Many of our clients are facing the challenge of where to start and how to scale with generative AI (gen AI). One promising concept is the use of intelligent agents, which automate and optimize processes, making gen AI implementation and scaling more manageable.

In simple terms, an intelligent agent is software that acts in a given environment to autonomously deliver a specific goal. Such agents are therefore supposed to do more than just automate routine tasks: ultimately, they are intended to be able to choreograph entire business workflows, learn from their past performance, acquire further knowledge, and then self-improve.

This blog explores a practical use case for leveraging such agents in middle and back office operations of investment banking.

Challenges to automating exception handling in Financial Services

Financial Services is a regulated industry. As a result, the middle and back offices of trading firms must ensure that all transactions are accurately recorded, reported, and stored. They also need to safeguard that all activities are conducted within regulatory guidelines, and that operational risks are appropriately managed.

However, even in areas that traditionally have high straight-through-processing rates in operations, there is still a considerable amount of manual effort required to resolve exceptions when they occur. Trying to automate such exception handling is not trivial and requires very specific capabilities.

To understand why, let’s examine the three most common barriers to automating exception handling:

• These activities often span multiple legacy systems and rely on often siloed data and processes within different business functions. This forces operations analysts to switch between numerous systems.
• The specific tasks required to resolve exceptions are often deeply ingrained in the knowledge of operations analysts, making automation efforts difficult, complex, and costly.
• There is usually a high volume of requests and clarifications that need to be sent to the front office and to counterparties, typically via email.

The question we have been asking ourselves is, can intelligent agents help to further automate the handling of such exceptions—in a controlled and auditable manner—to drive new efficiencies and reduce the overall costs across the trade lifecycle?

Leveraging work orchestration to steer intelligent agents

While the vision of intelligent agents is to one day choreograph all the steps of a business outcome, today the tasks performed by intelligent agents still need some careful steering—especially if processes require a clear audit trail. In other words, there is a need to set boundaries of what an agent should do and what data it can leverage. This is where work orchestration solutions come in: embedding the respective agents in a well-defined, end-to-end orchestrated process is currently essential.

For Accenture, this goes beyond theoretical discussion. We operate a capability that provides back-office operations as a service, so we frequently grapple with the same questions our clients face. In collaboration with ServiceNow and DeepSee, we have recently created an intelligent work orchestration platform that aims to deliver operational efficiencies by also leveraging intelligent agents.

The platform is able to create a digital twin of a business process; it can coordinate task allocation to agents; and it provides a library of out-of-the-box agents to perform specific tasks within a business domain (for example, automatic email categorization, routing, and drafting).

How to fix a Standard Settlement Instruction Mismatch automatically

Here is an example of how this platform realizes the power of intelligent agents: a common exception in the cash securities settlement process that requires manual intervention is a Standard Settlement Instruction (SSI) Mismatch that prevents a transaction from being closed in an automated way.

Traditionally, an operations analyst would receive an exception from the core trading platform and would need to perform several tasks to identify the root cause, determine the next best action, agree on the resolution with the counterparty—usually over email—and then resolve the issue back in the core trading platform.

Within our platform, a set of agents, using Large Language Model (LLM) powered Natural Language Processing (NLP) and Machine Learning (ML), perform those tasks automatically: if an SSI Mismatch event occurs from the core trading platform, those agents work together to solve the issue.

• The platform tasks the fails predictor agent to determine the likely root cause of the mismatch based on historical behaviors, i.e., is it a bank issue or a counterparty issue?
• It tasks the SSI recommender agent to find the correct SSI.
• The platform sends a templated email to the counterparty requesting them to update the instruction to the correct SSIs (along with the recommended SSI).
• If required, an email agent will send a reminder email to the counterparty.
• An email watchtower agent reviews all emails, categorizes them using the email categorizer agent (and if required uses an email structuring agent to interpret any unstructured attachments).
• An SLA watchtower agent reviews the trade status and confirms no further action is required when the deal is completed.

The work on this platform leads me to conclude that before firms can start to deploy intelligent agents today, they would need to spend time looking at the process definitions and encoding tasks at the keystroke level into a digital twin of a business process. In my opinion, using intelligent agents controlled by a fully orchestrated business process embedded in responsible AI principles is the most promising way to achieve the efficiencies needed to significantly reduce the costs of a trading firm’s middle and back office with gen AI and enable human talent to focus on complex problem-solving and strategic decision-making.

Interested to learn more or talk about our practical experience? Feel free to reach out to me on LinkedIn and let’s have a chat.