AI agents put corporate payment controls under pressure
by Ben Poole
Software is beginning to spend company money. By 2030, Deloitte estimates AI agents could influence $17.5 trillion of commerce, including $15 trillion of transactions between businesses. The same technology is helping fraudsters manufacture identities, impersonate companies and move stolen funds before anyone can stop them.
That collision runs through Deloitte’s ‘Future of payments: Insights for 2027’ report. Its five predictions cover fraud, automated shopping, programmable money, embedded finance and Big Tech’s growing grip on the payment interface. They all land on the same problem for finance: once a machine can choose what to buy and release the payment, its authority has to be defined in terms that software can understand and humans can audit.
The numbers come from a mix of Deloitte forecasts, a May 2026 survey of more than 1,500 US consumers and business leaders, earlier research among 500-plus respondents, and market data from other organisations. Most concern the US. But the questions travel much further.
Fraud provides the report’s sharpest warning. Deloitte expects US losses from authorised push payment scams to approach $15bn by 2028 under its baseline scenario, up from $8.3bn in 2024. Yet autonomous payments are moving quickly into ordinary commercial activity. Companies need to decide what an agent can do before any money moves, and who picks up the loss if it goes wrong.
Fraud starts before the payment does
Generative AI has lowered the cost of building a convincing false identity. Fraudsters can combine stolen data with fabricated documents, faces and voices during account opening. Deloitte expects synthetic identity fraud to cause at least $23bn of losses by 2030.
Payment controls at the end of a transaction can’t carry that burden alone. Deloitte says firms need to check identity and intent throughout the customer relationship. Typing speed, navigation patterns and other session-level signals can help expose an automated or manipulated login. Visa’s pending $2.4bn purchase of behavioural biometrics provider BioCatch shows how much the industry is prepared to spend on this layer of defence.
Now there may be a second actor at the checkout: AI agents buying on the customer’s behalf. A merchant may receive a valid payment credential from software acting for a person or company, without the buyer touching the checkout. The issuer still needs evidence that the agent was appointed and followed valid instructions. Deloitte points to Mastercard and Google’s Verifiable Intent work as one attempt to create that record for merchants and issuers, as well as for any disputes that follow.
Real-time push payments leave almost no time to pull stolen money back. Under Deloitte’s baseline forecast, US authorised push payment fraud losses rise to $14.9bn in 2028. Its conservative scenario reaches $12.4bn. The aggressive case puts losses at $18.2bn. AI-enabled investment fraud, impersonation and romance scams account for much of the projected increase.
The UK makes sending and receiving payment firms share reimbursement costs for eligible victims, with refunds generally due within five business days. US banks don’t face the same general requirement for authorised payments. Even there, research cited by Deloitte found that a bank’s explanation and success in identifying the fraudster affect whether a customer stays.
Confirmation-of-payee checks and stronger login controls are only part of the answer. An agent that can create a supplier, change bank details or approve a payment would concentrate several familiar fraud risks in one place. Its permitted actions need to be narrower than the access available to the employee supervising it.
The customer may be a bot
Automated shopping is already more familiar than the phrase “agentic commerce” suggests. Research cited by Deloitte found that 73% of shoppers use AI somewhere in the shopping journey, and Visa found 47% of US shoppers use it for support such as comparisons and recommendations. Much of that activity still helps a human make the final decision. Fully delegated purchasing is at an earlier stage.
But the numbers get much bigger once agents start completing purchases. Deloitte estimates that AI agents could influence $17.5 trillion of gross merchandise value by 2030, split between $15 trillion in B2B commerce and $2.5 trillion in consumer transactions. Around 30% of digital commerce transaction value could be influenced by agents by then. Half of the businesses in its earlier survey said they would work with autonomous agents to improve efficiency and personalisation.
This changes how a company sells. A product page written for a person may be less useful to software comparing price, delivery, returns and payment terms through an application programming interface. Merchants may need machine-readable catalogues. They’ll also need authentication that admits an authorised buying agent without opening the door to a malicious bot. An agent may rank offers by stated rules instead of habit or brand attachment.
AI use is rising faster than people’s willingness to hand over the wallet. A 2025 measure cited in the report found 70% expressed some comfort with agents making purchases. By June 2026, 55% said they were uncomfortable with that autonomy. Security and data privacy worried 58%, while 57% were concerned about poor decisions or unauthorised action. Another 39% questioned the reliability of the output.
Consumers were also clear about what might win them over. Some 59% wanted reimbursement for AI-related mistakes together with the power to review and override an agent. Another 57% wanted full control over where and how it could act. Real-time notifications and an explanation of the agent’s decisions appealed to 37%.
Money also changes how people feel about handing over control. Those who felt financially secure were about three times as likely to be comfortable using agentic AI in their finances as those who felt insecure, at 32% against 10%. US card balances had reached $1.25 trillion and serious auto-loan delinquency stood at 5.5%, leaving some households little room to absorb an automated mistake.
Corporate buyers will ask the same basic questions at a different scale. Procurement agents need an approved counterparty list and a spending ceiling. Their authority should also expire after a set period. And there must be a route back to a person when price, delivery or contract terms fall outside the mandate. The biggest share of Deloitte’s forecast is B2B, which makes those controls a finance issue well before bot shopping becomes routine for consumers.
The three rail problem
Autonomous agents need a payment method that can keep up with their decisions. Deloitte sees cards, stablecoins and tokenised bank deposits developing as different settlement routes, each suited to a particular kind of transaction.
Cards remain the familiar choice for consumer purchases. Stablecoins can support external payments where round-the-clock settlement or programmability matters. They may also make sense for transactions too small for traditional payment methods. Tokenised deposits offer similar speed inside the banking system while retaining the characteristics of a deposit.
This is already happening in machine-to-machine payments. The x402 protocol lets AI agents pay in stablecoins for computing capacity or API access. It processed more than 178.3 million agent-initiated transactions between its May 2025 launch and the report’s cut-off.
Cross-border payments and automated treasury are already live uses. So are continuous settlement and supplier payments, with escrow and microtransactions developing behind them.
Stablecoin circulation has grown roughly thirtyfold since the start of 2021 to about $300bn. Deloitte projects that US retail payments enabled by stablecoins could rise from $8bn in 2026 to $25bn in 2027 and $65bn in 2028. Its forecast then reaches $135bn in 2029 and $233bn in 2030. Agent-led shopping is expected to be one of the drivers.
For banks, stablecoins and tokenised deposits have very different balance-sheet consequences. Under the US GENIUS Act, stablecoin issuers can’t pay interest and must keep reserves in safe, liquid assets. Funds converted into a stablecoin leave the deposit base against which a bank can lend. A tokenised deposit stays on the balance sheet.
JPMorgan’s Kinexys platform handles about $7bn a day through its tokenised-deposit rail, according to the report. Swift began testing a blockchain-based system with 17 banks in July 2026 to move tokenised deposits between institutions around the clock.
Whatever settles the transaction, the agent first needs a verifiable identity and a defined mandate. In Deloitte’s survey, banks were the most trusted provider for this role, selected by 75% of consumers and 46% of business leaders ahead of any fintech. Both groups wanted approval rights and spending limits. They also wanted a clear answer on who is responsible when something fails.
Treasury should start with one small payment flow
Round-the-clock settlement changes the working day for treasury. Payments can leave outside bank operating hours, and machine-triggered transactions may occur far more often than manually initiated ones. Cash positions and funding rules have to follow the same clock. So does exception monitoring.
Deloitte’s advice is deliberately small to begin with: “Corporate treasury should pilot one bounded, agent-initiated flow now, before volume makes mistakes expensive.” A recurring payment to an established supplier, with a low limit and a clear human escalation route, would show whether the permissions work and whether treasury can reconcile the result. Any gaps in the underlying data would surface quickly too.
Then there is the cash sitting idle on a 24/7 rail. Stablecoins don’t earn a return under the US regime described in the report. Deloitte suggests testing automated sweeps into a tokenised money market fund or another yield-bearing product, then returning the money when cash is required.
Treasury would still need separate controls for the product and its custodian. The authority granted to the agent executing the sweep can’t be assumed either.
Segregation of duties matters even more when one system can do so much. A system that selects a supplier, negotiates terms and pays the invoice shouldn’t validate every step with the same unchecked instruction. An independent record must show who set the rule and what the software decided. It should also explain why the payment passed its controls.
Embedded finance can leave responsibility blurred
Synapse’s 2024 collapse supplies the report’s clearest warning about divided responsibility. More than 100,000 people lost access to over $265m held through fintech platforms after the intermediary’s records diverged from balances at partner banks. Federal deposit insurance would have applied if a partner bank had failed. It didn’t cover the collapse of the intermediary routing the money.
The sums moving through these arrangements are getting larger. Deloitte cites a forecast that US embedded-finance revenue will reach $51bn in 2026, up from $40bn in 2025, on about $7 trillion of transaction volume.
Money transmission is still governed through separate regimes in 49 states and Washington, DC. California added its Digital Financial Assets Law in July 2026 as the federal Consumer Financial Protection Bureau withdrew guidance and proposed narrower supervision of nonbanks.
The UK draws a clearer line of responsibility. An authorised principal remains responsible for regulated activity carried out by an appointed representative. The framework gives regulators a more obvious firm to hold accountable.
Some fintechs are seeking bank charters. Revolut received conditional US approval in September 2026, with four requested products still under review. Nubank’s conditional approval took 121 days. With a charter, the fintech takes settlement and reconciliation liability onto its own books instead of leaving it between itself and a partner bank.
Any company using an embedded payment service needs similar clarity in its contracts. It should know which entity holds the funds and whose ledger is authoritative. The contract also needs to say who must restore access after a reconciliation break.
Deloitte calls for independent, real-time reconciliation with fintech partners and liability agreed before onboarding. A glossy interface can’t answer those questions after money has gone missing.
Big Tech wants the moment of approval
Big Tech’s advantage sits at the front of the transaction. Technology groups already control wallets and consumer devices. Many also provide identity services. The report says digital wallets account for 83% of digital payments. That gives the platforms the behavioural data and customer access needed to decide which payment choice appears first.
Attempts to become the financier have proved less durable. Technology companies have launched cards and buy now, pay later services. They’ve also experimented with cryptocurrencies. Several of those products have since been closed or reworked with established providers.
Wallets and digital identity have fitted better. Google Wallet’s move into digital driving licences leaves public authorities with the regulatory role while the technology company owns the interface.
Deloitte expects the next contest to centre on verification. A platform that confirms an AI agent’s identity and authority before the request reaches a bank could become the gatekeeper for agentic commerce. Banks and payment networks would still process the money and carry regulated risk. But they could lose more of the customer interaction to the company that approves the agent.
Merchants may gain another sales channel, though with less control over discovery and customer data. Payment routing could also move further into the platform. Banks do retain one powerful advantage: customers still trust them. That leaves a role as the regulated infrastructure behind the interface.
Whoever defines the authorisation standard will have considerable influence over this market. They’ll also need to prove what the agent was told to do when a transaction is challenged.
By 2027, moving the money may be the least difficult part. Companies still need to establish who instructed the agent and how far its authority extended. Then there is the question of which participant carries the loss. Without those answers, instant settlement could simply make a control failure arrive faster
Like this item? Get our Weekly Update newsletter. Subscribe today
