EuroFinance

The pace of change in treasury technology has accelerated sharply. Use cases are emerging, systems are being overhauled, and for the organisations with the resources and appetite to move, the question has shifted from whether to adopt AI to how, and in what order.
But most treasury teams are not there yet. Adoption is uneven, foundational data work remains unfinished, and the organisations pushing hardest on automation are discovering a new set of problems: who is accountable when a model produces an output no one can fully explain, and what happens when systems stop advising and start acting?

Not one speed, but many

Salesforce, a cloud-based software company, was an early adopter of treasury management systems. Yet on artificial intelligence, progress has been mixed. “When I think about AI and other things like that, we’re a mix between fast and slow—because we were doing all these other things, we don’t have the resources to do a lot of stuff,” said Linda Xu, senior treasury manager at Salesforce. “You might be quick on something but slow on others.”

At Lyft, a global mobility company operating across six continents, the approach is more deliberately sequenced. “We tend to invest in technology when we can align the automation to the business return on investment (ROI),” said Patrick Lee, treasury manager at Lyft. “That could mean operating efficiencies, cost savings, derisking operations or helping the organisation scale.”

The data foundation

Most technology transitions in treasury face a fundamental challenge: data. Systems proliferate, formats diverge, and the result is a landscape in which useful information exists in abundance but is frustratingly difficult to consolidate and act upon.

Salesforce adopted a single platform last year to retire its separate TMS and consolidate onto a single platform. That ambition has not yet been fully realised. “We are currently still on two systems,” said Xu. “There is duplicative work. It is challenging.” Product gaps remain; backlog items accumulate. As a bridge, Salesforce uses a business intelligence platform to impose a degree of uniformity across outputs from disparate systems.

The goal Xu articulated is broader: “The data lake would be the game changer—having all the data in one place and being able to mould it.”

At Lyft, the same logic drives a TMS implementation to replace a collection of ERPs, cash-forecasting applications and the spreadsheets that connect them. The TMS provider recently introduced AI-powered cash forecasting, but the company has not yet activated it. “The quality and structure of our input data comes first — what you put into the model determines what you get out.”

What treasury actually wants from vendors

When evaluating vendors, treasurers increasingly find that the most important variable is one that rarely features in marketing materials: support.

“What can make our lives easier? What can make things go faster?” said Xu. Salesforce regularly needs to integrate new bank accounts and banking relationships at speed. “Our current TMS doesn’t have the greatest support. Every time we’re talking to new vendors, it’s: what is the support model like? How will you support us in times of really urgent payments?”

Lee framed the evaluation similarly—as an exercise in looking beyond the sales pitch. “More and more, we need faster access to data and real-time visibility into our bank accounts.” he said

Xu pointed out that: “No system is perfect. No matter how organised we try to be and how thorough we are with our requirements, it sometimes isn’t enough. There’s a lot of give and take.” The most practical mitigation, she says, is the quality of the consultant engaged from the outset.

“The consultant we had during the backlog items was amazing. If we had had him at the start, we would probably have hashed these things out sooner and started talking about those gaps earlier.”

AI: where it is actually being used

The AI applications gaining most traction in treasury today are narrower and more practical than the transformative ones most often discussed.

At Salesforce, the cash forecasting team has been the most active adopter, using tools to reformat TMS outputs into shapes suited to their models. “All the systems spit out reports in all these formats and you spend more time massaging it to be what you need it to be,” said Xu. AI eliminates that reformatting work.

Banks: AI in the background

Banking partners appear to be deploying AI internally before exposing it to customers, constrained by the same data and regulatory concerns that slow adoption elsewhere.
The most visible application is fraud detection, though it operates largely below the surface. “Payment detection—things like a payment we’ve never made before, we would get alerts for that and they would ask us to validate,” said Lee. “Connecting the dots, perhaps that’s where AI is being deployed in the backend.”
The darker dimension is that the same tools available to legitimate users are available to bad actors. Deepfakes and voice synthesis are making business email compromise increasingly convincing. “Fight AI with AI,” as Xu puts it.

The accountability problem

The most fundamental tension in AI adoption for treasury is not technical but professional: if a model produces outputs that are accurate but cannot be fully explained, who is accountable when things go wrong?
Xu challenged the premise. “If you know it’s more accurate, then you should be able to explain why it got to be more accurate. That’s where I’m having a hard time with the question.”

Lee is more direct about why Lyft has paused on AI cash forecasting despite the capability being available. “As a user, even though AI is generating all these outputs and it could potentially be more accurate than I am—if I can’t explain it, I ultimately have the accountability.” Auditability matters independently of accuracy. “What are the audit trails behind the logic that goes into producing these outputs?”

Lee is explicit that he sees potential for AI cash forecasting to be valuable precisely because rideshare businesses by nature are influenced by external variables — weather, major events, seasonal patterns — that are difficult to incorporate accurately through manual methods. “Ideally, we’d have a solution that incorporates real-time external signals like weather and events directly into our cash forecasting models. ”

Governance and the risks of acting quickly

Beyond accountability, two concerns dominate the treasury’s thinking on AI adoption: governance and cybersecurity.

“Regulations around AI are very fragmented depending on geographic region,” said Lee. “The US AI regulation is still a work in progress. The onus is very much on us corporates to ensure that any technology and AI applications we use—that the data are highly secure and protected and the proper guardrails are in place.”

As treasury moves toward agentic AI—systems that not only advise but act—that responsibility grows. “As we’re moving into agentic processes where they’re actually executing on our behalf, that risk also grows exponentially.”

For Xu, the governance concern blends into a resourcing one. “How are we going to do our day jobs and also be innovative and do these things? It sometimes feels like it doesn’t line up.”

The API opportunity, and its limits

Bank API connectivity represents one of the clearer near-term opportunities in treasury—faster data delivery, real-time account visibility, and reduced dependence on legacy file-transfer protocols such as SWIFT.
“API integration is going to continue to be a high-value add for any organisation,” said Lee. “Getting that data is much faster than our traditional methods.”

Xu pointed to the gap between the technology’s potential and its current reach. “A lot of banks, if they do offer something, it’s only in the US.” Global API coverage would be transformative, she suggests, but remains some way off.

There is also a practical management challenge in running hybrid connectivity models. “This one does API, this one doesn’t,” said Xu. A coherent strategy for connectivity is as important as the technology itself.

What agentic AI could finally solve

Looking further ahead, both Xu and Lee are cautiously optimistic about agentic AI—systems capable of executing multi-step processes autonomously rather than simply generating outputs for human review.

Xu’s optimism is tempered by experience. “I think it’s smart for us to start slow and see where everything is. We need a lot of time to test things out. I’m excited to see more use cases come out.”

Lee’s wish list is specific. He would like agentic AI to address two of the treasury’s most persistent pain points: electronic bank account management, and bank fee analysis. “Comparing fees across banking partners is a known challenge — the structures are rarely fully comparable. That’s the kind of problem I’m excited for this technology to solve.” Lee also cites KYC (Know Your Customer) as another area where agentic AI could have a significant impact and one he is particularly excited about.