Signal Scanner · FINANCIAL SERVICES & FUTURE OF MONEY · 19 October 2026

Agent Payments Are Being Built Before Anyone Decides Who Pays When the Agent Exceeds Its Mandate

UK and euro-area authorities are opening agent-initiated payments without saying who bears the loss when an AI agent exceeds its owner's mandate. Issuers, acquirers and merchants face unpriced dispute risk through 2027 and 2028.

Produced using Shaping Tomorrow’s six-stage process for converting external change into decision-ready intelligence.
Scanned and filtered by AI. Stress-tested and judged by humans.

The consensus reads agentic payments as a race to build rails. HM Treasury wants the UK to be "at the forefront of the global development of agentic payments" (HM Treasury, 14/07/2026), and the ECB is recruiting firms to test AI agents on the digital euro (ECB, 28/09/2026). Beneath that race, and on the available evidence, no authority has yet said who pays when an agent spends beyond what its owner meant. The UK consultation asked the question and closed on 6 October with no draft rule. Issuers and merchants will settle the first cases themselves during 2027.

Signal Identification

A regulatory gap rather than a pivot: rules written around a customer who authorises each payment still decide who bears an agent's out-of-mandate loss. Clear fraud follows existing token rules (Worldpay, 30/07/2026); the open ground is everything short of fraud, such as an agent misreading an instruction.

Time horizon: 1-3 years, and it binds 2027 to 2028, shown when HM Treasury's consultation response names who bears an agent's out-of-mandate loss, or a card network publishes a dispute rule specific to agent-initiated transactions. (Milestones: UK consultation closed 6 October 2026; ECB applications due 9 November 2026; ECB experiments January to June 2027.)
binds 2027 to 202820262028
Probability band: Medium-High (60 to 80%)
Probability band: Medium-High, 60 to 80%, the chance the signal binds within the window shown above.LowMediumHigh
Geographic / Jurisdictional Scope: Primary: UK and euro area. Spillover: the US and other card-network markets.
PrimaryUKEuro area
SpilloverUSCard-network markets
Sectors exposed:
Retail and commercial banksCard issuersAcquirers and payment service providersOnline merchantsAI agent platform providers

What's Changing

HM Treasury's consultation asks whether provisions on "authentication and consent of payments transactions, and liability for unauthorised payment transactions, need updating" (HM Treasury, 14/07/2026). It concedes that "the PSRs were designed before the development of AI and may not fully facilitate the use of agentic AI" and gives no date for the wider reforms. The ECB's call, with applications due 9 November 2026 and experiments from January to June 2027, covers "AI agents" but excludes "cybersecurity, fraud prevention, compliance, risk management" (ECB, 28/09/2026).

The FSB's final report, which sets out 12 sound practices, asks institutions to "define clear roles and responsibilities" (FSB, 15/10/2026), which governs the firm, not the customer's loss. Market practice is no further on: "No liability shift exists yet", and issuers "tend to side with the cardholder" where no signed mandate exists (Worldpay, 30/07/2026). In the US, "There's no clear legal rule for who bears the loss when an AI tool exceeds its mandate" (Payments Dive, 09/10/2026).

Dates are set for the pilots; none is set for the loss rule

Dated steps so far July 2026 to June 2027 UK consultation 14 Jul to 6 Oct ECB call 28 Sep to 9 Nov ECB experiments Jan to Jun 2027 Who answers who pays Analyst reading of scope HM Treasury Asks the question ECB call Fraud and risk out of scope FSB report Firm governance only Card networks No liability shift yet

Sources: HM Treasury (14/07/2026); ECB (28/09/2026); FSB (15/10/2026); Worldpay (30/07/2026). The right panel is the analyst's reading of each document's scope; the left panel shows the dates the documents state.

Disruption Pathway

Stage one, to mid-2027, is the gap: pilots run while the consultation response is pending, and issuers decide the first disputes case by case. Stage two, 2027 to 2028, is the first written answer, a Treasury response or a network dispute rule, neither dated (HM Treasury, 14/07/2026). Stage three is hardening, when the answer fixes what evidence of a mandate counts.

Stress concentrates in three places: non-fraud disputes where an agent misreads an instruction; refund and reimbursement regimes built for a human payer, where on one reading neither allocates loss to the party that supplied the agent (Bratby Law, 22/07/2026); and the agent provider, which sits outside payment regulation. Two adaptations follow. Operationally, issuers and merchants should record each agent's mandate, limits and action log, since the allocation turns on how consent and limits were recorded at the outset (Bratby Law, 22/07/2026). Financially, acquirers should price agent disputes into merchant terms.

Why This Matters Now

Heads of payments, fraud and disputes, and merchant acquiring own this, with chief risk officers behind them. A dispute reserve, merchant agreement or card product built for human authorisation will be tested by the first agent that buys the wrong thing. Regulation E and Regulation Z turn on whether a transaction was authorised (Payments Dive, 09/10/2026), and the UK consultation has not yet moved that test (A&O Shearman, 17/07/2026). The decision to revise is the evidence standard: what proves a mandate, who holds the record, and who reserves for the gap. Boards should ask for a loss estimate under both a cardholder-favouring and a merchant-favouring allocation, and treat a published rule as the trigger for commitment.

Decision-action posture for this signal: Prepare: mandate records and dispute pricing are worth setting up now, but no capital should be committed until a regulator or network writes the allocation.

Counter-Argument

The strongest objection is that the gap is smaller than it looks. One payments lawyer argues the existing regulations still answer most cases: "Whether the payer consented under regulation 67 determines whether the provider must refund under regulation 76" (Bratby Law, 22/07/2026), and an acquirer says fraud "liability generally follows existing authenticated tokenized-transaction rules" (Worldpay, 30/07/2026). If agents authenticate through tokens, losses fall where today's rules put them.

The lawyer concedes that "What the regulations do not settle is the scope of a delegation exercised by software" (Bratby Law, 22/07/2026), and calls that "an open question rather than a finding that the current rules have ceased to apply" (Bratby Law, 22/07/2026). This report's reading is that disputes short of fraud are where the volume will sit, so the existing rules answer the easy cases and leave the contested ones.

Implications

On the available evidence the gap is durable for at least two years: it persists until the Treasury responds to its consultation or a network writes a rule, and neither is dated (HM Treasury, 14/07/2026). The inflection window runs from the ECB experiments in early 2027 to a first written allocation in 2028. Firms that hold mandate records and have priced the risk gain, because a first rule will most likely reward whoever can prove what the agent was told to do. Firms that wait lose negotiating position.

Early Indicators to Monitor

Disconfirming Signals

Strategic Questions

Keywords

Agentic payments; AI agent liability; unauthorised transactions; Payment Services Regulations; strong customer authentication; dispute resolution; chargebacks; mandate and consent records; digital euro; Financial Stability Board

Bibliography

Source tiers: Tier 1, governments, regulators and intergovernmental bodies. Tier 2, think-tanks, academic institutes, major consultancies and quality data providers. Tier 3, quality journalism and specialist trade press. Tier 4, vendor, company and practitioner sources, used only as directional corroboration.

How this was made

The Decision-Grade Intelligence Method

Shaping Tomorrow’s six-stage process for converting external change into decision-ready intelligence.

Scan · Filter · Stress-test · Judge · Trigger · Actionable Intelligence

See full methodology → · How we verify → · Track record →


Prepared by Shaping Tomorrow: 19 October 2026