An EY report claimed that agentic AI could handle 80% of routine treasury tasks. ^[1] Agent memory is defined as information written to external storage during runtime and retrieved in later calls, distinguishing it from system prompts, conversation history, and static knowledge bases. ^[2] The author states that every AI agent built today eventually encounters a wall caused by memory limitations rather than a lack of reasoning power or coding ability. ^[3] The EY report noted that companies must first establish reliable data systems. ^[4] The same EY report said this would let treasury professionals focus on strategic work. ^[5] Agentic AI can execute multiple operations across various systems within seconds, unlike humans whose impact is limited by reaction speed. ^[6] Eyal Alkalay, CIO of Verbit, stated that the solution to AI security concerns is not to slow down the application of AI. ^[7] If Agentic AI possesses excessive authority and receives malicious instructions, gets malware injected, or simply makes a wrong decision, the consequences will spread very rapidly. ^[8] The challenge for the industry and Verbit is to ensure that security, identity, authority, and monitoring develop at a speed equivalent to AI agents. ^[9] Agentic AI systems are defined as networks of specialized AI agents. ^[10] The article asserts that real-world tasks are rarely solvable by a single agent. ^[11] The author states that when a single AI model is given a task, it performs quite well. ^[12]