By Eric Newcomer, CTO and Principal Analyst, Intellyx
Financial services organizations are developing agentic AI systems that autonomously and reliably execute human tasks, make decisions, and take actions proactively.
The problem is that the majority of agentic AI systems are not yet ready for production. Organizations are developing individual AI agents and multi-agent systems as initial steps, and trying out agentic AI for internal systems.
But the race is on, and agentic AI is rapidly becoming a competitive issue.
According to a recent article, for example, BNY currently uses AI agents as employees, giving them logins, email addresses, and managers for supervision.
A BNY agent can autonomously detect and fix issues in code, although a human still approves a change before it’s deployed.
In January, Citi announced that it’s working on agentic AI systems for the “do it for me” economy, and in July started deploying agentic systems internally for developers to automate software patches and upgrades.
Citi predicts applicability to a broad range of additional use cases, including personalizing new product offers to customers, virtual financial planners, automating routine operational tasks, real time risk profiling for loans, cashflow forecasting, customer onboarding, and fraud detection.
Agentic AI systems have tremendous potential to reshape the financial services landscape, but challenges remain in moving from current “human in the loop” systems to the fully autonomous systems that can reason and take independent action.
Relying on AI agents
As financial services firms deploy more and more AI agents, and build up the components of agentic systems, they increasingly need to ensure the agents are reliable and trustworthy.
A big challenge is the agent’s reliance on human language to interact with the large language models (LLMs) that power the generative AI agents use.
Interacting with computers using human language greatly simplifies the way in which people interact with computers, reducing time for research, analysis, and reporting. But using human language also introduces new risk and uncertainty.
Human language is ambiguous and subject to interpretation, which can produce incorrect results and hallucinations. Until these issues are resolved for agents, a human has to be in the loop to confirm the results.
Defining a set of service level agreements (SLAs) that measure agent performance is a good way to build trust – not only for typical latency, reliability, and availability goals, but also for the new gen AI challenges of accuracy and safety.
Breaking up a workflow or business process into a series of discrete steps also helps resolve the issues by reducing the scope of an agent’s interaction with an LLM, focusing the evaluation process on fixing issues within that scope.
Multi-agent systems
Multi-agent systems perform complex tasks, such as onboarding a new customer, processing a payment, or responding to an incident or outage.
Multi-agent systems communicate with each other to complete a workflow or business process. Individual agents reason and make decisions based on available data, but they still do not take independent action. Humans are typically still in the loop.
Multi-agent systems communicate among agents in various patterns (such as peer to peer and hierarchical), share data among agents, and use a persistent memory system to store the prompts and results from multiple conversations with LLMs.
Such agents typically need to access organizational tools such as ServiceNow, JIRA, or Splunk, and access internal data for training on regulatory compliance, preventing data leakage, flagging identity spoofing, or detecting fraud.
Access to external data sources and tools can be defined using the Model Context Protocol (MCP).Communications between agents can be defined using the Agent to Agent (A2A) protocol.
Agentic AI systems
Simply put, an agentic AI system automatically executes a series of steps, including reasoning and adaptability, to take action to achieve a specific business goal. An agentic system can remove the “human in the loop” or significantly reduce human interaction.
Starting with individual agents and building up multi agent systems are great ways to prepare the organization to achieve the final, critical goal of adopting agentic AI systems.
Agentic AI systems actively learn, adapt, and autonomously take action, based on what they learn and decide. Agentic systems help financial services organizations become more cost efficient and deliver more personalized customer experiences.
In contrast to AI agents that follow predefined workflows or respond to user prompts, agentic systems are proactive, goal-driven, and capable of taking coordinated actions across tools, data sources, within defined constraints.
Organizations typically deploy agentic systems initially for internal activities, such as scheduling meetings, booking business travel, or deploying application patches.
The Akka Agentic Platform
Akka, used by dozens of financial institutions as the backbone to business-critical distributed systems, provides an agentic AI platform to develop, deploy, operate, and evaluate AI agents individually, in workflows, and ultimately in agentic systems.
Capital One, Icon Solutions, Judopay and dozens of other banks depend on Akka technology for payment systems, loan management, and more.
Akka’s compute model reduces infrastructure cost and token processing charges. Their simplified developer experience increases the velocity of change.
Components of the Akka Agentic Platform include:
- Orchestration - define and control long-running systems
- Agents - create agents, MCP tools, and APIs
- Memory - durable, in-memory and sharded data for LLM conversations
- Streaming - high performance data streaming
The Akka Agentic Platform evaluates agents and agentic systems against a complex set of SLAs to build trust in gen AI -- including performance, reliability, availability, accuracy, and safety.
The Platform supports long-running, multi-agent processes, handling large volumes of shared memory, and real-time streaming data for ambient, adaptive, and edge AI use cases. It’s easy to get started with simple systems and evaluate and build toward more complex systems over time.
The Intellyx take
The potential impact of agentic AI on financial services organizations is immeasurable. The potential efficiencies and productivity gains are already obvious, as are the tremendous strategic business improvements.
Better customer experience is driving adoption as well, whether improving customer service automation or simplifying account opening and loan processing.
Akka has a long history of providing reliable, scalable distributed systems for enterprise applications. Akka brings this history and expertise to bear for agentic AI systems – solving challenges in development and operations, and most importantly, trust and safety.
Financial services organizations embarking on the agentic AI journey should definitely evaluate an AI agent development framework such as the Akka Agentic Platform.
Copyright © Intellyx B.V. Intellyx is editorially responsible for this document. No AI bots were used to write this content. At the time of writing, Akka is an Intellyx client. Image by ThisIsEngineering from Pexels.com.
