EuroFinance

Artificial intelligence has quickly become a priority for treasury teams. Budgets are being allocated, use cases identified and pilots launched. Yet, despite the momentum, many organisations remain stuck in experimentation, unable to translate early promise into scalable outcomes.

According to Shivank Rai and Suhail Shaikh, Senior technical product manager at a big tech company, the problem lies not in AI’s capability, but in how it is being approached.

Too often, teams begin with a solution rather than a problem. Under pressure to adopt AI, they look to vendors or tools that promise quick results. But without a clearly defined objective, those solutions rarely deliver. “Don’t start with a vendor or with a solution that you’ve predefined,” said Rai. “What you want to be starting with is what’s the problem you’re solving… define the decision, the outcome, the experience you’re trying to drive, and work backwards from there.”

This seems straightforward. In practice, it requires a shift in mindset. AI is not something that can simply be layered onto existing processes, it forces organisations to rethink how those processes work.

Reinforce the foundations first

If the first challenge is defining the problem, the next is less appealing: data.

For many treasury teams, data is fragmented, inconsistent, and relies on old tools. In that context, expectations of AI can become unrealistic. As Shaikh noted, before thinking about AI, organisations need to ask a more basic question: are their data ready?

“Before you jump into the AI bandwagon… you need to first understand… Are you data ready?” he said.

That readiness, he explained, involves knowing what data exists and what is missing, making it accessible for systems, and enriching it with context. Without this, AI may produce incomplete and misleading outputs. “If you have only 60% of your data… whatever responses AI is providing is biased to that 60%,” he added.

This limitation is particularly visible in cash flow forecasting, one of treasury’s most sought-after AI use cases. The difficulty is not simply about improving model accuracy, but about reframing the problem itself.

“Cash flow forecasting at scale is fundamentally a signal detection problem,” said Rai. “You’re not predicting the future from history. You’re aggregating leading indicators that exist right now.”

In this view, AI’s role shifts from predicting outcomes to identifying patterns and signals in real time, something it is far better suited to do.

From tools to systems

That shift in thinking extends to how AI solutions are designed. The idea of a single system capable of solving multiple treasury challenges is appealing, particularly for teams with limited resources. But in practice, it rarely works, Rai and Shaikh added.

Shaikh noted that general systems produce broad answers, often falling short of the level of accuracy the treasury requires.

More effective approaches rely on multiple specialised agents, each responsible for a specific task, whether that is gathering data, analysing it or generating insights. These agents operate together, with their outputs combined into a final recommendation. What appears to the user as a single system is, in reality, a coordinated set of processes working behind the scenes, according to Shaikh

Even then, trust remains a central concern. Treasury is a function where mistakes have immediate consequences, and reliance on automated outputs requires high confidence. To address this, organisations are building validation mechanisms into their systems. Shaikh described the use of “judge agents”, which assess outputs before they reach the user, compare them to past benchmarks and correct them where necessary. Over time, these systems improve through feedback, becoming more accurate with each iteration.

Despite these layers of automation, humans will make the final decision. “The chain ends with a human making a decision faster and better than they could ever before,” Rai said.

This human-in-the-loop model is not just a safeguard—it is essential for governance. In treasury, where accountability and auditability are critical, decisions must be explainable, with a clear record of the data, signals and reasoning behind them.

For organisations looking to move beyond pilots, the lesson is practical. Many AI projects fail not because the technology is lacking, but because the foundations are weak. “The model usually isn’t the problem,” Rai noted. “It’s the data layer that wasn’t input to the model.”

The most effective approach, therefore, is to start small—focusing on well-defined, repeatable use cases where data is available and outcomes can be validated. From there, teams can build confidence and gradually expand.

AI may transform the treasury. But as Rai and Shaikh suggest, that transformation will not come from technology alone. It will depend on how carefully and how patiently organisations build around it.