Agentic payments need agentic accountability

For two years, agentic payments were a conference slide. In 2026 they became an infrastructure program. Card networks shipped tokenized mechanisms that let AI agents initiate transactions against a user's preapproved payment methods without touching the underlying credentials. Shared protocols emerged so agents could discover merchants, compare offers and manage post-purchase logic. And in April, the International Monetary Fund published a note, How Agentic AI Will Reshape Payments, that did something rare for an institution of its size: it told the industry exactly where the line should be.

The line matters more than the hype. Because underneath the demos, a quieter question is compounding: when software starts initiating transactions at machine speed, who is checking the economics of all that movement, and how?

Short answer

Agentic payments are transactions initiated by AI agents acting under delegated user authority. The IMF warns that the core risk is adaptive systems making irreversible payments without proper controls, which means accountability, verification and audit evidence must run continuously at machine speed rather than through periodic manual review.

What the IMF actually said about agentic payments

The note's authors, Sonja Davidovic and Hervé Tourpe, identify a structural tension. Payment infrastructure earns trust through predictable rules, legal certainty and clear accountability. Agentic AI reasons probabilistically: the same inputs can produce different outputs, and the logic behind an individual decision is not always reproducible or legally interpretable after the fact.

Their conclusion is not that agents should stay out of payments. It is that the architecture has to keep the two natures separate. The IMF proposes a three-layer framing: agents can add real value in the upstream intent and orchestration layer, where they interpret goals, plan tasks and coordinate across services. Authorization and settlement, the layers where money actually moves and becomes final, must remain rule-based. In their words, the main risk "does not come from using probability-based reasoning itself, but from letting adaptive systems make irreversible payments without proper controls, checks, or accountability."

Read that sentence again, because it generalizes beyond authorization. It is a statement about where deterministic logic belongs in an AI-native financial system: anywhere an output must be reproducible, auditable and defensible. Execution is one such place. The economics of execution are another.

The accountability gap nobody is pricing

Consider what agentic payments do to transaction volume and shape. Agents negotiate, split, retry and route. A single human intent, "restock inventory at the best landed cost," can fan out into dozens of transactions across providers, corridors and currencies, each carrying its own fees, FX treatment and settlement terms. The per-transaction amounts shrink. The transaction count explodes. The fee surface area grows with it.

Now hold that against how most companies verify payment costs today: a finance analyst, a spreadsheet and a monthly invoice review that checks whether the total looks roughly like last month. That process already misses fee leakage of 0.2 to 0.5 percent of payment volume at current, human-initiated volumes. It has no chance against agent-initiated flow. The charges will be individually tiny, structurally complex and arriving continuously. Nobody will reconstruct the expected charge for any of them by hand, so nobody will know which ones were wrong.

This is the accountability gap. The industry is building the ability for software to spend money far faster than it is building the ability to prove the spending was priced correctly. Every basis point of unverified drift, multiplied across machine-speed volume, becomes a permanent tax paid to whichever counterparty's billing system happens to err in its own favor.

Regulation is converging on evidence

The regulatory direction makes the gap harder to ignore. The EU AI Act imposes transparency and documentation obligations that phase in through 2026 and 2027, and the IMF explicitly points to it as a key risk mitigation lens for agentic applications. In the United States, the GENIUS Act of July 2025 pulled payment stablecoins, one of the rails agents are most likely to use, into a federal framework with defined issuer obligations. Supervisors are already discussing know-your-agent standards to sit alongside know-your-customer.

The common thread is evidence on demand. Whether the examiner is asking who authorized this agent, why this transaction executed, or whether this charge matched the agreed terms, the acceptable answer is a record, not a recollection. Firms that cannot produce transaction-level lineage will discover that "the agent did it" satisfies no regulator, no auditor and no counterparty in a dispute.

What agentic accountability looks like

If the IMF's architecture is right for moving money, its mirror image is right for verifying it. The verification core has to be rule-based: contracts translated into executable pricing logic, an expected charge computed for every transaction from the clauses that govern it, and a comparison against what was actually billed, settled or deducted. That math must be reproducible, because it is the evidence. Every discrepancy needs the full chain: source transaction, contract clause, expected figure, actual figure, variance.

Agents then operate on top of that core, not inside it. They investigate discrepancies, assemble dispute cases, chase resolution with counterparties and monitor provider performance continuously, work that human teams cannot do at machine-speed volume and should not want to. Humans keep the judgment calls: which disputes to escalate, which providers to renegotiate, which terms to change. Intelligence at the edge, rules at the core, people in charge. The same design the IMF prescribes for payment execution, applied to payment economics.

The symmetry is the point. Agentic payments and agentic accountability are not two trends. They are one trend seen from both sides of the ledger, and adopting the first without the second is how a company ends up automated, fast and quietly wrong about its own costs.

The bottom line

Every prior leap in payment automation followed the same arc: the transacting capability arrived first, the control capability arrived late, and the gap between them was paid for in leakage, fraud and write-offs. Agentic payments offer the first chance to break that arc, because the same technology that initiates transactions can verify them, continuously and with evidence, from the moment the volume arrives.

That is the bet behind Bluefyn, and behind the broader shift it belongs to: an economy where software moves money will only stay trustworthy if software also proves the math. The companies that let agents spend will need agents that account. The ones that build both sides now will scale through the agentic era with their margins, and their auditors, intact.

Frequently asked questions

What are agentic payments?

Agentic payments are transactions initiated by AI agents acting under authority delegated by a person or business, rather than by a human explicitly approving each transaction. The agent interprets an objective, plans the steps and initiates payment through preauthorized mechanisms.

What did the IMF say about agentic AI in payments?

In an April 2026 note, the IMF argued that agentic AI can add value in the intent and orchestration layer of payments, but that authorization and settlement must remain rule-based. The core risk it identifies is adaptive systems making irreversible payments without proper controls, checks or accountability.

Are AI agents allowed to make payments today?

Yes, within emerging guardrails. Card networks and payment providers have introduced tokenized authorization that lets agents initiate transactions against preapproved payment methods without holding raw credentials. Regulatory frameworks, including the EU AI Act and stablecoin legislation like the GENIUS Act, are shaping the obligations around those flows.

How do you audit AI-initiated payments?

The same way any payment should be audited, at higher frequency: reconstruct the expected charge for each transaction from the governing contract, compare it to the actual charge, and keep an evidence-grade record of every variance. Because agent-initiated volume is continuous, the audit has to be continuous too.

What is the biggest risk of agentic payments?

Loss of accountability. Probabilistic systems making irreversible transfers without reproducible logic create disputes nobody can adjudicate and costs nobody can verify. The mitigation is architectural: keep execution and verification rule-based, and let agents operate around that core.

What is know-your-agent?

Know-your-agent refers to emerging frameworks for verifying the identity of an AI agent and the scope of authority its user delegated to it, analogous to know-your-customer checks for humans. It is one of several controls regulators are exploring as agent-initiated transactions grow.