Perspective · August 2026 · 5 min read
The agent buys. The acquirer pays.
Three builders are each constructing one quarter of the same map — mandate, evidence, recourse, reconciliation. None of them can finish it alone, and until someone does, loss allocation rolls downhill the way it always has.
I wrote recently that the Agentic Payments Alliance is not an innovation club. It is a liability negotiation.
This week the negotiation showed up in my comments and my inbox.
Three builders wrote to me. Each is building one quarter of the same map: mandate, evidence, recourse, reconciliation.
None of them can finish the job alone. That is not a criticism. It is the finding.
The mandate builder
A policy-engine founder is building governance for agent payments. Rules before the transaction. Escalation to a human when the transaction lands in a grey zone. A cryptographically signed decision after the policy check, bound to agent, vendor, amount, nonce, and expiry.
Serious work. And he named his own gap without flinching: enforcement depends on the processor actually checking for the signature. A processor that has never heard of him will process the payment anyway.
He is issuing mandates into a world with no one obligated to read them.
The evidence engineer
A backend engineer who has built payment systems for a decade wrote the line I have not stopped thinking about. Dispute evidence was never a query problem. It was a recording problem. By the time anyone asks, the row has been updated in place and the reason is gone.
Then he went one step further. With agents in the flow, the field nobody captures is the mandate itself: what the user authorized, at what scope, at what moment.
A reconciliation gap can be repaired next quarter. A mandate that was not recorded at authorization does not exist. Ever.
The recourse founder
A third builder is working on the quarter nobody on the Alliance founding list is building: recourse. Collateral pledged before the transaction, paid automatically to the harmed party on proof of breach. No arbitrator. No sixty-day cycle.
Elegant. And it inherits everything upstream. Automatic recourse settles only as fast as the evidence behind it, and evidence settles only as well as the mandate beneath it.
Fast recourse against weak records is not justice. It is a faster way to be wrong.
Meanwhile, the standards tables multiply
Google's AP2 defines signed mandates as verifiable credentials, with Mastercard and Amex at the table. Visa runs its own trusted-agent track. OpenAI and Stripe have a third. The new Alliance seats several of these players at a fourth.
Count what exists: at least three mandate vocabularies.
Count what does not: a single rulebook that says who eats the loss, and a single party obligated to reject a transaction that arrives without its papers.
Formats are free. Enforcement is expensive.
The part I want acquirers to sit with
You already know how loss allocation ends, because you live at the bottom of the hill. The cardholder disputes. The issuer pulls the money. The merchant fights or folds. And when the merchant folds, the acquirer pays. Underwriting exists because recourse rolls uphill until it finds someone solvent.
Agent transactions do not change that gravity. They remove the rulebook that softened it.
When an agent buys the wrong thing, was the cardholder present? Does the chargeback right attach? Whose fault is a mandate the merchant never saw and the issuer never checked?
The honest answer today: nobody knows. And when nobody knows, history says the acquirer pays while the lawyers find out.
So do not wait for the standards. Four things are worth doing this quarter, and every one of them pays for itself even if agents never reach your portfolio.
Record something at authorization for every agent-originated transaction, even before a standard tells you what. A signed record you invented beats a perfect record that does not exist.
Make dispute evidence write-once. If your systems update the row in place, you are deleting your own defense in real time.
Count the person-dependent steps between a bad transaction and a complete reconstruction of it. Every person in that path is latency you will pay for at dispute time.
Decide now which agent-originated volume you will decline without a mandate attached. The first acquirer with that policy will write the template everyone else copies.
And the part for software companies
If you embed payments in your platform, you are holding the one enforcement point that actually works, and most of you do not know it.
The processor will not check a third party mandate. The network standard will take years. But your platform decides whether the agent ever holds a usable credential.
Do not give the agent an instrument until the mandate exists. A single-use credential, bound to vendor, amount, and expiry, minted only after the policy check. Then bypass is not a violation you detect. It is a payment that cannot exist.
The mandate protocols are the paperwork. The credential is the lock.
Platforms that understand this will not wait for the standards bodies. They will be the enforcement layer the standards bodies eventually describe.
This is the work I do: settlement, evidence, underwriting, and the operating model underneath agent readiness, for acquirers who will otherwise inherit the loss, and for software platforms that want to be the lock rather than the loophole.
The standards will decide who is liable. Your architecture will decide who actually pays.
Guillermo Lizasoain is the founder of Gleaning Capital LLC. He spent twenty-five years building and modernizing payments platforms — global acquiring, processor integrations, settlement and reconciliation, and underwriting automation — and now advises investors and operators on the same systems.