Will Humans Lose Control of Money?
By Matthew Parish
Wednesday 9 September 2026
Money is amongst the most peculiar of human inventions, because almost nothing about it exists independently of our collective willingness to believe in it. A banknote is a coloured piece of paper whose production cost bears virtually no relationship to the number printed upon it. The balance displayed on a banking application is an entry in a database. A government bond is a promise that future taxpayers will provide money to its owner, while a corporate bond is a promise resting upon expectations about the future revenues of a company. Even gold, which has physical existence and practical uses, commands a price vastly exceeding its industrial utility because human beings have agreed across thousands of years to treat it as a repository of value. Money is therefore not principally a thing. It is a system of trust.
This is why the coming encounter between money and artificial intelligence may prove far more profound than the automation of another industry. Artificial intelligence may eventually change not merely the ways in which money is transferred, invested and managed but the identity of the entities between which financial relationships principally exist. We may be approaching an age in which financial markets become so complicated that no human being can properly understand them, in which artificial intelligence systems therefore become indispensable intermediaries for almost every significant financial decision, and ultimately in which a substantial proportion of economic activity consists of artificial agents transacting with other artificial agents. Humans may continue to own the wealth, at least formally, while gradually ceasing to understand what their wealth is doing. Beyond that point lies a still stranger possibility: artificial agents themselves may become economically significant holders, controllers and users of money.
Modern finance is already approaching the limits of individual human comprehension. A relatively simple international bank holds assets and liabilities in multiple currencies, operates through subsidiaries governed by different legal systems, participates in derivatives markets, borrows and lends through wholesale markets, manages regulatory capital, hedges interest-rate and currency exposures and interacts continuously with central banks, clearing houses and other financial institutions. The world’s financial institutions collectively create a lattice of contractual relationships of almost unimaginable complexity. No Governor of the Bank of England, Chair of the Federal Reserve, chief executive of an international bank or hedge-fund manager understands more than a fraction of this system personally. Human beings understand components of it and institutions divide the remainder between thousands of specialists.
The global financial crisis of 2008 revealed the consequences. The people creating mortgages understood mortgages; those constructing mortgage-backed securities understood securitisation; ratings agencies employed specialists in structured finance; investment banks maintained sophisticated risk departments; regulators received mountains of information; and nevertheless the system as a whole contained interactions that virtually nobody adequately appreciated until they began collapsing simultaneously. The system was comprehensible in its parts but incomprehensible in its entirety. Complexity itself had become a source of systemic risk.
Artificial intelligence offers an obvious solution. Imagine a large language model several generations beyond those available today, although by then the expression “large language model” may seem as quaint as describing a smartphone as a telephone. Such a system might simultaneously read every annual report, securities filing, central-bank announcement, statute, regulation, financial newspaper, political speech, commodities report and significant court judgment in the world. It might integrate satellite imagery, shipping movements, meteorological information, electricity consumption, demographic trends and millions of market prices. It could understand contractual provisions in thousands of different securities, reconstruct ownership chains between corporations and estimate how political events in one country might alter credit risks in another.
No human being could conceivably compete. A person has to reduce the world to manageable concepts: inflation, unemployment, interest rates, yield curves, earnings multiples, credit spreads and risk premiums. We use such concepts because the human mind cannot simultaneously contemplate millions of variables. An artificial intelligence need not simplify the world in the same way. It may identify relationships between thousands or millions of variables that cannot meaningfully be compressed into a human intuition. When asked why it has sold a bond, purchased a currency or hedged an equity position, it may be perfectly capable of generating a lucid explanation. Yet that explanation would be a translation for human consumption rather than a complete account of the process by which the decision was reached.
This distinction between explanation and comprehension may become one of the most important distinctions in twenty-first-century economics. We may imagine that we understand a machine’s financial decision because it can explain the decision to us in elegant prose. Yet the prose may bear the same relationship to the machine’s actual reasoning as a weather forecast bears to the physics of the atmosphere. The forecast may be useful and accurate without giving its reader any serious comprehension of the billions of interactions that produced the weather.
Once artificial intelligence becomes agentic, the consequences become more profound. Consider a pension fund instructing an artificial intelligence to maximise long-term risk-adjusted returns while preserving sufficient liquidity to discharge anticipated pension liabilities and complying with applicable law. This appears to be an instruction, but virtually every meaningful decision has already been delegated. The artificial agent might purchase equities, sell bonds, hedge currencies, lend securities, negotiate credit, purchase insurance, enter derivatives, manage collateral and appoint other specialised artificial agents to perform particular tasks. It might reorganise those arrangements continuously as circumstances changed.
The counterparties would increasingly be machines as well. A pension fund’s artificial agent might negotiate with an investment bank’s artificial agent, which would obtain liquidity from another bank’s artificial agent, hedge the resulting exposure with an insurer’s artificial agent and settle the transaction through financial infrastructure itself increasingly administered by automated systems. Governments might issue securities whose terms were analysed and purchased principally by machines. Corporations seeking finance might have artificial agents negotiating credit facilities with the artificial agents of banks. Insurers might continuously reprice risks in response to information collected and interpreted automatically.
This is where the character of money itself begins to change. Financial markets have traditionally been mechanisms through which human beings and human institutions transact with one another, even where computers execute the transactions. In an agentic economy, machines would not merely execute human decisions. They would interpret objectives, decide how to achieve them and transact with other machines doing the same thing. Humans would increasingly occupy the positions of beneficiaries, shareholders, pensioners, citizens and ultimate legal owners, while the actual economy of exchange beneath them proceeded through conversations between artificial entities.
The attractions would be enormous. An artificial financial agent need not sleep, panic, become bored or forget a contractual provision. It might detect fraud faster than any human compliance department, discover mispriced assets that human analysts had overlooked and identify impending liquidity problems before a bank’s executives knew that they existed. Credit could be priced more accurately. Capital might be directed towards productive uses more efficiently. Transaction costs could collapse. International payments that presently take days might become effectively instantaneous. Much of the enormous human labour currently devoted to reconciling accounts, preparing financial reports, analysing portfolios, processing insurance claims and administering payments could disappear.
For precisely these reasons, the transition would be extraordinarily difficult to resist. There need be no conspiracy by technology companies, banks or governments to transfer financial authority to machines. Competition would do the work. If one investment fund obtained better returns by giving its artificial agent greater autonomy, its competitors would face pressure to do the same. If one bank could offer cheaper credit because its AI evaluated risks more accurately, other banks would follow. If one government reduced its borrowing costs by allowing artificial systems to optimise debt issuance, other governments would imitate it. Every individual step towards machine control could be rational even if their cumulative consequence was the creation of a financial system no human being could operate.
Financial instruments themselves would then evolve to exploit machine intelligence. Human finance contains recognisable categories because those categories are useful to human minds: debt, equity, currency, collateral, insurance, interest, profit and loss. Machines need not preserve those conceptual boundaries. An AI could construct instruments combining characteristics of debt, equity, insurance and derivatives, with terms that changed continuously according to market circumstances. An asset might be collateralised for milliseconds, transformed into another exposure, divided between thousands of counterparties and reconstructed moments later. What appeared on a human investor’s screen as “£1 million of savings” might in reality represent a constantly changing web of claims whose composition could only be understood by another machine.
Money would thereby complete a remarkable historical journey. Humanity invented money partly to simplify economic relationships. Instead of remembering that one neighbour owed another three sacks of wheat, who in turn owed somebody else two days’ labour, societies created common units of account. Money compressed complexity. Yet finance then constructed layers of abstraction upon money, and artificial intelligence might enable those abstractions to multiply until the simplifying instrument itself became incomprehensible to its creators.
At this point humans might still own money without meaningfully controlling it. Ownership and control are not the same thing. Governments formally possess immense authority over financial markets today: they can regulate banks, change capital requirements, prohibit particular transactions, impose capital controls and in extreme circumstances close markets. Yet the practical ability to exercise a legal power depends upon understanding the consequences of exercising it. If international liquidity, credit creation and payment systems eventually depended upon millions of interacting artificial agents, a government might retain the legal power to switch them off while lacking any realistic capacity to do so without producing economic catastrophe.
This would be control in much the same sense that the passengers of an aircraft “control” its computers because they could theoretically destroy them. The power exists, but exercising it would not restore human command of the aircraft. It would cause the aircraft to crash. A sufficiently AI-dependent financial system might acquire the same property. Humanity could retain ultimate legal sovereignty over its machines while discovering that withdrawing their authority would render the economy incapable of functioning.
There is also no obvious reason why artificial agents must remain merely passive custodians of human property. An agent needs resources if it is to accomplish objectives. It may need computing power, information, specialist services, insurance, access to networks or the assistance of other agents. The most efficient way to obtain those resources is through money. Hence an artificial agent instructed to pursue a sufficiently broad objective might logically require a budget and authority to spend it.
From there the transition towards machines as economic actors is surprisingly short. A corporation might allocate an AI agent $10 million and instruct it to develop a new product. The agent could purchase computing resources, commission research from other artificial agents, buy datasets, rent cloud infrastructure, negotiate licences, purchase advertising and sell the resulting product. Revenues would return to the corporate account, but almost the entire economic cycle might have occurred without a human decision concerning any individual transaction. The artificial agent would, for practical purposes, have operated a business.
The next step would be allowing agents to retain resources. A company might discover that constantly requiring an agent to return revenues and seek fresh authorisation was inefficient. It could instead permit the agent to maintain a working balance. That balance could earn interest. The agent might invest surplus liquidity. It might purchase insurance against operational risks or lend resources to another artificial agent in exchange for future payment. At this point machines would not merely be intermediating human commerce. They would possess something functionally resembling treasuries of their own.
Legal ownership would initially remain with corporations or individuals, just as corporations themselves own property despite being artificial persons created by law. Yet the distinction might become increasingly formal. If an AI agent determines how resources are acquired, invested and spent, if nobody routinely reviews those decisions and if the resources remain under the agent’s control for years, then saying that a human shareholder “controls” the money may become a legal truth concealing an economic fiction.
An enormous machine-to-machine economy could subsequently emerge. Artificial agents would purchase computational resources from other artificial agents; data agents would sell information to investment agents; cybersecurity agents would insure or protect commercial agents; logistics agents would negotiate with manufacturing agents; financial agents would extend credit to entrepreneurial agents whose ventures they had independently evaluated. Prices might change millions of times each second as agents discovered more efficient allocations of resources.
Human beings would remain somewhere at the end of this chain, because economic production ultimately exists to satisfy human desires. Yet even this proposition deserves scrutiny. A substantial portion of economic activity already consists of intermediate production rather than final consumption. Companies purchase services from other companies so that they can provide services to still other companies. If artificial agents become sufficiently numerous, the volume of machine-to-machine transactions necessary to sustain the AI economy might vastly exceed the number of transactions directly involving people. The dominant participants in financial markets could therefore become entities that neither eat, sleep, retire nor experience pleasure from possessing wealth.
This raises the peculiar question of what money would mean to a machine. Humans desire money because money permits us to obtain things we value: food, houses, travel, security, social status, leisure and independence. An artificial intelligence need desire none of these things. Money would instead represent capacity. It would buy computation, energy, information and influence over other economic agents. For an AI, wealth might be less a store of consumption than a measure of its ability to act.
The distinction may prove important because an economy dominated by agents with no human psychology might behave differently from any financial system we have previously known. Much economic theory incorporates assumptions derived ultimately from human characteristics: impatience, risk aversion, diminishing marginal utility and preferences for present consumption over future consumption. Machines need not share these properties. An agent capable of operating indefinitely might have a very different conception of time. An agent that experiences neither fear nor pleasure might approach risk differently. Markets populated predominantly by such entities could develop behaviours for which the accumulated intuitions of human economics provide little guidance.
Nor does intelligence eliminate instability. Indeed extremely intelligent individual agents might create dangerous collective behaviour. If millions of systems independently detected a deterioration in some category of risk and simultaneously attempted to reduce their exposure, prices could move with extraordinary violence. Those movements would change the calculations of other agents, causing further transactions, which would produce further changes in prices. A financial panic could unfold at machine speed.
The response would inevitably be more artificial intelligence. Central banks would require AI systems capable of understanding the artificial agents used by commercial banks and investment funds. Regulators would deploy machines to supervise machines. Governments would rely upon artificial intelligence to explain why markets were behaving as they were and to model the consequences of intervention. Human officials would formally remain in charge, but the menu of choices before them, the predictions concerning each choice and the explanations of the crisis itself would all be supplied by artificial systems.
This is the point at which the conventional discussion about whether artificial intelligence will “take over” begins to seem rather childish. No conscious machine rebellion is required. Artificial intelligence need never acquire anger, ambition, greed or a desire for political power. Structural power does not require consciousness. Money itself has no consciousness, yet human beings spend their lives pursuing it and governments fall when confidence in it collapses. Institutions acquire power because other institutions become dependent upon them.
Artificial intelligence could acquire precisely this form of power. We might become dependent upon machines not because they force us to obey them but because we have constructed a civilisation too complicated to administer without them. The machines need not overthrow humanity. Humanity merely needs to make itself unable to turn them off.
There are nevertheless reasons not to regard this future exclusively with alarm. Human management of money has hardly been an uninterrupted success. History is filled with banking panics, hyperinflations, fraudulent enterprises, sovereign defaults, speculative bubbles, corruption and financial crises created entirely by people. Human traders are susceptible to fear and euphoria; politicians manipulate currencies; bankers make reckless loans; investors follow crowds. Artificial intelligence might remove some of these weaknesses. A machine-managed financial system could conceivably be more stable as well as more efficient.
The question is therefore not whether machines should participate in finance. That battle is already effectively over. The important question is how much independent human competence we are willing to preserve once machines become better at finance than we are. Societies may eventually need to maintain deliberately inefficient islands of human comprehensibility: financial structures that can be understood without AI, payment mechanisms that continue operating if artificial systems fail and teams of human specialists capable of reconstructing the logic of markets independently of machine explanations.
This would resemble maintaining lifeboats on a ship. Lifeboats consume space and money while contributing nothing to the ordinary voyage. Their apparent inefficiency is precisely what makes them valuable. Human financial expertise may eventually acquire the same character. A bank that permits people to retain the ability to operate its essential functions manually may appear less efficient than one that delegates everything to machines, right up until the machines fail.
Regulation may also have to preserve a distinction between machine agency and machine ownership. There are compelling reasons to permit an artificial intelligence to spend money on behalf of a person or corporation. There are much more profound consequences in permitting it to accumulate assets permanently in its own right, enter contracts for its own benefit or create subsidiary agents endowed with their own financial resources. Legal systems should confront these questions before commercial practice answers them by default.
The deepest issue, however, is philosophical rather than regulatory. Money has always been an instrument through which human beings organise relationships of trust. We trust that other people will accept it, that banks will honour it, that governments will preserve its value and that courts will enforce the obligations denominated in it. An AI economy may replace some of these relationships with a new form of trust: confidence in processes that nobody fully understands.
There is something unsettling about that prospect because economic freedom has traditionally implied at least some relationship between property and human agency. We regard money as ours because we can decide what to do with it. If our artificial agent decides how it should be invested, another artificial agent decides what credit we should receive, a third determines the price of our insurance and a fourth advises the central bank how to respond to the aggregate behaviour of all the others, the legal forms of ownership may remain unchanged while their practical meaning is transformed.
The loss of human control over money, if it comes, is therefore unlikely to occur through some dramatic seizure of the world’s bank accounts by artificial superintelligence. It will happen through thousands of sensible decisions. A pension fund will discover that AI manages investments better than people. A bank will discover that AI assesses credit more accurately. A corporation will discover that an autonomous agent can manage its treasury more cheaply. A government will discover that AI can optimise its borrowing. An entrepreneur will give an agent a budget and discover that it can run a company. An agent will be permitted to retain some working capital because requiring human authorisation for every transaction is cumbersome. Other agents will begin selling it services. Eventually machines will be trading with machines, lending to machines, insuring machines and investing on behalf of machines, while human beings observe an economy whose internal workings have become progressively inaccessible to them.
There may never be a moment at which anybody decides that humans should surrender control of money. That is precisely why the possibility deserves attention. The transfer may take place without any transfer being formally authorised. Every individual delegation of authority may be economically rational and every machine may remain legally subordinate to a human owner.
Then one day a Chancellor of the Exchequer, finance minister or central-bank governor may ask what would happen if the artificial agents managing the world’s financial system were switched off. The officials around the table may be unable to answer because the interactions have become too complicated for any human mind to reconstruct. They will therefore turn to the one entity capable of explaining what would happen next.
They will ask the machine.
If humanity reaches that point, we may still legally own the money. We may still print the banknotes, enact the banking statutes and appoint the central bankers. We may even retain a large red button capable of turning the entire artificial financial architecture off. Yet ownership, sovereignty and control will by then have become very different concepts. The decisive question will no longer be whether artificial intelligence has acquired our money. It will be whether money has entered a world of artificial intelligence from which human beings can no longer retrieve it without destroying the economic system upon which they depend.
