What Money Becomes: Confessions of a Head of AI at a Global Bank

Imagine a phone on a bedside table at half past three in the morning, Nairobi time. The screen shows one word from a colleague on another continent, then a line nobody wants to read twice. Suspended.

In my world that word is usually followed by a quiet hour of forms. This time it was an exchange's word, and the exchange was Binance. According to Euronews, deposits and withdrawals were suspended at about 00:30 GMT on Friday, after the company announced a hack. The estimated impact was around $6.3 billion, or €5.5 billion.

I should say the awkward thing now. I am a composite, and so is the bank. The bank is not named, the risk committee I describe is partly imagined, and the narrator is stitched together from several rooms I have sat in and several late nights I would rather not itemise. Where I report a fact, it carries a link. Where I offer a judgment, I say so. Where a scene did not happen in quite this form, I write "Imagine." Keep those three registers apart. I have watched them blur in slide decks for a career.

Six point three billion dollars does not fit inside a normal bank's sense of scale. Imagine the question on every screen that morning: not how to recover, but what, exactly, had been taken.

Authority Is the Real Key

Start with the picture most people in my industry carry. When they hear "custody," they picture a vault and a key. The vault is the bank's stone. The key is the thing you guard. That picture is reassuring and, after this month, incomplete.

Euronews reports that Binance's initial finding pointed to a compromised laptop belonging to a DevOps engineer. The unauthorised party is said to have reached a service container and then executed withdrawal transactions. Binance, as quoted by the same outlet, said it had "found no evidence of any private key leaks from wallets." I take that sentence seriously, and I read it the way I read every statement from an incident team on day one: as an early one. Binance has promised a detailed technical analysis. Until it arrives, the early account is what we have.

The lesson survives the specifics. A private key is a secret. An authority is a process. If the system that decides what gets signed, when, and for how much is compromised, the secrecy of the key does very little for you. In my reading, the network was built to check that the ink is real, not to ask who is holding the pen.

Attribution is contested, so I will be exact. As Euronews reports, TRM Labs has said it has "high confidence" that TraderTraitor, a group it links to North Korea, orchestrated the attack. That is an assessment by an analytics firm, not a court finding. Binance has made no attribution of its own in the reporting. Repeating an attribution as settled fact is a risk a bank should not take.

The question at three in the morning is never whether the money is safe. It is who was allowed to ask for it.

Money Was Never Only the Bank's

I have stopped believing that money lives in banks. The story suits the people who run banks, because it flatters them. The truth is messier and older.

Consider M-Pesa, the mobile money service Safaricom runs in Kenya. Its everyday use is the clearest example I know of money that never needed a branch. Imagine a vegetable seller at a Nairobi kiosk, taking a payment, sending a receipt, and never once thinking about a correspondent bank. Her money is already digital, already shared, and already trusted by many people who never opened an account at a bank branch. In my judgment, whatever "money" means in 2030, it will not be a rule that only banks may keep ledgers.

Now the newer layer. In October, JPMorgan's Kinexys said it had brought its deposit token, JPMD, live on Ethereum, available to J.P. Morgan's institutional clients. This is the distinction I most want non-specialists to understand. A deposit token is a bank's own liability. It moves on a public-style ledger, but it is still a claim on the bank that issued it. A stablecoin is a privately issued promise. Both can be fast. They are not the same thing, and the difference becomes very important on the morning one of them stops behaving like money.

That morning is what the Bank for International Settlements has been circling. Its 2026 Annual Economic Report, in a chapter titled "Anchoring trust in money: innovation beyond stablecoins," says that stablecoins "by construction, they cannot currently ensure exchange at par across issuers and blockchains under all conditions." Par is the whole game. A coin worth a dollar on Tuesday is a fine thing. Suppose the same coin is worth ninety cents on Wednesday, because a bridge jammed or a reserve was slow. That is a different instrument with a different name. The BIS paper also proposes a unified ledger, in which central bank reserves, tokenised commercial bank money and other supervised private money could sit together. It says such a ledger could uphold the attributes of money if the safeguards are strong enough. I agree with that sentence in a way that is not comfortable. My industry's future looks less like a fight between banks and crypto, and more like a plumbing negotiation nobody asked for.

Rules Arrive in Layers

Paperwork comes next, and it is where much of finance's money is quietly decided.

In the United States, the Office of the Comptroller of the Currency issued its final rule under the GENIUS Act on 21 October 2026. According to Sidley Austin's summary, the rule and the Act take effect on 18 January 2027. The sentence I keep pinned above my desk comes from the rule, as Sidley quotes it: a permitted issuer "must ensure that reserve asset value equals or exceeds outstanding issuance value at all times." Read it twice. "At all times" is a phrase that will make an engineer and a treasurer hate each other.

Then there is model risk, the discipline I am paid to run. For fifteen years the American benchmark was SR 11-7. On 17 April 2026, the Federal Reserve and its partners issued a new supervisory letter, SR 26-2, which supersedes SR 11-7 and is framed on a risk basis. The letter itself does not mention generative or agentic AI, and this piece does not claim the guidance covers them. My instinct is that regulators will write the agent rules after the agents have started paying bills, because that is how rules have usually arrived in finance.

Agents Who Pay

Agents that pay are the idea I was most eager to believe, and the one I am now most suspicious of.

In April 2026 Visa launched Agentic Ready, a programme that lets banks and payment partners test agent-initiated payments with live cards and real merchants. Visa's own release says it was live with more than 20 partners in the UK and Europe, with a rollout to more than 85 partners across Asia Pacific and Latin America planned. That is an announcement, and I have sat in enough launch rooms to know what an announcement is worth on a Tuesday. The question that matters for a bank is not whether a demo works. It is what happens when an agent is wrong or generous, the way a human employee can be, only faster and at three in the morning.

Imagine a risk committee at two o'clock in the morning, after a weekend of nervous conversations about an incident on another exchange. The chair asks a simple question: what is the maximum an autonomous agent may authorise, to whom, and who can revoke that authority in under a minute? Nobody in the room has a clean answer. Someone proposes a threshold. Someone else asks whether the threshold is attached to the agent, to the customer, or to the network token that carries the credential. We spend forty minutes discovering that our identity model was designed for people who sleep.

This forecast is composite opinion, extrapolated by eye from Visa's Agentic Ready testing programme described above. No model produced it, and nothing has calibrated it. By 2030 I would put roughly even odds that a meaningful share of routine business-to-business and card-not-present payments at large banks will be initiated by software acting for a person or a company. The bank's central question will change from "who is this customer?" to "which agent holds this mandate, with what limit, and who can switch it off?" I would not trust a bank that answers the second question with a policy document alone.

The Voice on the Phone

Fraud is where the agent meets the telephone. The most convincing thing in a fraudster's kit is now a human voice.

Vendors publish fraud statistics constantly, and many of them are surveys, samples or market estimates. Surveys of expectations are not measured losses, and a market-size estimate measures something else again. I have stopped quoting any fraud figure without its method attached, and I would ask you to do the same. Losses are harder to pin down than the headlines suggest, because counting reported cases and estimating a market are different exercises.

In my judgment, the attack that matters most for a mobile-first economy is the one that arrives by phone, in a familiar accent, asking for a small urgent transfer. Such an attack is built to be absorbed by routine, right up until the routine stops working.

Imagine a fraud analyst in a Nairobi operations room with headphones on, listening to a recording of what sounds like a regional director asking for a same-day wire. The voice is right. The cadence is right. The only wrong thing is the context, and the analyst has about ninety seconds to decide whether context is enough. The bank that trains that analyst to ask for a second channel, a callback to a number already on file, and a human who can say no, is doing something no model will do for it.

A Forecast, Held Loosely

The forecast comes last. I will keep it short, because forecasts are cheap at my level and expensive for the people who believe them.

My judgment, uncalibrated, is that the models will converge. Within a few years, the foundation models that large banks rent will be close enough in capability that the model itself becomes a commodity line item, bought in bulk and swapped without ceremony. What will not be commoditised is the data a bank holds about what its customers actually do, the trust earned by years of paying people correctly, and the relationship that tells a bank which phone number belongs to whom. In that world the advantage is not the cleverest model on the desk. It is the owner of the customer's mandate, the record of the customer's behaviour, and the permission to say "wait, let me call you back."

I believe money in 2030 will be more programmable, more tokenised, more regulated in layers, and more often initiated by something other than a person. I also believe trust will be the scarce commodity, because a payment can now be perfectly formed and perfectly wrong. The Binance episode did not prove that crypto is doomed, or that banks are safe. One incident cannot establish a general rule. I take it to suggest that the people and systems allowed to ask for a signature are as valuable as the keys, and as dangerous.

The habit I have tried to build in my own team is the only advice I am certain of. When an unexpected request comes in, however real it sounds, hang up and call back on the number you already have. Make the machine explain why its mandate is valid. Keep a human on the phone who is allowed to refuse.

Imagine it is still half past three. The phone lights up again with the same word, now followed by a question mark. I hang up, and I dial the number already on file.

Editorial note: This essay expresses the founder’s ideas, developed with AI-assisted research and writing (parallel Claude Haiku 5.5 agents) and checked against the linked sources on 30 October 2026. Convictions, proposals and forecasts are identified in the text; scenes framed as imagined are illustrative. The narrator is an imagined composite voice, not a real person or any named bank; facts and sources are cited as they are in the text.

Cover photograph / illustration: Axel Tschentscher. CC BY-SA 4.0. Original source. Illustrative image: the Financial District of Lower Manhattan at night, seen from Brooklyn Bridge Park. Photograph: Axel Tschentscher, via Wikimedia Commons. CC BY-SA 4.0. This image does not document the events described.