By: Eleanor Hill
Faster payment rails, digital currencies, and increasingly autonomous agents are being sold as the seamless future of finance. Yet the smoother the experience becomes, the easier it is to miss the fault lines underneath – until access disappears or delegated logic acts beyond its limits. Here’s what to look out for along the journey.
Adapted from a keynote delivered by Eleanor Hill at the SAP for Treasury and Working Capital Management Conference, hosted by TAC Insights in Rome in June 2026.
In the third century BCE, the Temple of Juno Moneta stood on Rome’s Capitoline Hill, honouring her role as protectress of the Roman state. Later, the city’s mint was built nearby, and over the centuries the goddess’s name drifted from temple to workshop to coin. That’s where we get words like money and monetary from – and tradition also ties the name to monere, to warn.
Of course, historians can keep arguing over the precise etymology, but the warning is worth heeding, regardless. Because money has never been neutral. In a nutshell, money is authority made portable. It’s also the embodiment of trust turned into infrastructure, and a promise that only works for as long as enough people keep believing in the institution behind it.
Treasurers, in that sense, are Juno’s descendants – not there simply to move money efficiently, but to protect the organisation’s ability to use it once conditions turn. And once autonomous technology takes hold.
After all, faster payments, stablecoins, tokenised deposits, central bank digital currencies, and AI agents will all change how value looks and moves, but none removes the questions that have always sat underneath. Who can issue the promise? Who can enforce or alter its terms? What has to keep working before the owner can actually use it? And who gets shut out when conditions change?
When access disappears
While the nature and form of money might be evolving, if you strip away the metal, the paper and now the code, money still rests on three things: authority, acceptance, and access.
Authority decides who can issue or validate a claim. Acceptance decides whether anyone else will take it. Access decides whether the owner can use the asset when and where they need to. Together these create trust – though access is usually the one that goes unnoticed… until it’s gone.
The US banking crisis of 1933 is an extreme, but pertinent, illustration. Bank runs had already forced widespread closures by the time Franklin Roosevelt took office. And just 36 hours into his presidency, at one o’clock in the morning, he suspended banking transactions across the entire country.[1]
Balances remained at the bank over the following week – but their owners couldn’t reach them, and had no idea when, or whether, their branch would reopen. The money was still there. What had gone was usability and confidence. In other words, ownership and access had been torn apart.
That same fault line runs through sanctions, payment-network exclusion, trapped cash, and banks simply changing their appetite for risk. Russia’s sovereign reserves didn’t stop existing when Western jurisdictions immobilised them after the full-scale invasion of Ukraine, and disconnected Russian banks didn’t lose legal title to their balances either.[2] What changed was whether those assets and that infrastructure could still be used on the old terms.
For a treasury team, the fallout is practical and immediate: failed settlements, liquidity you can see but can’t touch, disrupted payroll, delayed supplier payments, and funding decisions made in a hurry with none of the usual time to think.
Spreading cash across several banks doesn’t automatically fix this, either. Each account can still depend on the same currency, correspondent network, cloud provider, jurisdiction, or settlement infrastructure. A backup only reduces risk if it breaks the dependency that failed – three routes back to the same point of control are not three independent paths.
Access can also narrow without a bank failing or a government imposing sanctions. Financial providers continually reassess the sectors and activities they’re prepared to support. One major European cash management bank, for example, announced in 2017 that it would stop financing and investing in tobacco companies, and market chatter suggested that those businesses were subsequently given six months to find new cash management arrangements.[3] In other words, legal has never automatically meant bankable.
Where cash is placed can carry consequences beyond continued access, too. Research from the Strategic Treasurers Alliance and Topo Finance modelled the emissions tied to cash deposits and money market funds (MMFs) across ten companies and more than 25 banking relationships. Estimated emissions intensity varied by more than 40 times between counterparties, while financed emissions from cash and money-market funds averaged 283% of the companies’ own reported carbon footprints.[4] Counterparty diversification therefore has to consider policy direction and the use to which liquidity is put, not just conventional credit strength.
The same scrutiny is also needed when and how money moves, not just where. A payment route can look new at the surface while remaining dependent on familiar institutions, jurisdictions and infrastructure underneath.
New rails, old dependencies
Cross-border payments have genuinely improved, but much of the plumbing underneath still belongs to an earlier era. Treasury now wants speed, transparency, richer data, and round-the-clock availability from systems that were largely built around bank messages, correspondent relationships and working-day settlement.
Swift is responding – and its shared ledger connecting banks’ tokenised deposits is designed to support always-on cross-border movement.[5] But stablecoins, tokenised deposits, and domestic instant payment systems have already spent several years exposing exactly where the traditional model runs out of road.
Whenever a new route appears, treasury needs to look past what’s on the screen. Who’s carrying the instruction? Which asset actually settles the obligation? Where does the liquidity sit? What compliance process can stop the transaction, and under whose jurisdiction? At what point does the “new” route simply reconnect with the infrastructure it was meant to replace?
Stablecoins also bring that problem into intense focus. They’re often pitched as a way around the established financial system, but the largest stablecoins might instead be reinforcing one of its oldest hierarchies. The market has grown past $300bn, more than 99% of stablecoin supply is denominated in US dollars, and adjusted transfer volume now exceeds $10tn over 12 months.[6] [7] The comparison with card spending isn’t like-for-like, but on-chain dollar movement is no longer confined to cryptocurrency trading.
The geopolitical point is sharper still. Dominant stablecoins are digital dollar claims backed mostly by cash and short-dated US government debt. In fact, BIS researchers estimate that dollar-backed issuers bought roughly $35bn of US Treasury bills in 2024 and another $33bn in 2025.[8] Every business holding a dollar stablecoin – or using one as a bridge between two non-dollar payments – extends the dollar’s reach rather than diluting it.
Reliance on one domestic bank might fall, but dependence grows instead on the issuer, its custodians, the treasury market and the rules governing redemption. A rail that looks stateless can still deepen reliance on a state-backed currency (worth keeping an eye on at a time of deep geopolitical tension and desired de-dollarisation).
What lies beneath
Elsewhere, digital money products can look almost identical sitting in a wallet while representing completely different underlying claims. A stablecoin asks the holder to trust a private issuer, its reserves, custodians, banking relationships and redemption mechanism. A tokenised deposit remains a liability of a commercial bank, while a central bank digital currency would be a direct central-bank liability carrying its own policy choices around privacy, access and control.
None of that is visible from the interface. In fact, the buying screen tells you almost nothing about which promise you’re holding – so significant due diligence is required.
USDC delivered a sharp reminder during the collapse of Silicon Valley Bank in March 2023. Circle disclosed that $3.3bn of the reserves backing the stablecoin were sitting at the failed bank, temporarily out of reach.[9] The tokens themselves never disappeared, but USDC briefly lost its one-to-one peg as the market repriced the reserves, redemption route, and banking access behind them. Despite the name and marketing claims, pressure had reached the structure holding the promise up.
Digital money therefore calls for an old treasury discipline: classify the liability before admiring the wrapper. Who issued the claim? Where do the supporting assets sit? Is redemption available at par and on demand? When does settlement become final? Can the asset be frozen? Where would the corporate rank if the issuer, bank or custodian failed?
When logic moves money
The next shift goes beyond the form money takes or the rail it travels on, to who authorises its movement. Programmable payments are already live: Siemens has applied them to cash allocation, intercompany funding, and liquidity concentration, while BMW has completed a programmable on-chain foreign-exchange transaction.[10] [11]
Taking it a step further, agentic systems introduce a different question because they don’t merely execute a condition set in advance. While a programmable payment follows a defined rule, an AI agent receives an objective, gathers information, chooses tools, decides its next move and adjusts its route as circumstances change, within whatever authority the organisation has handed it. One automates an instruction, but the other decides how to pursue an outcome.
One of the biggest concerns here is that an agent doesn’t need to collapse or produce an obvious hallucination to become unreliable. Old facts can remain in its memory after they’ve stopped being relevant, for example. Or a one-off exception can begin functioning as precedent. Likewise, repeated summaries can shed important conditions without anyone noticing. And a preference learned in one context can shape a decision it was never meant to touch.
Research is beginning to quantify this deterioration, which is referred to as agent drift. One benchmark of evolving data-analysis work found accuracy falling by almost 47 percentage points between the early and late stages of a task, not through one dramatic error but through the gradual loss of the correct analytical state.[12]
Separate research found that stored observations such as impatience, cost-consciousness, or risk tolerance could alter parameters in later, unrelated tool calls.[13] Throughout that drift, the system can remain fluent, responsive, and highly convincing.
Even more worrying is the fact that human vigilance tends to move in the opposite direction. In one widely cited study, operators detected only around 28% of automation failures under constant high reliability, compared with just over 80% when reliability varied and users remained more alert.[14] Agentic AI sharpens that established weakness because the system can choose tools, retain memory, make connected decisions, and take action.
Every organisation introducing agents therefore needs a delegation map covering what each system can observe, remember, infer, recommend, initiate, and execute. A human reviewer who sees only the final recommendation can be technically in the loop while lacking everything needed to challenge it. Meaningful control depends on seeing which sources were used, what the agent was allowed to remember, whether an earlier instruction had expired, and why a particular action was selected.
A human in the loop isn’t a safeguard if that person has been left out of the context.
Find the cracks first
Of course, none of this is an argument against innovation. Faster payments can free up liquidity and cut friction, digital money could solve genuine settlement and cross-border problems, and AI can bring earlier signals into treasury, surface patterns manual analysis would miss, and remove enormous volumes of administrative work.
What’s important to note is that issuing a warning isn’t a rejection – it’s a request to pay proper attention before the conditions change.
Before relying on a new form of money, payment rail, agent, or provider, treasury should ask four connected questions:
· Who controls it – who sets the rules, changes them, or withdraws access?
· What’s the promise – which legal claim is being held, who issued it, and how does it behave under stress?
· What has to work – which banks, rails, correspondents, models, data sources, cloud systems, compliance processes, and people sit between instruction and settlement?
· Who carries the loss – the issuer, bank, provider, corporate, or customer?
One further test runs through all four:
· Can the organisation pause the process, override the decision, and reconstruct what happened? A pause needs a real trigger, a named owner, and a route back to a known safe state. An override must be technically possible, not just theoretically allowed. Reconstruction means identifying which data, memory, rule, and authority produced the outcome.
Returning, then, to where we started in Rome, let’s remember that Juno Moneta stood for protection and warning long before digital currencies, payment rails, or AI agents existed. Although the technology has changed beyond recognition, the responsibility underneath it hasn’t. If anything, it has intensified with scale and speed.
The real test of money has never been what it promises when everything is working normally. It’s what still works when something starts to crack – and it’s the treasurer’s responsibility to find those fault lines before pressure widens them, ensuring the organisation retains access, control, and a route to safety when it matters most.
Endnotes
[1] Federal Reserve History, “Bank Holiday of 1933”. Roosevelt issued the proclamation at 1am on 6 March 1933, 36 hours after taking the oath of office, suspending banking transactions nationwide for the week.
[2] Council of the European Union, “EU sanctions against Russia explained”, and “Immobilised Russian assets: Council decides to set aside extraordinary revenues”. The measures prohibited transactions involving Central Bank of Russia reserves, immobilised around €260bn across G7, EU and Australian jurisdictions, and excluded a number of Russian banks from SWIFT.
[3] BNP Paribas, “BNP Paribas announces new measures regarding the financing of tobacco companies”, 24 November 2017. The two six-month cash-management cases described in the article come from the author’s own reporting.
[4] Strategic Treasurers Alliance and Topo Finance, Financed Emissions Guide for Treasurers, 2026. Report to be published Q3 2026. The ten-company analysis covers more than 25 banking relationships. Its figures are modelled indicators derived from disclosures and proxies, not audited Scope 3 totals.
[5] Swift, “Swift’s blockchain ledger ready for use as 17 banks set to pioneer tokenised cross-border payments”, 9 July 2026. The shared ledger provides an orchestration layer for bank-issued tokenised deposits, enabling overnight and weekend movement before final settlement through existing systems.
[6] Marco Carapella, Marco Lubis and Alexandros Vardoulakis, “Stablecoins in 2025: Developments and Financial Stability Implications”, FEDS Notes, 8 April 2026; and Visa, “Stablecoins and the future of onchain finance”. The Federal Reserve recorded aggregate market capitalisation of $317bn on 6 April 2026; Visa reports that more than 99% of supply is US-dollar denominated.
[7] Visa, “Stablecoins and the future of onchain finance”. Visa’s adjusted on-chain measure filters bot-driven and other non-organic activity and recorded $10.2tn of transfer volume over the preceding 12 months.
[8] Rashad Ahmed and Iñaki Aldasoro, “Stablecoins and safe asset prices”, BIS Working Papers No. 1270, revised June 2026. The authors estimate net Treasury-bill purchases of roughly $35bn in 2024 and $33bn in 2025 by dollar-backed stablecoin issuers.
[9] Circle, “$3.3 billion of USDC reserve risk removed, dollar de-peg closes”, 13 March 2023. Circle said $3.3bn, around 8% of USDC reserves at the time, had been held at Silicon Valley Bank.
[10] J.P. Morgan, “Siemens Treasury’s Digital Transformation”, covering programmable workflows for cash allocation, intercompany funding and liquidity concentration.
[11] J.P. Morgan, “BMW Group completes first programmable on-chain FX transaction using Kinexys”, December 2025.
[12] Yinghao Xu et al., “LongDS-Bench: On the Failure of Long-Horizon Agentic Data Analysis”, arXiv:2605.30434, 2026. This preprint benchmark covers 68 multi-turn data-analysis tasks and reports an almost 47-percentage-point decline from early to late stages; it is not a production treasury failure rate.
[13] Abhinav Dabas et al., “Memory-Induced Tool-Drift in LLM Agents”, arXiv:2605.24941, 2026. The MEMDRIFT preprint tests whether stored personality observations, including cost-consciousness, impatience and risk tolerance, alter parameters in later unrelated tool calls.
[14] Raja Parasuraman, Robert Molloy and Indramani L. Singh, “Performance Consequences of Automation-Induced ‘Complacency’”, The International Journal of Aviation Psychology, 3(1), 1993, pp. 1–23. In the multitask condition, detection was around 28% under constant high reliability and just over 80% when reliability varied.
