Banking Sector - The “6-Second View” of AI Agents in Banking: Part 1 of 4

Introduction

In this four-part series, we’re going to examine the current shift towards AI agents through a 20-year lens of practical experience and a bit of personal nostalgia. The example I’ll use to help us was the most fun and enduring application that I ever wrote while working as a software engineer at a retail bank. We called it, “6-Second View”. This was the vision of then CIO, Brian McMullan. Brian made a lasting impact on my life, and remains a good friend to this day.

6-Second View was ambitious for its time. It integrated multiple “systems of record” across all points of contact to serve up a concise customer profile to every customer-facing employee. It was a valiant effort to “Know Your Customer”. Everyone had an appropriate level of access based on their role and all applicable regs applied. By exploring how an intranet application like 6-Second View was displaced by later core system add-ons, third-party SaaS products, and how we might imagine it today as an “agentic” app, you’ll learn why AI agents:

• Help software vendors preserve their core value while reimagining its delivery.
• Work with employees and each other to lighten the burden of repetitive tasks.
• Delight customers and members with better service and overall experience.

The series title is, admittedly, a bit of a bait-and-switch. It will, in fact, take seven seconds to read it.

From ATMs to AI Agents

When I first met Brian, I knew precious little about banking. I liked money, but had no idea how it was managed in a banking context. Like many, I thought banks and credit unions were government agencies — shrouded in mystery. Bank processes lived inside of that big, suffocating vault. I would have assumed that financial institutions had everything figured-out and would benefit little from someone like me. Over the next four years of working at that bank — and especially later as the technical co-founder of a fintech concern — I would receive an education to the contrary. For their similarities and for their differences, neither retail banks, nor credit unions operate like government agencies. But like any agency, they’re made up of good people — just doing the best they can with ever-increasing complexity. FIs are actually quite dynamic and continue to encounter unique challenges as both technology and consumer expectations grow. That proved to be a lot of fun, then, and it is still exciting to think about today.

The first challenge handed to me in 2007 was to find relevant data in a core system of roughly 3,000 tables, with limited documentation, and to then integrate with other apps across a fragmented landscape of vendors. From a teller system running on MS Access with nightly ports to SQL Server, to an early example of SaaS-based CRM, “6-Second View” had to quickly capture and convey it all. Identity verification, balances, warnings and alerts, next best offer, recent call center activity, and the like would need to be surfaced with zero friction. Change would come quickly, too, as the landscape expanded to include mobile banking, eStatements, behavioral rewards, and more. Because of its usefulness and versatility, 6-Second View would endure for roughly a decade until a core conversion to FIS made it unnecessary. Before it was decommissioned, I maintained it as a consultant long after leaving the bank. In a way, taking 6-Second View offline felt like saying goodbye to an old friend.

To revisit and reimagine “6-Second View” as it might look with today’s emphasis on agentic — AI agent driven — design, we need to set the stage a bit. Please bear with me for just six more seconds...

The Time Before RESTful APIs, SaaS, and Cloud Infrastructure

AI agents would have hated the early-to-mid 2000s, when a typical community bank or credit union’s technology stack was almost entirely on-premise, thick-client, and point-to-point. A Windows desktop teller system and deposit platform were connected over leased lines to a proprietary core that lived in the data-center, the basement, or maybe even a closet. Separate servers ran ingress for Internet Banking, IVR, imaging, and loan origination. Integration often meant screen-scraping or exporting nightly flat-files through batch processes. This meant that each channel kept its own copy of customer (or member) data. New features for many systems could require months of vendor professional-services, hardware procurement, and training. Clunky XML-based interfaces existed, but they were limited, expensive, or “unsponsored”. This reinforced vendor lock-in and slowed any attempt at a unified customer view. Direct access to the core was possible with ODBC or OLE drivers, but this was generally frowned upon. The difference between read-only and live access to the core was a single connection string parameter. A misconfiguration could spell disaster and certainly raise auditor eyebrows. This would have proved to be a rather miserable time for AI agents had they emerged sooner. If they existed in 2007, and if they had shoulders with which to shrug in disillusionment, they would have done so and found a cave to live in for the next 20 years.

The following table hints at what systems generally looked like for FIs when I started work on 6-Second View in late-2007.

Advancements & Disruptive Innovation in the 2010s

In the early 2010s, the industry pivoted further toward API-first implementations and SaaS deliveries. This inched us closer to being able to take advantage of something like the agents emerging today. Core providers exposed RESTful endpoints, fintech startups offered cloud-hosted point solutions, and banks began to assemble best-of-breed ecosystems through integration hubs rather than forklift upgrades. There was a lot of “disruptive innovation” happening. Digital banking, card processing, CRM and fraud analytics increasingly ran as multitenant SaaS, with secure APIs or iPaaS connectors feeding data back to on-premise or hosted core systems in near-real-time. APIs became a superior method for facilitating omnichannel experiences, and they framed-up at least one critical support mechanism for agents today, which we’ll describe later as the “tools” that agents use. The ability to integrate applications more seamlessly at scale began to shrink development cycles and shift the focus from infrastructure management to customer experience.

Continued Progress in the 2020s

As we entered the 2020s, core providers like FIS, FiServ, and Jack Henry were unbundling long-standing cores and redeploying them as micro-services on public cloud infrastructure. According to Jack Henry’s 2024 research, 43% of U.S. banks and credit unions now have a public, cloud-native core in their strategic roadmap, up from 33% the year before. Most independent SaaS providers requiring real-time access to the core achieve it through APIs provided by the core vendor, and those who don’t may still receive scheduled data dumps through sponsored transports.

Does this mean that we’ve come to a single source of truth in final form? That might be a stretch, but with specialized data and functions more commonly exposed through documented APIs — whether available from core system providers or independent SaaS vendors — the industry may be well-positioned to reimagine a tech stack with AI agents. That is intriguing to ponder considering that a full service institution might now have to integrate 25-plus distinct systems of record to give staff the same 360° picture that 6-Second View did less than two decades ago.

The table below helps us to think about the types of systems and integrations that may be involved in 2025 as we reimagine our 6-Second View.

Understanding the Watershed Moment We’re In Today

To move us through the remainder of 2025 and beyond it, we need to help decision-makers understand just what’s happened. We need to answer the question of why this development is so earthshaking. We need to describe the anatomy of an AI agent and how it differs from what we’ve known for decades. We must arrive at, and acknowledge, a mind-blowing truth…

Agents Are People Too…

Okay, no. They’re not people. But if you give me just six more seconds, then I’ll tell you why I chose the heading. The easiest way to think about AI agents is to think of them as people. Doing so is essential to understanding the paradigm shift from the software you’ve always known and deployed to these “things” we call agents. All of the systems we’ve mentioned so far require specialized rules — implemented as code. Code has traditionally been written by human beings, but even that’s changing now… I can assure you that agents have become excellent code writers. The requirement for executable application code alone does not distinguish a “traditional” system from an “agentic” system. This is because traditional software has always required a very real human-being to interact with it through some interface to get work done — be it to get insights, perform a task, or schedule a predefined task that happens later. Rarely is business software intended for a single user alone. It’s often intended for multiple users working in a team. That team of people may choose to communicate face-to-face, via phone, email, SMS, video conferencing, or other in-app messaging to coordinate efforts, make the right decisions, and to get the job done. Communication may be in words, intonation, idioms, sarcasm, jokes, and even through facial expressions. People operating collectively to get work done usually have designated roles and separate responsibilities. Likewise, there is always a superior to report to — whether it's a manager, a VP, an executive, the Board, or a regulatory agency. Agents work in this same manner because they can “think”. They do this in both similar and dissimilar ways to us humans. Agents interpret and express thought through “natural language”, but can interpret other forms of sensory input including what they see. They can distinguish human “sentiment” in what they read, hear, or visualize. They depend on information to make decisions like we do. They can toss out what they think is bad information. They can use tools to act on good information. They rely on each other's specialized abilities and authority to ensure that their decisions and actions are viable. This helps them to achieve their individual or collective goal. And, agents that are purposed to do different types of work will have different requirements for aptitude and experience level — just like human employees do.

“But how can this be possible?”

If you’ve used ChatGPT, Gemini, or Claude, for example, then you’ve interacted with a Large Language Model (LLM). Large Language Models have become the brains that power most modern agents. When you asked a question and it responded intelligently, the LLM made an informed “decision” to do so. It did not use “code” as we’ve understood it to make that decision. It evaluated the context of the conversation. It considered its own form of experience. That experience is provided by a training dataset composed of basically the universe of everything digital up to that training date. It hasn’t necessarily “remembered” everything in that training dataset to respond to you, but rather it has “learned the pattern” for everything it's trained on, and makes predictions about how to respond. It does this astonishingly well.

“Will this replace me?”

In an agentic architecture diagram, agents may appear to replace people because the intelligence required to make decisions is built in, tools are made available to agents so that they can act on their decisions, and other agents are made available for questions and the delegation of tasks. While I can tell you this is an extraordinary thing to watch play out between agents — and a lot of fun to orchestrate — the likelihood of job displacement isn’t quite as dramatic as you might fear — not yet. AI, and even its anticipated leap to AGI, does not have to replace “you” or the “team” that you work with. It is, however, critical to see agents turning traditional Software as a Service (SaaS) on its ear. The era we’ve moved into promises Service (delivered) as a Software. In other words, the software has become the worker because it can make rational decisions on its own and act on them. So, the best way to think of agents is absolutely as people — giving you the ability to onboard intelligence as a utility, augment your team, and offload less desirable work. Thinking of agents like people is useful from a “systems perspective”. I’ll leave you to ponder the philosophical implications.

“But answering a question is a long way from getting human work done, isn’t it?”

Getting any kind of meaningful work done starts by asking and correctly answering the right questions. It's part of how we plan “what to do” and what “tools” we’ll need to use to do it. We ask questions, too, when conditions change and we’re faced with ambiguity... “What do I do now?” Maybe you asked one of the leading AI chatbots about a current event that predated its training data. Did it search the Internet before responding to you? If so, then it first used a tool. That’s significant for at least three reasons: (1) It made a decision on its own — reasoning that it might not be able to answer correctly otherwise. It didn’t use code to make that decision. There were no if/then/else conditions or decision trees that guided it to a tool choice. It may have been influenced by additional instructions, but if so, those instructions were stated in natural language — plain speak. (2) It was able to interact with another system and repository of knowledge to get the information it needed on its own. It did not need you to choose and type search phrases, hit the ENTER key, click a button, or filter out any irrelevant search results by reason. (3) To date, only living creatures with a notable level of intelligence have demonstrated tool use.

You can interact with an agent. An agent can interact with you. An agent can interact with other agents — asking them for answers, for second opinions, or to take some action. An agent can select from tools provided to it so that it can perform a task that helps it to achieve its goal. Agents choose tools that allow it to interact with its environment and often this means to interface with other systems. Tools commonly wrap traditional, coded functions and API methods, but they may even use additional AI models through helper functions. Tools can be custom made for agents or they may be discovered and integrated via support for emerging standards like Anthropic’s Model Context Protocol (MCP). In the not-so-distant future, agents may code, test, and deploy their own tools to achieve their goal. Agents can get confused, just like people. They can suffer from a certain kind of “information overload”. They can make mistakes, but when properly made aware of each other, they will seek out help and approval from other agents based on their defined roles. In other words, working together, agent outcomes are better — just like in human teams. This is why it's important to carefully separate agents by specialization. Doing so is just the latest expression of an age-old requirement we call “separation of concerns”. Delegation is as important for AI agents as it is for humans. We will increasingly see this facilitated by emerging standards like Google’s Agent-2-Agent protocol (A2A). Proper orchestration and oversight is as vital for agent teams as it is for human teams. And finally, human feedback and approvals are still very much a requirement, but with far less burdensome volume.

Wrapping Up

The more you think about agents “as people”, the better foundation you’ll have upon which to build something that can “serve people” — and the closer it will bring you to reimagining a traditional app like 6-Second View. In Part 2, we’ll begin with the extraordinary opportunity available to vendors.

Part 2: Vendors Preserve Value By Reimagining Delivery