Banks and insurance companies were quick to recognize the potential of generative AI. By the middle of last year more than three out of four had implemented AI projects and the vast majority plan to increase their investments.
And yet, despite FS firms already notable accomplishments, agentic AI is about to create a wave of innovation that will transform the industry.
As I mentioned previously, AI agents operate within an ecosystem—the agentic AI architecture—comprising of an orchestrator, super agent(s) and multiple utility agents. Each has a specific role in the digital team with the ability to reason logically, plan, decide, act and communicate in pursuit of a goal. Agents can access multiple large language models (LLMs) and other data sources, and understand and reflect on both structured and unstructured information. Their long-term memory, together with feedback loops, allow agents to learn and continuously enhance their performance.
While it may work autonomously, agentic AI is designed for close collaboration. Like bees in a hive, agents interact with each other—with oversight by higher-level agents or, ultimately, human beings—to manage complex workflows. Agents are also adaptable: skills can change or advance as required and capabilities can be reused repeatedly. With multi-system agent interoperability, like Accenture introduced with Trusted Agent Huddle, agents can seamlessly work together in a single platform, allowing organizations to select the right agents for the need.
The benefits are significant. Agentic architectures can increase the speed, thoroughness and accuracy of most FS processes. In banking, agents could be highly applicable for use cases like supporting relationship managers and handling routine customer enquiries to reviewing credit applications, providing personalized financial advisory and detecting and preventing fraud.
A leading retail bank recently used agentic AI to accelerate a large data migration and legacy technology replacement program. It deployed an agentic architecture—an ecosystem of AI agents—to support its developers, including a software development agent to write the new code and several different critique agents to review the code quality and provide feedback. The agents loop until an acceptable level of quality is reached, with a human software developer following along and capable of directing the agents. The early results have been staggering, with 30% efficiency on development, millions in cost savings and increased code review frequency.
Key elements to get right
In an industry where trust can never be taken for granted, there are four attributes that every potential user of agentic AI should prioritize: explainability, maintainability, data maturity and security.
AI agents operate in more complex, dynamic environments than traditional AI. Their ability to figure out, on their own, how best to achieve their goals, means they could evolve in ways that are not obvious to their human handlers.
The future of agentic AI in financial services depends on banks and insurers retaining full control of their processes. They need to understand why their AI agents are making specific decisions and taking specific actions. They also need to be able to fix, update and improve their functionality. Both explainability and maintainability can be challenging when the systems are evolving autonomously.
Another element that can’t be overlooked is data, specifically the need for a modern data architecture and data maturity model. The data layer is a critical input as firms look to tailor these agents to the customer environment and the way they operate. Without the right level of maturity and continuous reinvention of data capabilities, businesses will struggle to move beyond using agents for isolated tasks.
Lastly, the agents have to be secure and act as intended. FS firms will have to develop identity access management and monitoring programs specifically for agents, with controls like an endpoint detection capability that can tip off security operations if an agent is doing something it’s not supposed to or if someone is accessing an agent who shouldn’t. This is vital as banks and others look to scale agents.
It should also be noted, however, that just because agentic AI can, doesn’t mean it always should. The focus should be on what’s really going to move the needle, finding areas along the value chain—whether it be operational efficiency or improving the client experience—to drive specific improvement, using agentic AI as an enabler.
From there, banks and insurers can develop agents that can be both reusable across multiple use cases and scalable. This element of reusability – creating utility agents that can be reused many times for different use cases and reinventions – is important to master.
How will employees and customers interact with agents?
Many workers already see value in working with AI, including the ability to automate routine and repetitive tasks, freeing them up to focus on higher-value activities that require human judgement and creativity.
These workers will play an important role in providing inputs and setting goals for the agents. They can also share feedback to help refine and improve the agents’ results and performance. In the coding example that I mentioned earlier, the bank deliberately involved its developers in the agent design process and rolled the agents out to its entire developer team. Colleagues should be empowered by their organizations to reinvent the work they know best.
Meanwhile there’s openness from bank customers to interact more with AI. Accenture’s 2025 Banking Consumer Study revealed that 62% of consumers would be willing to use an AI-powered financial assistant to help them manage their finances.
While many big techs are investing heavily in agentic AI, fintechs are also enabling these architectures. For example, most of the AI startups recently selected for the New York FinTech Innovation Lab’s 2025 class have an agentic AI focus:
- Lyzr has an agent infrastructure platform that automates entire job functions as well as workflows.
- Multimodal’s AI agents are trained on company data, making them not only accurate and efficient but less prone to hallucinations.
- DeepSee’s flagship workflow automation application, DeepRecon, is regularly audited and continuously monitored to ensure transparency, data security, accuracy and resilience.
The future is bright
Agentic AI’s ability to enhance scalable, always-on intelligence across the enterprise holds the promise of new benchmarks for operational performance, even when it comes to the most complicated processes. It could help simplify costly areas like risk and compliance, going further than traditional AI to not only find and analyze data, but execute follow-up workflows, make recommendations and draft suspicious activity reports.
However, there’s much careful work to be done before the industry can reach these milestones. Just as importantly, firms seeking to capitalize on the technology must recognize that they cannot simply layer AI on top of their existing systems, workflows and data. Instead, they need to fundamentally rethink their processes before automating them and lay out their data in a different way. These are complex software deployments and need to be treated and maintained as such.
If FS firms get the foundation blocks right and take the caveats seriously, agentic AI could indeed be a game-changer. The reimagining of key business processes, executed by AI agents and integrated with human oversight, can generate transformative results.
Read the full article here
