Agentic AI and Treasury: Lessons Learned from Years of Automation
Terece Tai, CPA, CA
8/11/20264 min read
Over the past three decades, treasury teams have experienced multiple waves of automation. Spreadsheets replaced manual calculations. ERP systems integrated business processes and banking platforms. For larger organizations, Treasury Management Systems (TMS) automated cash visibility, payments, and risk management workflows.
Today, Agentic AI promises another leap forward.
The productivity gains can be significant when AI is applied to the right problem. In many cases, AI can reduce hours of work to minutes.
A cleaner path to success is to learn from past mistakes of others.
Establishing the right goal is critical. The goal is not to maximize automation. The goal is to maximize return on automation
Protection over optimization
One of the first questions is often Which processes should be automated? Some processes are naturally suited for automation while others are not.
The best candidates are repetitive, homogeneous activities with predictable inputs and outcomes. Examples include reconciliations, recurring reporting, standard calculations, and routine data transfers.
However for treasury, the primary consideration should not be efficiency but risk exposure. Protection comes ahead of optimization
For example, fraudsters often target corporation’s payment processing employees by requesting wire info changes via hacked vendor’s or colleague’s email. This is one of the most common fraudulent scenarios where a lot of corporation falls victim. Can you imagine automating the vendor bank info change request process and have AI processed all incoming requests without a proper callback procedure? The time saving should not be shortening fraudster’s time to access your cash.
Focus on Return, Not Automation
Many organizations evaluate automation by asking whether a task can be automated.
A better question is whether the automation creates enough value to justify the effort required to build, support, and maintain it.
During one ERP implementation, I observed a proposal to automate a process that required roughly five minutes per day to perform manually.
At first glance, the idea appeared sensible.
After reviewing the proposed design, however, the automated process would actually require approximately ten minutes of daily intervention through additional button pushing, monitoring, and cross-checking.
The automation created more work, not less.
Had transaction volumes increased tenfold, the economics might have been different. But at the existing volume, the benefits simply did not justify the added complexity.
The lesson was simple: Good automation is not about replacing manual effort. It is about creating a meaningful return on the effort required to implement and support the solution. Quality over quantity. Speed should not compromise accuracy, diligence and sound judgment.
Learn to Walk Before You Run
Automation can be compared to upgrading from walking to running.
Running is faster, but only if you know where you are going.
For instance, when I used to oversee credit risk and accounts receivable collections for an Italian subsidiary, the amount of overdue receivables was significant, which is not uncommon in Italy. Many people believed automating payment reminders and collection emails would accelerate collections. After further investigation on why invoices remained unpaid, the issue was not a lack of collection follow-up.
The underlying problem was operational and originated from order administration.
Many disputes were caused by mismatches between sales orders, customer-requested order changes, delivered quantities, invoiced amounts, customer acceptance records, and carrier’s approval to invoice. A small misalignment would result in payment being held.
Sending more reminder emails would not solve those issues. The fix was indeed tightening the matching of sales order, order changes, delivered items and invoices.
The collection problem originated from upstream challenges, not from insufficient communication downstream.
This is a lesson that applies equally to AI initiatives today. Technology itself rarely fixes a broken business process.
AI Readiness Matters
The success of an AI initiative depends heavily on the readiness of the organization implementing it.
Companies with mature processes, reliable data, and clearly defined workflows are generally well positioned to benefit from AI.
By contrast, organizations with evolving processes, incomplete data, or unclear responsibilities often struggle to achieve sustainable results. In those environments, the risk of AI hallucination snowballing can increase and cause errors to multiply exponentially.
Before asking, "How can we use AI?" organizations should first ask:
"Are our processes and data ready for AI?" For treasury, a key question is whether a new agentic AI flow would expose the company to higher risks of fraud or errors. If the risks are higher than what the company can tolerate, then additional controls must be established to mitigate the risks to an acceptable level.
Where Human Judgment Still Matters
Large Language Models (LLMs) perform exceptionally well when dealing with common scenarios.
Treasury management rarely operates in a standardized environment.
Every company has its own legal structure, banking relationships, foreign exchange exposures, risk tolerance, commercial realities, and internal control environment. For MNCs, the complexity would increase with cross-border transactions involving multiple jurisdictions and tax implications.
AI can provide useful insights and recommendations, but determining whether those recommendations are appropriate for a specific organization still requires experience, context, and professional judgment developed through years of practical application. Common sense is a control.
Final Thoughts
Agentic AI will undoubtedly become an important tool for treasury teams.
But the organizations that benefit most will not be those that automate everything. They will be the ones that understand their processes, maintain quality data, and apply technology where it genuinely adds value.
Key takeaways
✅ Not every process should be automated.
✅Technology rarely fixes a broken business process.
✅Protection should come before optimization.
✅Common sense is still a control.
✅Successful AI adoption is often more about business readiness than technology.
✅LLMs excel in handling common scenarios, but not niche expertise.
This is where specialist advisory firms continue to create value.
Every organization has unique business processes, control environments, banking relationships, risk exposures, legal structures, and strategic objectives. While AI can provide useful insights and recommendations, applying those recommendations appropriately still requires practical experience and judgment.
At FX-cient, we believe successful AI adoption is not a technology challenge but a business readiness challenge. Organizations achieve the greatest return from AI when processes, controls, and data are properly aligned. Our treasury, foreign exchange, and internal controls expertise helps clients assess and strengthen that readiness before investing in automation.
Disclaimer: The information and tips provided in this article are for general informational and educational purposes only. While we strive for accuracy, the content should not be taken as professional financial advice. Ultimately, your own research and personal judgment should drive your financial decisions. In no event shall FX-cient be held liable for any financial losses, damages, or decisions made based on the use of these tips.
Disclaimer: This article offers general guidance only. Financial decisions should ultimately be driven by the reader's own judgment. FX-cient accepts no liability for any actions taken or outcomes resulting from the use of the content.
© 2026. All rights reserved.