Work in Treasury revolves around communication, with reading, writing, discussion, and analysis serving as the primary tools. Information is exchanged through various formats—spreadsheets, documents, and meetings—each playing a vital role in how we collaborate and deliver results.
Given this, AI appears to be a promising tool for us, as it is inherently multimodal (capable of processing and generating text, images, and other types of data) and can handle the varied nature of our tasks. However, while AI excels in certain areas, it also has notable shortcomings, particularly in the context of treasury workflows. Ethan Mollick referred to this phenomenon as the “jagged edge” (the uneven performance of AI, excelling in some areas while falling short in others).
To understand AI's role in treasury, it’s important first to recognise where AI truly excels. For instance, AI can process lengthy documents and distil them into concise, useful formats, such as summarising key points from regulatory papers or converting them into memos for non-experts.
Additionally, AI is skilled at analysing structured data and generating insights, often without the need for data downloads—simply capturing an image allows AI to interpret it.
Similarly, you can take a snapshot of a risk report, have AI explain it, and even tailor the explanation for a specific audience, all within a short time. Fact-finding is another area where AI helps; tools like Perplexity (an AI search tool that provides concise answers to complex queries) can sift through web content and provide summaries that directly address your questions. Moreover, AI can summarise meetings, producing transcripts or key documents that capture what’s discussed and agreed upon.
It can even conduct virtual surveys, acting as a customer and providing feedback on your product—something that traditionally would require hours and significant expense. These capabilities illustrate how quickly AI has advanced in a short space of time. However, it’s crucial to acknowledge AI’s limitations.
One noticeable issue is inconsistency in its responses—you might ask the same question multiple times and receive different answers. This variability can be both advantageous and frustrating, especially when you’re relying on AI to generate scenarios based on its training data. Often, these scenarios result in familiar approaches rather than novel ideas. To unlock more creative outputs, you need to prompt the AI repeatedly, which requires time and patience.
In addition to scenario generation, AI can struggle in other areas. Despite recent improvements in AI’s ability to handle basic tasks in Excel, the process remains cumbersome and error-prone. Consequently, relying on AI to build Excel models isn’t practical yet, which is disappointing given how crucial spreadsheets are in treasury work. (AI is said to be much better at coding, something I don’t have the skills to judge).
Reflecting on my experience with AI over the past 18 months, it’s been a mix of trial and error. Sometimes I stumble upon surprisingly effective solutions; other times, I spend hours trying to get AI to perform a task, only to be met with failure.
The challenge in treasury is that most professionals lack the time or inclination to experiment extensively with AI, given the demanding nature of their roles. Nevertheless, this hasn’t deterred many from using AI at work even when it’s not officially available. There is considerable evidence that employees use AI privately to tackle work-related challenges. This trend raises concerns about sensitive information being uploaded without adequate oversight. It also leads to less internal discussion and collective learning about AI’s capabilities and limitations—a missed opportunity.
Given these dynamics, how should AI be introduced in treasury?
Simply providing employees with access to tools like ChatGPT or Claude and expecting them to figure it out on their own isn’t sufficient. While some might successfully adopt AI this way, the process is generally slow, laborious, and often doesn’t yield the best results. To fully benefit from AI, users need to explore its capabilities thoroughly, which requires a more structured approach.
A practical starting point is to experiment with pilot use cases (small, controlled experiments) that identify straightforward and time-consuming tasks—prime candidates for AI assistance. Analysing your workflow, particularly repetitive tasks that AI could handle, is key. For example, writing, idea generation, and creating strategic documentation are areas where AI could make a significant impact. Collaborating with colleagues on these pilot projects allows for the comparison of results and helps build a solid set of use cases to understand how AI can best support your work.
However, before these pilot projects, it’s important to ensure that your team is comfortable with using AI. As noted earlier, experimentation is necessary, but not everyone will feel at ease with it right away. Therefore, training is not only helpful but necessary to build confidence. The goal is to foster a mindset where AI is seen as a helpful co-worker, albeit one whose output can range from brilliant to average. I've found many attempts with AI may not work, but some will, and these will offer huge returns. It’s much like the 80/20 rule (a principle suggesting that 80% of outcomes come from 20% of efforts).
Encouraging this trial-and-error approach and sharing insights across the team will help build a strong case for AI adoption in treasury.
Moreover, employees with limited treasury experience can achieve tasks that would typically be beyond their reach, as long as they understand that AI is an aid, not a replacement.
This brings us to the issue of hallucinations (AI generating incorrect or misleading information). While these were a significant concern a year ago, they are less frequent now, though they still occur. The problem is that AI can be a convincing source of misinformation, and only by carefully reviewing its output can you detect inaccuracies. Thus, the skills required for using AI include not just flexibility in experimentation and prompting but also thorough analysis to ensure accuracy, particularly in the area we work in.
Some might argue that it’s better to wait for AI to improve and become more integrated into existing products. However, I believe this would be a mistake. As AI evolves, delaying its adoption leaves you behind, risking obsolescence without realising it. The key is to think ahead—where do you see AI in three to five years?
Although predictions are challenging, there’s a strong case that AI will soon be able to handle complete workflows (end-to-end processes managed by AI), a concept referred to as “agency” (the ability to delegate tasks to AI, with humans overseeing the results).
We will need to rethink what it means to work in Treasury because much of what we currently do will be done by AI. In this world, advanced skills, experience, and understanding of how to work with AI will become increasingly important.
In the shorter term, the upcoming release of GPT-5 (the next version of OpenAI’s language model), anticipated in the next few months, could bring significant advancements. I plan to revisit some past use cases with the new model to see how things have progressed.
Comments