve a goal might use those systems in technically correct but contextually inappropriate ways.
Hallucinations, language models generating plausible but false information, caused particular problems in customer-facing applications. Agents sometimes provided confident but incorrect information about policies, compatibility, or timelines. The agents weren’t lying; they generated responses that sounded authoritative based on patterns in training data but weren’t grounded in actual facts.
Cascading failures worried system architects. When agents depend on other agents, failures propagate. A bug in one agent’s output could feed into another agent’s decision-making, compounding errors across the system.
The industry’s response focused on governance frameworks rather than trying to eliminate risks entirely. The recognition grew that agents, like humans, would make mistakes. The question became how to detect, contain, and recover from those mistakes rather than prevent them absolutely.
Last updated: March 12, 2026