The prediction: neither Visa nor Mastercard is likely to publish a binding, generally applicable dispute and chargeback rule for agent-initiated card transactions with an effective date before the 2026 holiday peak. The pattern in the last month points instead to a first binding network rule landing no earlier than the April 2027 release cycle, with the October 2026 releases confined to identity, token and authentication plumbing. For merchants, that implies one concrete thing: agent-initiated disputes over Black Friday and Cyber Monday 2026 will be adjudicated under rules written for a human pressing a buy button.
In short
- The prediction: no binding agent-dispute or chargeback liability rule from Visa or Mastercard takes effect before the 2026 holiday peak; the first one likely arrives in the April 2027 release cycle at the earliest.
- Signal 1: Mastercard’s own Signals report, published on 3 August 2026, still frames liability and dispute resolution as an unsolved industry-coordination problem, and puts consumer willingness to let an agent buy autonomously at just 10%.
- Signal 2: HM Treasury’s payments consultation, published 14 July 2026, reopens authentication, consent and liability for unauthorised transactions, and does not close until 6 October 2026.
- Signal 3: the agent-attributed loss data networks need to price a liability rule does not yet exist at scale: AI-referred sessions remain well under 1% of e-commerce traffic and cluster at comparison rather than checkout.
- What to do: treat agent-initiated orders as an unpriced dispute category this quarter, capture agent metadata now, and negotiate the gap in your acquirer contract rather than waiting for a network to fill it.
Why this matters now
Most coverage of agentic commerce has focused on the front of the transaction: which protocol wins, which credential is presented, who controls the checkout surface. That framing has largely been settled over the past year, and shopappy has tracked it as agentic checkout became a named sales channel in merchant reporting. The unglamorous half of the transaction, what happens sixty days later when a cardholder says they did not authorise the purchase, has attracted far less attention.
That asymmetry is about to become expensive. A dispute rule is not a technical nicety; it is the document that decides whether a chargeback lands on the issuer, the acquirer, the merchant or the agent platform. Every other layer of the agentic stack has shipped something in the last eighteen months. The liability layer has shipped frameworks, protocols and white papers, but no binding rule.
The timing question is therefore not academic. Peak season 2026 will be the first holiday period in which agent-initiated purchase volume is material enough for merchants to notice in their own dashboards. If the rules governing those transactions are still the general card-not-present rules, the default allocation is well understood and unfavourable to merchants.
Three signals observed since mid-July point the same direction, and they come from three genuinely independent places: a card network’s own research arm, a national treasury, and the adoption data itself. None of them is a leak or a rumour. All three are publicly checkable, which is the point.
It is worth being precise about what a dispute rule does that a protocol does not. A protocol decides how a payment is presented and authorised at the moment of purchase. A dispute rule decides, weeks or months later, whose money moves back when the purchase is contested, and under what evidentiary standard. The two are routinely conflated in agentic commerce commentary, and the conflation is why the liability gap has stayed largely invisible.
Signal 1: Mastercard’s own August report still calls liability an open question
On 3 August 2026, Mastercard published a Signals report titled Encoding Trust: The Race for Intent, Consent and Control in the Agentic World. The framing matters as much as the content. A network that was weeks away from publishing a binding dispute rule would not typically release a research paper describing the problem as a race still being run.
The report identifies verifiable intent, agent identity and clear permission frameworks as the building blocks of trusted agentic commerce, and argues that AI platforms, banks, merchants and payment networks need shared approaches to identity, permissioning, liability and dispute resolution. The operative word is “need”. Liability and dispute resolution appear in the report as requirements to be met, not as rules already drafted.
The consumer data in the same report reinforces the timing read. Mastercard puts consumer openness to collaborating with an AI agent to find the best option at roughly 85%, and willingness to let agents complete specific commerce tasks on request at around 74%. But willingness to let an agent complete a purchase autonomously sits at only about 10%, with roughly a third open to an agent choosing what to buy provided the human retains final payment approval.
That 10% figure is the analytically important one. Fully autonomous purchase is precisely the scenario a novel dispute rule would need to govern, because it is the scenario in which no human pressed a button. If only one consumer in ten is currently willing to operate that way, the volume of genuinely agent-authorised, human-absent transactions is small, and the commercial urgency to write a bespoke rule for it is correspondingly low.
Mastercard’s report also describes agentic tokens as able to carry task-specific authority, restricted by agent, merchant, category, spending limit, timeframe or usage rules. That is real infrastructure, and it is the necessary precondition for a dispute rule. It is not the rule. The distinction between building the evidentiary rails and publishing the adjudication standard is the gap this piece is about, and it echoes the earlier build-out of tokenized agent identity as the gate for agentic checkout.
The report is also, notably, the second half of a pattern. Networks that intend to change liability generally telegraph it through mandates and effective dates in merchant bulletins. Research reports describing shared industry challenges tend to precede that stage by quarters, not weeks.
There is a second-order tell in how the report handles evidence. Encoding intent into a token is useful precisely because it anticipates a future dispute in which intent must be demonstrated. Building that capability now, while explicitly describing the adjudication framework as unfinished, is the behaviour of an organisation sequencing evidence before rules rather than shipping both together.
Signal 2: the UK reopened payment liability law in July, and the clock runs past peak
On 14 July 2026, HM Treasury published its consultation on modernising the UK’s payment services and electronic money regulatory framework, part of the National Payments Vision agenda. The consultation explicitly recognises that the Payment Services Regulations were written before agentic AI and may not fully accommodate it.
Critically for this prediction, the consultation asks whether the regulations should be updated in areas including authentication, consent and liability for unauthorised transactions, in order to support agentic payments safely and at scale. Respondents have until 6 October 2026 to reply.
Two things follow from that date. First, the substantive UK policy position on agent liability will not exist until after the consultation closes, which is already past the point at which peak-season rules would need to be locked. Second, and more consequentially, a live regulatory consultation on exactly the question a network rule would answer creates a strong incentive for the networks to wait.
Publishing a binding commercial liability rule while a national regulator is actively consulting on statutory liability risks two bad outcomes: the rule is superseded, or it is read as an attempt to pre-empt the regulator. The prior precedent here points clearly. Networks generally sequence their rule changes after regulatory direction is visible, as they did through the strong customer authentication build-out, where commercial standards followed the regulatory technical standards rather than leading them.
The UK is not the only jurisdiction in motion, which strengthens rather than weakens the read. The Competition and Markets Authority published research on agentic AI and consumer harm on 9 March 2026, taking the position that businesses are responsible for their agents’ actions in much the same way as for an employee’s actions. The EU AI Act’s transparency obligations began applying on 2 August 2026. In the United States, the Regulation E authorised-or-unauthorised binary remains an awkward fit for agent scenarios and has not been formally reconciled.
A network writing one global dispute rule has to satisfy all of these at once. That is a multi-quarter drafting problem even with full internal agreement, and it is a reasonable explanation for why the networks have shipped credentials first and adjudication later.
None of this suggests the networks are moving slowly by their own standards. Reconciling one commercial rulebook against a UK consultation, an EU statutory regime and an unresolved US position is genuinely difficult, and doing it badly would be worse than doing it late. The observation here is about sequence and timing, not about diligence.
Signal 3: the loss data needed to price a rule does not exist yet
Card networks do not write liability rules from first principles. They write them when they have enough attributed loss data to price the risk and allocate it without destabilising either side of the interchange relationship. That is the mechanism that makes this prediction more than a guess about corporate speed.
The adoption data from the last several months shows a channel growing very fast from a very small base. AI referral traffic to US retail sites grew sharply year over year through the first quarter of 2026, and Shopify has reported AI-referred orders to its merchants rising by an order of magnitude year over year. Those growth rates are genuinely striking.
The absolute levels are not. AI-driven sessions still sit at a fraction of a percent of total e-commerce traffic on most public measures. A channel at well under 1% of sessions, even growing at several hundred percent annually, generates a correspondingly thin tail of disputed transactions.
The composition of agent usage compounds the problem. Public surveys of agent usage put product comparison at roughly 62% of agent interactions, checkout at around 23%, and post-purchase activity at about 19%. The dispute-generating end of the funnel is the smallest slice of the smallest channel.
An independent research report published on 17 August 2026 covering agentic commerce adoption trends and execution limits reached a compatible conclusion: agent capability is concentrated in discovery, with completion and post-purchase handling lagging materially behind. The gap between what agents recommend and what agents actually transact is the gap in the loss data.
This is also why the merchant-side control decisions of the past year matter. Where retailers have kept the final transaction on their own surface, as tracked in the shift toward retailer-controlled agentic checkout, the resulting transaction looks like an ordinary card-not-present sale in the network’s data. It does not get flagged as agentic, which further starves the very dataset a rule would be priced from.
There is a reflexive quality to this that is easy to miss. The thinner the attributed loss data, the weaker the case for a bespoke rule; the longer the rule is deferred, the more agent transactions are settled under general card-not-present codes and never labelled as agentic at all. That feedback loop tends to delay the threshold rather than accelerate it.
The likely breaking point is therefore not a particular percentage of traffic. It is the first period in which agent-attributed disputes become visible enough in issuer loss reporting to prompt a question from a risk committee. On current trajectories, that condition looks more like a 2027 event than a 2026 one.
What the pattern suggests
Read together, the three signals describe a system that has built the plumbing, opened the policy question, and not yet accumulated the evidence. That combination has a fairly predictable output: continued publication of protocols, token specifications and voluntary frameworks, with binding liability reallocation deferred.
The release-cycle mechanics sharpen the timing. Visa announces most dispute rule changes on April and October release cycles, with the April 2026 core rules published on 18 April 2026. An October 2026 release would need its substantive content locked well in advance of publication, and would in practice take effect too close to peak for merchants and acquirers to implement.
| Signal | Date | Source type | What it implies | Typical lead time |
|---|---|---|---|---|
| Mastercard Encoding Trust Signals report | 3 Aug 2026 | Network research | Liability and dispute resolution still framed as shared problems to solve | Quarters, not weeks, before a mandate |
| HM Treasury payments consultation | Published 14 Jul 2026, closes 6 Oct 2026 | Regulatory consultation | Statutory liability position unsettled until after peak | 12–24 months to enforceable rules |
| Agent adoption and execution-limit data | Through 17 Aug 2026 | Market data | Insufficient attributed loss volume to price a rule | Depends on channel share crossing a threshold |
| Amex Agent Purchase Protection | 14 Apr 2026 | Issuer programme | Closed-loop protection is possible without a network rule | Immediate, but not portable |
The precedent set is consistent on lead times. Network liability shifts historically move on multi-year clocks from first announcement to effective date, because they require issuer systems, acquirer systems and merchant systems to change together.
| Prior precedent | Nature of change | Approximate announcement-to-effect lead | Relevance here |
|---|---|---|---|
| US EMV liability shift | Reallocated counterfeit fraud liability to the least-secure party | Several years | Closest analogue: a liability reallocation, not a credential launch |
| Strong customer authentication under PSD2 | Authentication mandate with liability consequences | Multiple years, with extensions | Commercial standards followed regulatory direction, not the reverse |
| Visa acquirer monitoring programme thresholds | Threshold tightening within an existing framework | Roughly two release cycles | Shows the fast path exists, but only for parameter changes |
| Card-on-file tokenisation | Infrastructure and credential handling | Phased over years | Infrastructure ships ahead of adjudication rules |
The pattern in that table is the core of the argument. Parameter changes inside an existing rule can move in two cycles. Genuinely novel liability allocations, which is what an agent-dispute rule would be, have historically taken considerably longer, and have generally followed rather than preceded regulatory clarity.
One further mechanical point supports the April 2027 read. Rules that change liability require acquirers to reconfigure dispute-handling systems and to communicate the change to merchant portfolios, which is work that acquirers strongly prefer not to schedule during peak trading. Even a network that had a finished rule in hand this month would face resistance to landing it in the fourth quarter.
Wider context: who actually holds the bag today
In the absence of a network rule, liability is not undefined. It is defined by whatever documents do exist, and those documents currently point in one direction.
The Agentic Commerce Protocol’s delegated payment specification is explicit that settlement, refunds, chargebacks and compliance remain with the merchant and their payment service provider, and that the agent platform is not the merchant of record. That is a clear allocation, and it is the merchant’s. The governance question around that protocol and its rivals is itself unsettled, as covered in the analysis of how ACP and UCP are likely to hand off governance.
American Express took a different route on 14 April 2026, launching its Agentic Commerce Experiences developer kit alongside Agent Purchase Protection, which it described as an industry first. The programme protects eligible card members from charges arising from AI agent error where authenticated purchase intent was transmitted. It is a meaningful move, and it is worth being precise about what it is not.
Amex operates a closed loop, issuing and acquiring within one network, which lets it absorb a loss unilaterally without renegotiating anyone else’s economics. Visa and Mastercard operate four-party models where any liability change has to be agreed across thousands of issuers and acquirers. A closed-loop protection programme is therefore weak evidence that an open-loop rule is imminent, though it is strong evidence that the problem is recognised.
Meanwhile the practical exposure is asymmetric. A consumer who used an agent can still dispute under existing consumer protection rules, while the merchant may have limited means to prove the purchase was genuinely authorised. The tokens now being issued do encode agent identity into the transaction record, which will eventually make attribution possible. Attribution without an adjudication rule simply tells you who to argue about.
It is also worth noting what the current arrangement means for smaller merchants specifically. Large retailers can negotiate bespoke dispute terms with their acquirers and can absorb an unpriced loss category for a quarter or two. Merchants without that leverage inherit the default allocation without the ability to alter it, which is the usual pattern when infrastructure ships ahead of rules.
Implications for retailers, platforms and payment providers
The practical consequence of this prediction is that agent-initiated orders should be treated as an unpriced risk category for at least the next two quarters. That is a manageable problem if it is named now and an unpleasant surprise if it is discovered in January.
For retailers, three moves look sensible regardless of how the prediction resolves. Capture and retain agent metadata on every order that arrives through an agentic surface, even where your systems currently discard it, because that record is the only evidence you will have in a representment. Segment agent-originated orders in your dispute reporting so you can see the loss rate separately rather than averaged into card-not-present. Review the dispute and liability language in your acquirer agreement now, because bilateral contract terms are the only lever available while the network rules are silent.
For platforms and marketplaces, the exposure is reputational as well as financial. If agent-initiated disputes cluster on a handful of sellers, existing seller-performance machinery will penalise those sellers for a failure mode they did not create and cannot control. Adjusting those thresholds ahead of peak is cheaper than unwinding the penalties afterwards.
For payment providers and acquirers, the window is a commercial opportunity. A PSP that offers a defined agent-dispute handling standard, even a contractual one, will be selling something the networks are not yet offering. That is a genuine differentiator through the 2026 peak and into the first half of 2027.
There is also a straightforward authentication read-across. The direction of travel on checkout authentication, tracked in the analysis of how passkey checkout follows regulatory bans rather than transaction volume, suggests that authentication changes tend to be driven by rule deadlines rather than by merchant demand. Agent dispute handling is likely to follow the same logic, which is another reason to expect movement on a regulatory clock rather than a commercial one.
| Scenario | What happens | Observable trigger | Assessed likelihood |
|---|---|---|---|
| Base case | October 2026 releases cover identity and token handling only; first binding dispute rule appears in the April 2027 cycle or later | October merchant bulletins contain no agent-dispute reason code | Most likely |
| Acceleration | A network issues a mid-cycle bulletin creating an interim agent-dispute condition before peak | An out-of-cycle acquirer notice in September or October 2026 | Possible but less likely |
| Regulatory forcing | UK or EU action after October 2026 compels a faster network response | Treasury response text signalling statutory change | Plausible in 2027, not in 2026 |
| Private resolution | Acquirers and large merchants settle terms bilaterally, and no public rule is needed for some time | PSP contracts adding agent-dispute clauses | Underrated |
A final implication concerns measurement discipline. Because agent-originated orders frequently arrive looking like ordinary card-not-present traffic, any internal estimate of agentic exposure built from payment data alone is likely to understate it. Cross-referencing referral and session data against order records gives a more honest picture, and it is the same dataset that would support a future representment.
Caveats: what could go wrong
This prediction has real ways to fail, and the strongest counter-signal is already on the record. American Express moved unilaterally in April 2026 with a protection programme that assigns a defined outcome to agent error. That demonstrates a network can act ahead of accumulated loss data when it judges the competitive positioning worthwhile.
If agent-initiated volume over Cyber Week 2026 materially exceeds current expectations, the competitive logic could change quickly. Visa and Mastercard both have an interest in not ceding the agentic narrative to a closed-loop competitor. A rapid, narrowly scoped interim rule is not implausible on that basis.
The release-cycle argument is also softer than it looks. Networks do publish out-of-cycle bulletins and mandates when circumstances warrant, and the April and October cadence governs the consolidated rules document rather than every operational notice. A targeted announcement in September or October would not be unprecedented, and would falsify the narrow version of this prediction while leaving the broader read intact.
A single high-profile failure could compress every timeline described here. If a widely covered incident involves an agent making large unauthorised purchases for many consumers at once, the pressure to publish something binding would be immediate and would not wait for a release cycle. Predictions built on institutional cadence are always vulnerable to a salient event.
There is also a definitional risk worth naming honestly. If a network publishes guidance that clarifies how existing reason codes apply to agent transactions, without creating a new liability allocation, reasonable observers could disagree about whether the prediction was right. The falsification test proposed here is deliberately narrow: a new or amended dispute condition, reason code or liability rule specific to agent-initiated transactions, with an effective date on or before 31 December 2026.
Finally, the private-resolution scenario would make this prediction technically correct and practically uninteresting. If large merchants and their acquirers simply write agent-dispute terms into contracts through the autumn, the absence of a network rule stops mattering much. That outcome is arguably underrated relative to the attention paid to the public rulebooks. Merchants can review the current consolidated rules directly in Visa’s publicly posted core rules document.
FAQ
What exactly is being predicted here?
That no binding, generally applicable dispute or chargeback liability rule specific to agent-initiated transactions from Visa or Mastercard will have an effective date on or before 31 December 2026. The first such rule is expected to appear in the April 2027 release cycle at the earliest, and possibly later.
How would someone check whether this was right?
Read the October 2026 and April 2027 network release bulletins and the consolidated rules documents. Look specifically for a new or amended dispute condition, reason code or liability allocation naming agent-initiated or agentic transactions. Guidance restating how existing codes apply would not count.
Does that mean merchants have no protection at all this holiday season?
Not quite, but close to it in practical terms. Existing card-not-present dispute rules still apply, and the tokens used by agentic protocols increasingly encode agent identity, which helps with evidence. What is missing is any rule saying who bears the loss when the agent, rather than the cardholder, made the error.
Is not Amex Agent Purchase Protection a counter-example?
It is the strongest counter-signal, and it is acknowledged as such. The important qualification is that Amex runs a closed loop and can absorb a loss without renegotiating with thousands of issuers and acquirers. Extending a comparable guarantee across a four-party network is a materially harder commercial exercise.
Why does the volume of agent transactions matter to a rulemaking timeline?
Because networks price liability rules from observed, attributed losses. With agent-driven sessions still well under 1% of e-commerce traffic and concentrated in product comparison rather than checkout, the disputed-transaction dataset remains thin. Thin data makes it difficult to set thresholds without either overcharging or underprotecting one side.
Could regulators act before the networks do?
Possibly, though not before the 2026 peak. The UK consultation does not close until 6 October 2026, and a government response followed by rule changes would realistically run well into 2027. The CMA’s March 2026 position that businesses are responsible for their agents’ actions is directionally useful but is research rather than binding rulemaking.
What is the single most useful thing a retailer can do this month?
Start capturing and retaining agent metadata on orders arriving through agentic surfaces, and segment those orders in dispute reporting. Both are low-cost changes, and both produce the evidence base needed either to defend a representment or to negotiate acquirer terms. Waiting until a rule exists means entering that period with no historical data.
Does keeping checkout on our own site remove the risk?
It reduces some risks and creates a measurement blind spot. A transaction completed on a retailer’s own surface generally looks like a standard card-not-present sale to the network, which is often commercially preferable. The trade-off is that the order may not be identifiable as agent-originated later, which is precisely when that label becomes useful.
What would most change this view?
An out-of-cycle acquirer bulletin from either major network in September or October 2026 introducing an interim agent-dispute condition. That would indicate the networks judge the reputational stakes of peak season higher than the benefit of waiting for regulatory clarity, and it would invert the sequencing assumption this analysis rests on.