EuroFinance

Treasurers are increasingly under pressure to manage cash and risk in real time as companies contend with volatile markets, complex global operations and rising financing costs. For many multinationals, that has exposed the limits of treasury processes reliant on spreadsheets.

At ASML, the Dutch semiconductors supplier, whose lithography machines are critical to advanced semiconductor production, treasury had to manage significant foreign exchange exposure.

“Our hedging programme was about 70% efficient,” said Jason Koeller, head of US treasury for ASML. “And it was also very manual, asking people to consolidate multiple forecasts, figure out positions before you hedge. Very spreadsheet-heavy, time-consuming, and prone to error.”

At Cloud Software Group, formed from the merger of Citrix and TIBCO in the year 2022, treasury teams faced a different challenge: gaining visibility over cash positions across the business. The company relied on employees across multiple regions to manually download bank statements, update spreadsheets and categorise transactions.

“It was very clear it was not going to be a sustainable process going forward,” said Robert Zavertnik, treasury manager. “Starting on Friday night, sending out all those spreadsheets, getting them back Monday, everyone updating — we would finish halfway through Tuesday.”

Laying the foundation

Both treasurers shared a conviction that had become a principle in modern cash management: before doing anything intelligent with data, you first have to find it.

For Zavertnik, the answer was to partner with treasury management platform to automate the entire data pipeline — bank activity, balances, and transactions — and classify each item by cash flow type. The process took two to three months to connect banks and onboard data, followed by another period of building classification rules. The payoff was immediate. A cash positioning statement that once consumed most of three working days could be produced in ten or fifteen minutes.

“We have this rich data set of all of our cash balances and cash flows,” Zavertnik noted. “In the future, we can leverage it for more advanced analytics — machine learning algorithms, AI agents, chatbots. It’s pretty exciting.”

The discipline of getting the data right also surfaced an unexpected organisational challenge. Because Cloud Software Group’s treasury team managed cash without originating all of it — accounts payable, payroll, and tax teams all touched disbursements — there was no standardised system for labelling outbound payments. Tagging transactions on the back end proved very complex.

Zavertnik’s solution was to push the classification upstream. He proposed to the CFO a standardised payment tagging protocol requiring every team that originated payments to embed category codes at source. The result: approximately 80% of the company’s disbursements arrived pre-labelled. “We were tagging it on the back end instead of on the front end,” he said. “Once we had the executive sponsorship, we made a lot of progress.”

Teaching the algorithm

At ASML, the foundation was not a single system but five years of actuals. ASML’s cross-functional team, working with an internal digital and transformation finance group, tested twenty open-source Python algorithms against that history, evaluating each for capturing trend, seasonality, and cyclical patterns in US dollar outflows. The model was progressively refined — first to two years of data, then further calibrated by comparing backward projections against known actuals.

The result was a forecasting model that had 96% accuracy for USD exposure — a figure that translated directly to economic outcomes.

“That’s $25 to $50 million of exposure we’re taking off the table,” Koeller said. “And that is real economic cost or gain.”

The model was not static. It retrained monthly, and a designated team member reviewed performance on that cadence to ensure the output remained within expected parameters. Having established the methodology for dollar flows, ASML extended it to the Japanese yen, with further currency pairs in progress. Koeller was candid that each extension required its own calibration. “Yen might not have as many suppliers as US dollar,” he noted. “You need to understand how the algorithm is working and continue to tweak it.”

A more recent addition to ASML’s toolkit showed how fast the frontier was moving. The treasury team built an agent that ingested counterparty emails submitted in response to deposit rate requests, extracted the rates automatically, and populated a structured table for comparison — a task previously done manually, email by email.

“That could save us time, and it gives us a more historical look at who’s coming to play,” Koeller said. “Who’s serious and who’s not.”

The consensus problem

Neither of these projects was technically straightforward. But both treasurers were emphatic that the technical obstacles were not the hardest ones.

At ASML, building internal consensus — particularly in a culture that, as Koeller noted, was characterised by Dutch organisational norms in which anyone could raise a concern — took longer than building the model itself. Accounting teams and executive finance leadership had legitimate questions about why, and how, the basis for hedging decisions was being changed. “You can’t just snap your fingers,” Koeller said. “It’s definitely consensus-building. And also getting people up to speed on what this is doing — when you start talking about Python algorithms, that’s usually not the nomenclature of treasury.”

At Cloud Software Group, it was the combination of limited resources  and institutional caution. Zavertnik’s team was small; the day job of funding operations left limited room to step back and evaluate new approaches. The IT security function, meanwhile, was wary of cutting-edge models and open-source agent tools following past data breaches — a posture Zavertnik regarded as reasonable rather than obstructive. The company used an enterprise-deployed instance of Google Gemini as its sanctioned large language model. “If our IT security people are saying I can put company data in there, then I’m comfortable with it,” he said. He had already found it useful for summarising lengthy bank entitlement PDFs — a task that required much manual effort.

What treasury learned

Asked what had made their projects succeed, both men pointed to factors that had little to do with the technology’s complexity.

For Koeller, the critical ingredient was a dedicated counterpart from the digital finance team — someone who worked with treasury to understand the data and the business goals. That relationship became a template. “We were one of the first in finance to really make that foray,” he said. “I think we have a good relationship, we have that rapport built. We know who to deal with and how to get things done faster.”

For Zavertnik, it was executive sponsorship and banking partnerships. The project to overhaul cash visibility was mandated from senior leadership — which made it straightforward to secure cooperation from other departments. He was equally clear that working with bank partners who had invested in API capabilities, and who were prepared to commit technical resources to the implementation, was essential for a team without internal IT bandwidth.

The economic case, both stressed, was easier to make than treasury professionals sometimes assumed. “You have cash all around the world,” Zavertnik said. “We don’t live in a zero interest rate world anymore. If we were able to get $2 or $3 million more in average deposits that are sitting around underutilised into money market funds, you could easily pay for an implementation like this.”

What came next

Both organisations’ plans were shaped as much by what they had come to know was possible as by any defined roadmap.

Zavertnik’s immediate priority was cash forecasting — using machine learning algorithms to generate run-rate projections from the transaction data flowing reliably through the system. Longer term, he envisaged a self-service chatbot through which executives could query cash positions directly, drawing on live data via the platform’s developer API. Whether leadership would actually use it, he admitted with some amusement, remained to be seen.

Koeller’s framing was broader. The hedging model and the deposit rate agent were, in his view, evidence for something larger — a demonstration that treasury data, precisely because it was controlled almost entirely within the function, was an unusually tractable starting point for AI adoption. “We weren’t doing anything that was going to change something in the ERP directly,” he said. “We control that data.” That containment made it easier to build trust, and that trust was being deployed across a wider set of initiatives as the system became integrated across ASML’s enterprise.

For both treasurers, the lesson of the past two years was less about any specific technology than about what it took to move an organisation. Data quality, executive agreement, cross-functional relationships, and a willingness to prove things incrementally before claiming them broadly — these were the conditions under which automation actually happened, as opposed to the conditions under which it was merely discussed.

Spreadsheets, Koeller added, were not going away. The point was not to replace the tools that treasury professionals understood. It was to make room for the work that those tools, increasingly, could not do.