Agentic AI moves practical

Banks are starting to use “agentic” artificial intelligence as supervised casework software, not as a machine that makes final compliance calls on its own. (financederivative.com) In the banking commentary that prompted this debate, Quantexa executive Alexon Bell wrote that traditional machine learning still accounts for about 85% of bank AI workloads, generative AI about 10%, and agentic AI 5% or less. He pointed to uses such as know-your-customer reviews, investigations and sales relationship management. (financederivative.com) The pitch is speed on messy internal work. Boston Consulting Group wrote on March 9, 2026, that banks are testing agents to extract information from documents, analyze individual cases and escalate exceptions while keeping “full audit trails.” (bcg.com) What banks cannot skip is governance. The Federal Reserve and Office of the Comptroller of the Currency said in Supervisory Letter SR 11-7, issued April 4, 2011, that banks need robust model development, validation, governance, policies and controls for systems used in business decisions, risk measurement and regulatory reporting. (federalreserve.gov) That old rulebook still frames a lot of the current argument. The Global Association of Risk Professionals wrote on February 27, 2026, that SR 11-7 remains “one of the few stable reference points” for model governance even as agentic systems test assumptions about fixed parameters and decision paths that can be reconstructed after the fact. (garp.org) In plain terms, an agent is useful only if a human reviewer can see where the answer came from. Bell’s example was a corporate investigation: if the system confuses an 11-entity global firm with a four-entity small business, the result is not just wrong but unusable in a regulated review. (financederivative.com) That is why data preparation keeps showing up in these discussions. Bell wrote that about 80% of a data scientist’s time is spent preparing data, and argued that generative and agentic systems fail in the same way older models do when the underlying records are incomplete or poorly prepared. (financederivative.com) Banks also appear to be moving faster than their internal controls. Wolters Kluwer said on February 26, 2026, that 31.8% of 148 financial institutions surveyed had AI or machine learning in production, but only 12.2% described their strategy as well-defined and resourced, and just 35.8% had internal policies for ethical AI use. (wolterskluwer.com) The same survey found 58.8% of respondents wanted more regulatory guidance, while 28.4% cited explainability and transparency as their top regulatory concern. That lines up with the practical version of agentic AI now taking hold in banks: software that assembles evidence, shows its work and hands the file to a person. (wolterskluwer.com)