• Quique Fernández, the Head of Treasury Transformation for the UK and Ireland at Embat


• 25.02.2026 08:15 am
#TreasuryAI #AI

The treasury dashboard blinks red as real-time balances stream across currencies and blockchain networks. An agentic AI flags a cash-flow anomaly; a stablecoin transfer settles instantly across borders before the CFO can even confirm it. The technology works, the question is whether the organisation does.

Gartner predicts that by 2028, one in three enterprise software applications will incorporate agentic AI. EY-Parthenon's 2025 Stablecoin Survey finds that over half of financial institutions not yet using stablecoins plan to adopt them within a year. The momentum is undeniable, but so is the failure rate. Over 40% of agentic AI projects are projected to be cancelled by the end of 2027, Gartner says.

The difference between systems that compound efficiency and systems that amplify chaos isn't the sophistication of the platform. It's whether the underlying treasury processes can actually support autonomous decision-making at scale. Speed means nothing if the rails beneath it are fractured.

Approval vs. Liability

The promise of agentic AI in finance is simple: let algorithms handle routine cash management while humans focus on strategy. But after three weeks of clicking "confirm" on 50 flawless recommendations, humans may stop looking. Then, approval becomes a rubber stamp, yet liability still sits with the human.

Boards aren't rejecting AI because they distrust algorithms; they're rejecting it because organisations haven't designed accountability processes that work with autonomous systems. Audit committees won't approve autonomous cash management when the audit trail can't explain why a decision was made or who is responsible when something goes wrong.

What's actually shifting in finance is humans moving from making individual decisions to designing the processes that govern how and when AI operates. You're no longer approving transactions; you're defining and setting programmatic guardrails (e.g., counterparty exposure limits or KYC/AML triggers) that the agent cannot override.

The shift is from probabilistic AI, which predicts the next word, to deterministic execution, where the AI is bound by hard-coded treasury policy. The AI functions as a policy executor, not a decision maker.

Again, rails only work if they're built on clean processes. An AI deployed into inconsistent reconciliation workflows or duplicate vendor records won't learn optimal behaviour - it will learn your organisation's bad habits and execute them at scale. 

When a supplier updates bank details and gets flagged as fraud, that's not an AI failure. It's a process failure that AI exposes faster than any human ever could.

Stablecoins: When Classification Beats Technology

Stablecoins promise instant settlement, 24/7 availability, no intermediaries. For treasurers managing liquidity across volatile currencies - Argentina, Venezuela, Nigeria - the value proposition is obvious.

But the biggest barrier to corporate stablecoin adoption isn't blockchain infrastructure or regulation - it's the absence of processes for ramp-on, ramp-off, and accounting classification. In stable-currency environments, conversion fees can outweigh settlement savings because no process exists to determine when stablecoins should be used.

Custody introduces another blind spot. If a company holds USDC in a self-custodied wallet, most audit frameworks can't document who controls the private keys. If the wallet is compromised, there's no FDIC insurance or established liability framework.

The companies succeeding with stablecoins aren't those with the best blockchain stacks. They're the ones that built processes convincing audit partners to treat regulated stablecoins as cash equivalents. Without policy alignment between MiCA issuance rules and GAAP/IFRS reporting, stablecoins act as an operational tax.

Mature treasury teams are building hybrid liquidity models where fiat and stablecoins coexist in a single dashboard and feed into the same ERP. That model collapses if your process architecture can't handle assets with different settlement speeds. 

Speed doesn’t matter if the data isn’t sound

Finance teams now choose between bank wires, instant payments, and stablecoins - each with different costs, speeds, and compliance requirements. When routing decisions must happen in real time, human judgment becomes the bottleneck.

This is where AI routing shifts from "nice to have" to "non-negotiable." An AI-driven router must navigate a 'multi-rail' environment, choosing between ISO 20022-compliant bank wires, Real-Time Payments (RTP), and on-chain stablecoin settlement based on cost, liquidity, and time. 

An AI that routes a payment incorrectly because vendor data wasn't updated hasn't failed at routing - it's exposed a data governance failure. That's when the real questions surface: Who updates vendor records? How often? What approvals are required when bank details change? Most organisations only confront these gaps once AI forces the issue.

Treasury is often the first place where AI makes real financial decisions without asking permission - because policies are clear and customer-level judgment is minimal. Success or failure here sets the tone for autonomous systems everywhere else.

Software Won't Fix Your Processes

Buying new finance technology won't fix broken processes. You have to fix the processes first, and that means data governance, deduplication, standardised vendor workflows, and real-time integrations between banks and ERP systems.

The typical failure pattern is predictable. A treasury team deploys an AI forecasting tool, but forecasts are inaccurate and the vendor gets blamed. But if the organisation has five definitions of "cash," inconsistent transaction categorisation, and reconciliation that lags by days, the platform only works when the data does.

The same is true for stablecoins. Companies that succeed invest time and money building frameworks for accounting, audit, tax, and compliance before moving money on-chain. Those that fail buy the platform first and discover the gaps mid-flight.

Sources:

Gartner, Inc., "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (June 2025)

EY-Parthenon, "2025 Stablecoin Survey of Corporates and Financial Institutions" (June 2025)

About the author

Quique Fernández is the Head of Treasury Transformation for the UK and Ireland at Embat. Formerly Treasury Manager at Northern Data Group and Head of Treasury Operations at Teya, Quique is a specialist in creating modern, scalable treasury functions. He helps businesses implement best-in-class strategies for cash visibility, FinOps automation, and bank connectivity, ensuring finance teams can achieve true digital transformation.