Cognitive Process Automation is here, and it’s powered by “Agentic Workflows”

Agentic-Workflows are applicable to a variety of business process like sales, marketing, customer support, HR, or Finance. For now, let’s consider the Swiss (or any) banking sector, which faces a constant influx of new regulations covering everything from anti-money laundering (AML) and Know Your Customer (KYC) to data privacy and capital requirements. Traditionally, staying compliant means manual review, extensive impact assessments across countless internal policies, and painstaking coordination to implement changes — a notoriously slow and costly process.

This is where Agentic Workflows shine, enabling what we can call an “Agentic-Regulatory-Radar”. We will introduce the Agent as a technical concept later, for now, just imagine it as an autonomous AI-helper. Something like a ChatGPT Model that works on its own based on a set of pre-defined instructions without you having to chat with it:

1. Research-Agent: This agent continuously monitors the websites of Swiss regulatory bodies (e.g., FINMA, SNB) and public news sources. It generates a weekly summary of key regulatory changes.
2. Compliance-Agent: The weekly summary then goes to the Compliance-Agent. This agent scans the bank’s internal databases (directives, contracts, manuals) to identify all affected products, services, systems, or processes, producing an impact report.
3. Policy-Writer-Agent: Both the summary and impact report are sent to the Policy-Writer-Agent, which drafts proposed policy amendments by cross-referencing the regulatory summaries with the identified internal assets.
4. Human-in-the-Loop — Compliance Officer Review: A human Compliance Officer receives a weekly brief detailing the changes, impacted policies, and suggested amendments. They provide expert feedback and request refinements, leveraging their deep understanding of legal and business implications.
5. The task then goes back to the Policy-Writer-Agent who then incorporates this feedback and resends the draft to the human expert for final review.

This example clearly illustrates how repetitive, cognitive tasks — like reading, analyzing, decision-making, and writing — could be partially handed off to AI agents, significantly transforming efficiency in critical business functions.
 

At its simplest, an Agent is an AI, typically powered by a Large Language Model (LLM), that is equipped with tools. Imagine an LLM that can not only understand your request but also independently decide which tool it needs to use to fulfill that request.

These “tools” are crucial. They can be:

• Access to a database: To fetch specific information for evaluation or decision-making.
• An API: To interact with external systems, retrieve real-time data, or trigger actions.
• Another asset: Any resource required to gather information or perform an operation.

The agent uses its LLM “brain” to select and execute the right tool at the right time. For example, consider our Research-Agent:

• First, it uses a web-scraper tool to scan regulatory websites and news portals.
• Next, it feeds the results into a summarizer tool to generate a standardized weekly report.
• If no relevant news is found, it may retry the web-scraper.
• After a second unsuccessful scan, it can skip summarization and instead trigger the email_sender tool to notify the Compliance Officer that no regulatory changes were detected.

This illustrates how agents can make context-aware decisions, sequence their tool usage, and handle exceptions — without manual intervention.

Now, imagine multiple such agents, each an expert in its own domain, equipped with specialized tools, working together! This is Multi-Agent Collaboration.

These agents are:

• Specialized experts: Each focuses on a specific part of a larger problem.
• Equipped with relevant tools: They have access to tools/data to get the job done.
• Context-aware: They can interpret that data within their specific area of expertise.
• Autonomous: They have the agency to execute actions or initiate processes within their defined boundaries.

An Agent-Workflow defines the entire business process where this collaboration between agents takes place. It’s not just a collection of independent agents; it’s a carefully designed orchestration that outlines:

• Sequences: The specific order in which agents interact and contribute.
• Exceptions: How the system handles unexpected situations or errors.
• Unhappy Flows: Defined paths for when things don’t go as planned, ensuring resilience.
• Human Review Triggers: Crucially, conditions and events where human oversight is explicitly required, ensuring a seamless human-in-the-loop process for critical decisions.

This structured approach ensures that even the most complex processes can be automated intelligently, with critical checkpoints for quality and compliance. 

Agentic workflows represent a significant leap forward in automation, moving beyond rigid rules to intelligent, adaptive systems. By combining the power of large language models with specialized tools and human oversight, organizations can tackle complex challenges in ways previously unimaginable. The “Agentic-Regulatory-Radar” for Swiss banking compliance is just one example of how these workflows can drive efficiency, enhance accuracy, and ensure continuous adaptability in even the most dynamic environments. As this technology matures, we can expect to see agentic workflows becoming a cornerstone of modern, intelligent enterprise operations, freeing up human talent for higher-value, strategic initiatives. The future of work is collaborative, intelligent, and increasingly agentic.

The AI@NCA Brief wants to hear your voice. Whether you have a burning AI question or a topic you’d like us to cover, we’d love to hear from you. Share your thoughts in the comments or reach out directly to our team 

Follow Ramona Beck on LinkedIn

Follow Pascal Wyler on LinkedIn