What happens when the choice of money is delegated to software?

The original proposition is deliberately provocative: autonomous AI agents tasked with protecting capital might examine UK commercial-bank deposits, see that they are not backed pound-for-pound by gilts, and prefer sovereign or fully reserved alternatives - that is a useful thought experiment. But does it describe how an AI financial agent would actually reason? Commercial-bank deposits are claims on banks, and banks create much of the money used in the economy when they lend. They are not designed to hold a matching gilt for every pound of deposits. Andrew Bailey, Governor of the Bank of England, made the trade-off explicit in July 2026: policymakers could in principle require bank assets to be risk-free, but that would be neither practical nor helpful for an economy that relies on banks to create credit. So, what would an AI optimise? “Sound money” is not an intrinsic machine preference - it depends on the objective function humans give the agent:

· a treasury agent might weigh credit risk, liquidity, yield, transaction costs, legal enforceability, deposit protection, collateral eligibility and settlement finality.

· a payment agent may value programmability and 24/7 availability more than yield.

· an individual’s agent may prefer an FSCS-protected deposit (currently protected up to £120,000 per eligible person per authorised firm) over an uninsured digital asset, even if the latter settles faster.

There is also no public evidence that the leading LLM providers have programmed their models to reject UK commercial-bank money because it is not 100% gilt-backed. Their published architectures suggest a more pragmatic contest between payment rails.

What are the leading AI platforms actually doing?

OpenAI’s Agentic Commerce Protocol, developed with Stripe, allows ChatGPT to help complete purchases whilst merchants remain merchant of record and use existing payment systems. In its initial model, users explicitly confirm transactional steps and payment credentials are constrained to specified merchants and amounts. That looks less akin to abandoning bank money and more like adding AI to established commerce. Google’s Agent Payments Protocol, AP2, goes further in formalising machine authority. It uses cryptographically signed “mandates” to prove what a user authorised and is deliberately payment-agnostic, supporting cards, stablecoins and real-time bank transfers. The architecture asks a more interesting question than which money is safest: can an agent prove it has authority to transact, and can every participant know who is accountable?

Microsoft is addressing the same problem from identity. Entra Agent ID gives AI agents distinct identities, permissions, lifecycle controls and auditable activity. Meanwhile, Anthropic’s work on trustworthy and increasingly autonomous agents, focusing on oversight, co-ordination and unintended action rather than prescribing a preferred form of money. None of these approaches establishes a universal AI preference for sovereign-backed cash equivalents.

Case studies already point to coexistence

In March 2026, Santander and Mastercard completed Europe’s first live end-to-end payment executed by an AI agent inside regulated banking infrastructure: the agent operated within predefined permissions and limits. That matters because it shows existing bank and card rails can accommodate autonomous transactions. And Mastercard’s Agent Pay for Machines widens the picture: designed for high-frequency machine-to-machine commerce, it supports settlement across cards, bank accounts and stablecoins. Visa is pursuing a similar multi-rail strategy, combining agent scoring and registries with stablecoin settlement and tokenisation. If the largest payment networks are preparing for several forms of money to co-exist, why assume AI will converge on only one?

The institutional experiments are equally revealing. BIS Project Agorá has demonstrated atomic cross-border settlement using tokenised commercial-bank deposits alongside tokenised central-bank reserves. Moreover, the Bank of England (which is working with banks on tokenised deposits) has proposed a regime for systemic sterling stablecoins and is supporting DIGIT, the UK’s proposed digital gilt. Rather than commercial-bank money disappearing, could it become programmable?

Could AI nevertheless change bank funding?

This is where the original thesis becomes more powerful when framed as a question rather than a certainty. AI agents may make money more contestable because they can compare alternatives continuously and switch faster than human treasurers or consumers. Imagine a corporate treasury agent that sweeps idle cash every few minutes between insured deposits, tokenised money-market funds, short-dated government securities and stablecoins according to yield, collateral requirements and payment needs. Deposits that once remained sticky, because moving them involved paperwork or operational friction, could become far more mobile. The BIS has warned that widespread stablecoin adoption could materially affect bank funding and credit provision. If deposits migrate to non-bank instruments, banks may have to compete harder for funding or rely more on wholesale markets. That could increase funding costs and, in turn, affect the price or availability of loans. But the opposite path is also possible: tokenised deposits could give commercial-bank money the programmability agents want whilst retaining the regulated banking relationship and the singleness of sterling.

What does this mean for businesses, the economy and individuals?

For businesses, agentic treasury could reduce idle cash, automate supplier payments, execute conditional settlement and make tokenised collateral easier to mobilise. The competitive question may become whether a bank account offers the same machine-readable functionality as an on-chain asset. Banks that expose deposits through programmable, interoperable infrastructure may remain central - those that do not could lose transaction flows even without losing customers entirely. For the wider economy, faster movement of money could release working capital and reduce payment friction, but it could also accelerate liquidity shocks. If thousands of agents respond to the same signal by moving deposits simultaneously, a slow loss of confidence could become a rapid digital run. Regulation may therefore need to consider not only the safety of each instrument but the speed and similarity of automated decision-making. For individuals, AI agents could automatically optimise savings rates, keep balances within deposit-protection limits, route payments cheaply and identify safer counterparties. But who is responsible when an agent moves money into an uninsured stablecoin, tokenised fund or overseas account? Will consumers understand when they have moved from protected money into an investment product? Programmability can improve choice, but it can also make financial decisions invisible.

The more useful question, therefore, may not be whether AI will abandon UK commercial-bank money. It is whether commercial-bank money will remain competitive when software can continuously compare every available form of sterling by risk, return, programmability, liquidity and legal protection. If tokenised deposits can combine the trust framework of banking with 24/7 programmable settlement, AI may not destroy the two-tier monetary system at all. It may force that system to modernise. The Bank of England’s stated aim is a multi-money system in which bank deposits, tokenised deposits and stablecoins remain interoperable at par. The real monetary contest could therefore be won, not by the asset with the purest theoretical backing but by the form of money that machines can verify, use and trust most efficiently.

This article first appeared in Digital Bytes (13th of October, 2026), a weekly newsletter by Jonny Fry of Team Blockchain.

How can digital money reshape the future of UK bank deposits?