Published: June 12, 2026
Agentic AI is already here, but treasury workflows won’t be transformed by novelty features or scattered pilots. Progress belongs to teams that build sturdy foundations across data, process, and governance, then prove value in the language boards recognise – time saved, risk reduced, and decisions improved. Here, three industry experts share their insights on how to leverage AI to do just that.
Although AI has arguably moved beyond pure hype, plenty of treasury teams still find themselves stuck between sprawling toolsets that promise everything – but cause confusion – or cautious postponement that leaves value on the table while peers learn at speed.
What 2026 demands, our experts believe, is a deliberate operating model where people, processes and systems reinforce one another, and where AI serves to deepen what treasury already does well rather than masking the fact that the groundwork is missing. This requires reflecting on a few cold facts.
“We’re past debating whether AI matters,” says Grant Walker, Senior Product Group Manager for Treasury, Pensions and Insurance, Nestlé. “It’s now really about how we scale it responsibly, globally – using automation and intelligent workflows to bring efficiencies, reduce manual effort, and support faster decisions while creating more capacity for value-added work.”
That framing is important because it grounds AI in treasury’s existing mandate rather than positioning it as a separate innovation agenda. The promise is not transformation for its own sake, but a rebalancing of effort – less time spent assembling information, more time spent understanding and acting on it. What changes is not the purpose of treasury, but the proportion of time available for judgment.
Yet scale also introduces exposure. Tools that work well in pockets can behave very differently when rolled out across regions, entities, and user groups, particularly when decision rights, data quality, and controls vary. Responsible scaling therefore becomes an organisational design challenge as much as a technical one, forcing treasury leaders to think hard about how intelligence is introduced, governed and explained.
One of the clearest risks to that ambition is agent sprawl. As agentic AI becomes mainstream, almost every vendor proposition now includes some form of autonomous assistant, often presented as a shortcut to insight or automation, but rarely as part of a coherent whole.
“Everybody has an agent that will do everything,” Walker observes. “But I don’t want my users having hundreds of agents that they have to call on. I want the experience to be simple – I ask my agent, it goes off, finds the answer, and delivers what I need.”
The concern here is not aesthetic, but structural. Multiple agents typically come with multiple permission models, interpretations of data, and various ways of triggering actions –all of which increase the risk of inconsistent outputs and erode confidence among users who are already (understandably) cautious by the nature of their role. What’s more, when trust weakens, treasury teams tend to fall back on manual workarounds that feel controllable, even if they undermine efficiency.
Aniket Kulkarni, Partner, Treasury and Commodity Trading, PwC, offers a useful way of reframing the challenge. “Your foundation should be to build your agent zero, and then you build all your other agents on top of it,” he says.
Agent zero is the foundation for Treasury’s AI strategy, which covers: data foundation, process foundation and clearly defined user roles and responsibilities. It orchestrates how requests are routed across systems, coordinates usage of data sources, and ensures user rights and permissions are respected.
Without that discipline, intelligence simply accelerates fragmentation. With it, agentic AI becomes something far more useful: a consistent ‘front door’ into a governed ecosystem.
Dr Arif Esa, Global Lead for Treasury and Working Capital Solutions, SAP, adds that as agentic AI moves from analysis to action, the importance of control frameworks only increases. He notes: “Segregation of duties and identity management must be in place. Everyone talks about agent-to-agent workflows, but identity has to remain clear.”
His warning is direct: “Do not leave the agent to decide on its own.” Rules and responsibilities must be inherited from systems of record, with data protection and privacy built in. “The system of record should lead, and the agent should be a consumer of those rules.”
Walker reinforces the same point from a vendor and integration perspective. “Don’t chase shiny features,” he advises. “Integration is what enables governance to scale and prevents a proliferation of tools that sit outside treasury’s accountability framework.”
Few topics generate as much anxiety in AI discussions as data, and with good reason. In treasury, poor data does not merely reduce accuracy, it undermines credibility.
Esa argues that the starting point must always be systems of record. “We should start with the application that produces the data – finance, supply chain, treasury – and make sure that data is collected in a way that is semantically useful, so that AI can be applied consistently.”
Semantic usefulness, in this context, is not academic. It means consistent definitions, harmonised structures and predictable behaviour across entities and geographies – the conditions that mean outputs can be explained and defended when challenged.
Esa is particularly cautious about blending internal and external data too early. “Do not start to combine external and internal data without having a covenant in place,” he says, pointing to the need for a defined layer that governs how data should look before it is exposed to AI. Only once internal and external data are harmonised “in one single data model” does intelligence become something treasury can stand behind.
He describes the relationship between applications, data and AI as a flywheel, where each reinforces the other. “Only if you have a close loop between the three will you be in a position to get the best value out of AI,” Esa states.
Walker echoes the same point from a corporate perspective, stressing that governance must be tangible rather than performative. “We invested early in master data governance,” he says. “Not just on the surface, but with measures and KPIs, because if you want to scale, you need confidence in the data that sits underneath everything.”
Once the right architecture and data governance is in place, the next question is where automation should actually run. For treasury, the answer is rarely at the level of a single task.
“Treasury is the most integrated function within finance, both internally and externally,” says Kulkarni. Cash flows originate in order-to-cash and procure-to-pay processes, then move through banks, markets, and internal controls, with dependencies that are structural rather than incidental.
This is why so many early AI initiatives disappoint. They optimise one step, then hand the problem to the next. Hyperautomation, by contrast, is about continuity – reducing handoffs, breakpoints and blind spots across the full flow.
Kulkarni describes growing interest among corporates in moving “from automation to hyperautomation”, particularly where agents can operate across processes rather than within them. A practical illustration comes from cash management. “Large corporates have hundreds of accounts that treasury needs to monitor and control,” he says. “It’s impossible for a team to look at all of these accounts and understand, at any moment, which balances are negative or excessively positive.”
Here, the value of agentic AI lies in consistency rather than sophistication. It can be used to define thresholds that reflect treasury’s liquidity strategy, surface exceptions as they occur, and enable people to focus on decisions rather than detection. “If the balance is negative one million or positive ten million, then the cash manager should get an alert from the agent, without having to lift a finger,” Kulkarni explains.
Crucially, though, that alert matters only if the process that follows is clear, governed, and understood. Hyperautomation forces treasury to confront questions about escalation, authority, and timing that manual processes often obscure.
Even with strong, well-governed foundations, AI adoption does not happen automatically. Treasury teams are cautious by design, and confidence has to be earned.
“Across finance and treasury, it starts with natural curiosity,” Walker says. Early experimentation plays an important role, particularly when it reveals limitations as well as benefits, helping users understand when to trust outputs and when to challenge them.
Scope discipline is central to that learning process. “We kept the scope small,” Walker explains. “We built our own little agents, designed to deliver information quickly and efficiently, but accuracy was non-negotiable.” In treasury, starting small is not a lack of ambition; it is how trust is built.
From there, the work becomes cultural. “It’s not just about having a PowerPoint or a Word document AI policy that never changes,” Walker says. AI becomes a living internal knowledge layer, updated as processes evolve and accessed when people need support, rather than consumed once and forgotten.
To spread capability, Walker’s team created a global finance AI Club, giving teams a space to share use cases, ask questions, and learn from one another, followed by a playbook so approaches could be replicated across markets. “Not everybody has the same level of curiosity,” he notes. “So, we had to think carefully about tiered enablement.”
Prompting emerged as another practical enabler. Through ‘Prompt Buddy’, teams shared effective prompts rather than relying on trial and error, anchoring adoption in visible outcomes. “Don’t just say, here’s a tool,” Walker says. “Show what it does for me, how it changes my day.” Time saved resonated most clearly: “It’s reduced 10, 20, 30% of the time I spent doing X, Y and Z.”
Kulkarni adds that fear often sits beneath resistance, particularly around job impact and dependency. Adoption improves when leaders can explain boundaries clearly and demonstrate that AI complements judgment rather than replacing it.
In summary, agentic AI will continue to advance quickly, and expectations will continue to rise. Treasury’s success, however, will be determined less by speed than by discipline.
“It’s not just a technology shift,” Walker reflects. “It’s a capability shift and a mindset shift.” Kulkarni is equally clear. “AI cannot be implemented without transformation.” Meanwhile, Esa brings the focus back to fundamentals. “Do not lose the connection between people, processes, and systems,” he says. “Without user confidence, adoption will not happen.”
Taken together, these themes point to a demanding but realistic agenda: design the architecture before adding intelligence, govern data before trusting outputs, enable people before expecting adoption, and keep systems of record firmly in charge.
Treasury has always balanced innovation with control, and agentic AI does not change that identity. It sharpens it – and for teams willing to do the groundwork early, it offers an advantage in the quality and confidence of everyday decisions.
This article is available to listen to
AI is working its way into almost every aspect of professional life. Sue Turner OBE, Founding Director, AI Governance Limited, and Professor in Practice – AI and Digital Technologies, University of Bristol Business School, offers an expert view on fostering an appropriate AI strategy for treasurers.
As digital assets mature, treasurers are reassessing their relevance beyond experimentation. Societe Generale’s Geneviève Douhet, Director of Digital Assets Programs, Global Transaction Banking, and Floor Meeuwis, Director of Liquidity and Virtual Account Management Advisory, Global Transaction Banking, join Stéphanie Cabossioras, Chief Strategy & Global Policy Officer, Societe Generale-FORGE (SG-FORGE), in examining what digital money can realistically deliver for treasury, and why regulated, bank issued stablecoins may represent a tangible solution.
