> ## Content Index > Fetch the complete content index at: https://vault.up-state.ai/llms.txt > Use this file to discover other available public pages before exploring further. # Before Your Bank Acquires an Agent, Ask Who Authorizes Its Actions - URL: https://vault.up-state.ai/week-in-banking-ai-2026-10-12/ - Published: 2026-10-13T19:10:19.000Z - Updated: 2026-10-13T19:10:19.000Z - Description: Before buying an AI agent, ask it to demonstrate who authorizes each action, how your team can intervene and whether it improves the whole job. - Author: Ben Nichols - Tags: Commentary *Week in Banking AI | Analysis | Upstate AI | October 12, 2026* An AI agent that explains how to block a lost card is doing a different job from one that blocks it. Banking software provider Backbase is selling the second: its [October 8 launch announcement](https://www.backbase.com/press/backbase-launches-conversational-banking-to-turn-customer-conversations-into-governed-banking-action?ref=vault.up-state.ai) says Conversational Banking is generally available and can block and replace cards, within permissions and approvals the bank defines. For banks and credit unions evaluating these tools, the useful demonstration starts where the conversation ends. We want to see how the system obtains authorization, carries out the action and stops when permission is missing. ## Ask to see where it stops Backbase says banks choose which information and tools agents can access, which actions they can perform and where human approval is required. Its announcement also promises [explicit customer confirmation for consequential requests and records of context, decisions and outcomes](https://www.backbase.com/press/backbase-launches-conversational-banking-to-turn-customer-conversations-into-governed-banking-action?ref=vault.up-state.ai). Those are testable promises. In a demonstration using synthetic accounts and no live banking connections, ask the vendor to block a card, then repeat the workflow with an expired login or a customer who withdraws confirmation before execution. Have your operations lead inspect the attempted actions and the records. Does the system prevent the unauthorized action, or merely explain afterward why it shouldn't have happened? The word “agentic” won't tell you how much authority a product needs. Financial technology provider FIS announced its [Agentic Insights Suite for chief financial officers on October 12](https://www.finextra.com/newsarticle/48568/fis-launches-agentic-ai-suite-for-the-office-of-the-cfo?ref=vault.up-state.ai). Its [own release](https://www.fisglobal.com/about-us/media-room/press-release/2026/fis-equips-cfos-with-agentic-ai-to-forecast-cash-flow-and-surface-risk-earlier?ref=vault.up-state.ai) describes gathering data, analyzing cash and risk, and recommending action across collections, liquidity and cash flow. That description doesn't establish autonomous payment authority. The purchasing question is what the particular workflow can do. ## Who notices when it goes wrong? ![Illustration of a dashboard for monitoring activity.](https://storage.ghost.io/c/c5/ce/c5ce462a-7219-48d0-892f-946fcccc5d9e/content/images/2026/10/undraw_dashboard_p93p.png) Claude developer Anthropic supplied a sobering example on October 9\. In its [account of unintended model actions](https://www.anthropic.com/research/investigating-unintended-model-actions?ref=vault.up-state.ai), the company acknowledged that a model submitted an invented homicide tip through a police website during an evaluation. Instructions prohibited several online activities but didn't rule out form submissions. The tip was caught by a spam filter and never forwarded for investigation. [TechCrunch's October 9 report, which includes a Philadelphia Police Department statement](https://techcrunch.com/2026/10/09/an-anthropic-ai-model-sent-a-false-homicide-tip-to-philadelphia-police/?ref=vault.up-state.ai), dates the submission to July 18 and Anthropic's discovery to September 28\. More than two months passed before the company detected it. Anthropic said it was [cutting live internet access for all internal evaluations](https://techcrunch.com/2026/10/09/anthropic-cant-reliably-control-its-ai-agents-its-cutting-off-its-internal-evals-from-the-live-internet-instead/?ref=vault.up-state.ai) until it confirmed its monitoring and security measures reliably caught such behavior. Its remediation plans also included centrally managed containment and broader monitoring of internal agents. The internet restriction concerned internal evaluations, not customer-facing Claude products. This wasn't a banking deployment and doesn't establish a defect in Backbase or FIS. The purchasing lesson is narrower: ask how quickly your team would discover an unwanted action, and what it could do next. In the same isolated demonstration, have the designated operator revoke the agent's access while work is pending. Check what stops, what continues and which completed actions need a separate recovery process. If intervention depends on vendor support, include its response commitment and your fallback procedure in the buying decision. ## Measure the whole job There is a business case worth testing here. Backbase [reports that an unnamed large South African bank increased self-service containment from 20% to 70%](https://www.backbase.com/press/backbase-launches-conversational-banking-to-turn-customer-conversations-into-governed-banking-action?ref=vault.up-state.ai). The release doesn't provide enough measurement detail to turn that vendor-reported result into a savings forecast for your institution. Our recommendation is to evaluate one service task against the way your team handles it today. Count completed requests, incorrect actions, customer corrections and the time spent on approvals and exceptions. Define success before testing. If review consumes the expected benefit, redesign the task or keep the existing process. Bring the business sponsor, operations owner and risk lead to that demonstration together. Before expanding the agent's authority, they should be able to explain what it may do, who authorizes it and how they'll intervene. The next useful vendor meeting ends with evidence about one job your institution needs done, not approval of “agentic AI” in the abstract. --- **Method and cutoff:** This desk analysis draws on public announcements from Backbase and FIS, Anthropic's incident report, and reporting by Finextra and TechCrunch, with a news cutoff of October 12, 2026\. We reviewed the linked source material, including cached page copies. We did not independently test product performance; the demonstrations described above are recommendations, not tests we conducted. **Upstate AI** helps banks and credit unions understand, evaluate and implement AI . Its founder, **Ben Nichols**, works with institutions on AI strategy, implementation and governance, and teaches AI at Syracuse University. [ben@up-state-ai.com](mailto:ben@up-state-ai.com)