A new IMF paper argues that while agentic AI could transform how payments are initiated and managed, the core infrastructure must remain deterministic to preserve trust, legal certainty and financial stability
A new paper from the International Monetary Fund (IMF) has pushed back against the prevailing narrative of increasingly “intelligent” payment systems, warning the core infrastructure underpinning transactions should remain deliberately rule-based, even as artificial intelligence becomes more autonomous.
In an April 2026 note, How Agentic AI Will Reshape Payments, authors Sonja Davidovic and Hervé Tourpe outline how a new generation of AI systems capable of acting independently on behalf of users could transform the payments landscape. However, they draw a clear boundary around where that intelligence should sit.
At the centre of the paper is a key tension: modern payment systems are built on deterministic logic, while agentic AI operates probabilistically.
“Payment infrastructures… rely on predictable rules, legal certainty, and clear accountability structures to ensure trust and financial stability,” whereas AI introduces “probabilistic decision-making systems capable of initiating financial actions at machine speed,” the authors write.
Rather than embedding AI deeper into payment rails, the IMF argues this distinction must be preserved.
Keeping payments ‘dumb’
The paper proposes a three-layer model for future payments architecture, separating AI-driven decision-making from execution.
Under this framework, AI agents operate at the “intent and orchestration” layer, where they interpret user objectives and coordinate transactions. However, the “authorisation and control” and “settlement” layers remain strictly deterministic, ensuring payments are executed according to fixed rules.
The reasoning being, as the IMF warns, that the risk does not stem from AI itself, but from allowing adaptive systems to directly execute irreversible transactions.
“The main risk with agentic payments does not come from using probability-based reasoning itself, but from letting adaptive systems make irreversible payments without proper controls, checks, or accountability,” the authors write.
In effect, the IMF is arguing that while payments may become smarter at the edges, the core system must remain “dumb” by design.
From ‘click-to-pay’ to ‘decide-to-pay’
Beyond architecture, the paper reframes the shift underway in payments as something more fundamental than faster rails or improved user experience. The authors highlight a transition from explicitly human-initiated transactions to agent-led decision-making.
“This evolution can be characterised as a shift from explicitly human-initiated transactions (‘click-to-pay’) toward agent-mediated decision processes (‘decide-to-pay’),” the IMF notes.
In practical terms, this means software agents could:
- Monitor conditions and trigger payments automatically
- Optimise routing across different payment rails
- Manage liquidity and foreign exchange decisions in real time
While early use cases are already emerging across e-commerce and cross-border payments, the IMF emphasises that adoption remains at an experimental stage.
The paper acknowledges agentic AI could deliver significant operational benefits, including lower costs, faster transactions and improved liquidity management. Agents could “coordinate and execute multistep workflows across financial distributed networks autonomously,” including selecting optimal cross-border payment routes and managing treasury functions.
However, the authors place equal emphasis on the risks, some of which extend beyond traditional payment concerns. One of the more notable warnings relates to the potential for correlated behaviour among AI agents. If multiple systems rely on similar models and signals, they could act in unison, creating stress across payment systems.
“Correlated agent behaviour… can lead to synchronised payment flows, thus increasing intraday liquidity demand and potentially straining settlement capacity,” the authors write.
This introduces a type of systemic risk more commonly associated with algorithmic trading than retail or commercial payments.
Legal and regulatory gaps emerging
The rise of agent-initiated payments also raises questions around authorisation and liability, which existing frameworks may not be equipped to handle. Current systems require a payment order to be traceable to a clearly authorised instruction from an account holder. That link becomes less clear when decisions are delegated to autonomous agents.
“Authorisation instead becomes structural and mandate-based, which raises questions about traceability, consent, and liability under existing legal frameworks,” the authors note.
The paper highlights broader uncertainty around accountability when things go wrong, particularly in cases where agents act within delegated authority but produce unintended outcomes.
Despite outlining a range of use cases, the authors stop short of advocating for widespread deployment of agentic AI in payments. Instead, they frame the issue as one of governance and system design.
“The challenge… is not whether to adopt agentic technologies, but how to integrate them into payment systems in a way that preserves trust, stability, and accountability,” the authors conclude.
In other words, the future of payments may depend as much on what systems choose not to automate as what they do.
