IRS Risk Chief Warns Regulators Cannot Match AI Agent Payment Speed

Dottie Romo, the Chief Risk and Control Officer at the US Internal Revenue Service, told a New York audience that regulators cannot match the speed of AI systems already moving money. Speaking at United Nations Headquarters during the Future of Money, Governance and the Law summit, Romo said oversight still runs at a pace that is not sustainable, and that some form of automated supervision will be needed. The warning lands at an odd moment, because machine-to-machine payments are still tiny in dollar terms yet already arrive in the millions.

Her argument in one line: regulators may soon need algorithms to supervise algorithms.

What the IRS risk chief said in New York

Romo joined a panel alongside Dino Cataldo Dell'Accio, Deputy Chief Executive of Pension Administration at the UN Joint Staff Pension Fund, and Julius Moye of Mastercard's financial crime solutions team. The session sat inside a summit run by the Government Blockchain Association and launched joint research on how AI, blockchain and quantum computing intersect with financial services. Romo's comments came at the close of a discussion that had already moved from encryption to transaction speed.

Her core claim was about tempo, not novelty. Regulators, she said, still lean heavily on periodic reports to surface fraud, control failures and other risks. That model was built for markets where filings arrive on a schedule people can read. Nobody in the room disputed that machines no longer work to any such schedule.

"Right now we're reviewing things at a very traditional pace," she said. "That's not going to be sustainable."

Why do regulators need algorithms watching algorithms?

Because oversight that reads periodic documents cannot catch fraud that machines commit in seconds. Romo said some form of automated supervision is now close to unavoidable, which in practice means algorithms monitoring other algorithms. "We're not going to be able to do that in a fast enough pace," she said of human-only review.

Her proposed shape is not full autonomy. Romo argued for more real-time monitoring across markets while keeping people in the loop on the decisions that matter. The distinction matters for anyone tracking compliance budgets, since detection shifts toward software while judgment stays with people, and the headcount argument stops being the main cost driver.

This is the uncomfortable part of the message. A regulator that wants continuous coverage has to accept continuous, machine-scale collection, which raises its own questions about data quality, model governance and who audits the auditors. The same official warning about machine speed was also sketching the infrastructure needed to answer it.

How large are AI agent payments right now?

Smaller in value than most people assume, larger in count than most oversight models were built for. Research published alongside the summit tracked 6.4 million x402 payment transactions carrying $119,947 in total between July 23 and August 26, across the Base and Solana networks. That works out to roughly two cents per transaction on average.

Volume, not value, is the number that deserves attention.

Agents are mostly buying individual machine inputs, such as data, API access or computing tasks. Each purchase is small enough to feel irrelevant to a human reviewer, which is exactly why aggregate behaviour looks nothing like the periodic reporting Romo described. Someone sampling a quarter of paperwork would never see 6.4 million micro-transactions moving through settlement rails in five weeks.

Why do the payments come out in fractions of a cent?

Because agents are buying metered digital resources rather than goods with a shelf life. In the tracked sample, 90.8 percent of transfers were worth less than one cent, while the AI agent involved recorded nearly 200 million settlement transactions since launch. Read that ratio carefully. A per-transaction fee is a rail design choice, and for machine calls it stops fees from swamping the underlying data or compute purchase.

Per-transaction fees are a rail design choice, not an accident.

The ratio also changes how risk scales. A loss that would be trivial on a human account becomes material once multiplied by hundreds of millions of executions, and each individual execution is small enough to fall under manual review thresholds. That is the part auditing models were never designed around.

What has the market already learned from automated blowups?

That unguarded automation fails in ways people remember. Mastercard's Moye pointed to the 2012 Knight Capital trading disaster and the Terra/Luna collapse as warnings about automated systems running without sufficient safeguards. Both turned a design or coding assumption into losses in the hundreds of millions, and both hardened the argument that speed can outrun controls.

Neither event involved an AI agent paying fractions of a cent for data. The control lesson still transfers.

The control precedent is older than the technology.

There is a quieter precedent already running at scale. Large institutional investors review enormous transaction volumes for anomalies every day, which is why monitoring vendors sell continuous surveillance rather than annual reviews. The UN pension fund made a related point about cryptography at the same summit, arguing there is no clean switchover date and that encryption will require continuous monitoring as threats evolve.

Who gets the kill switch when code moves money?

Humans, at least on the decisions that count. Moye described a layered model in which machines detect unusual behaviour and automatically contain the problem, with serious incidents escalating to people. "It's really using AI and machines to apply the tourniquet and stop the bleeding and then have humans come in to do the surgery," he said.

Dell'Accio pushed on the other half of the problem, arguing that automation cannot erase responsibility. He said accountability should trace back through three questions: who developed the code, who implemented the code, and who oversees the code. None of those questions mention the model, and all three have a human name attached.

That framing lands directly on wallets, agents and protocol teams. If a coded agent can move funds without a human approving each step, then the audit trail becomes the control itself. Counterparties and supervisors will want a named owner for the code, a documented change process, and a way to halt execution that does not depend on the agent's own cooperation.

What should traders and policy watchers track next?

Four items stand out. First, whether real-time monitoring language turns up in actual supervisory guidance rather than conference rooms. Second, whether agent payment volumes keep compounding while average ticket sizes stay small, which is the combination that would turn a rounding error into a systemic channel. Third, how compliance vendors price continuous surveillance for institutions that currently budget around periodic review.

None of it is a tradable catalyst on its own.

Fourth is the piece crypto markets would feel first: stablecoins and API-based rails becoming standard plumbing for autonomous software. If machine payments settle on chains like Base and Solana rather than card networks, the addressable volume for tokenized settlement grows without any retail user changing behaviour.