An agent now decides which chargebacks are worth fighting, and Apple wants to pay publishers by the use

Maverick Payments has put an AI agent inside its dispute workflow, and the agent is the thing that decides which chargebacks get contested [1].

I have sat on the losing end of that decision more than once. Contesting a chargeback costs more in staff hours than a lot of disputed transactions are worth, so the working answer in most merchant back offices is that nobody fights the small ones at all. Nobody wrote that down as a policy. It is arithmetic doing the job of one, and it has been setting dispute strategy for years.

The agent picks its battles

PYMNTS reported on 13 August 2026 that Maverick Payments has integrated Findustry AI's Chargeback Agent into its white-label unified payments platform, so partners can offer it to merchants through the platform where they already process payments, without navigating network rules or writing new code [1]. The agent gathers evidence from merchant systems, reads what the card network requires for the reason code in question, assembles the rebuttal and files it [1].

Jonathan Razi, Findustry's founder and chief executive, put it this way: "Like a good coworker, our AI agent will decide which cases to fight, which to automate, and which to surface for you" [1]. Guy Dimaggio, senior vice president of operations and strategy at Maverick Payments, framed the reason in cost terms: for the high-risk verticals the company serves, chargebacks are a significant operational and financial burden [1].

The load-bearing word in Razi's sentence is decide. You buy a rebuttal generator the way you buy any other tool. Something that ranks disputes and declines the ones it judges unwinnable is doing triage, and triage needs an owner a regulator can name. When a card network asks why a merchant never contested a run of fraud claims, "the model scored them low" is only an answer if somebody can produce the scores, the threshold and the person who approved it. Dispute systems are built to record what they file. The decision not to file leaves no artefact unless somebody built one on purpose.

Apple would rather pay by the use than by the licence

The Wall Street Journal reported, and MacRumors and TechCrunch carried on 12 and 13 August 2026, that Apple has approached publishers about supplying current news to the rebuilt Siri, and has proposed multiyear agreements that pay a publisher when its content is actually used rather than a flat fee for broad access [2][3][4]. Apple has discussed a nine-figure budget for those payments [3]. The new Siri is due this autumn in iOS 27, iPadOS 27 and macOS 27 [3].

A flat licence is easy to sign and easy to forget about. A metered one needs somebody to define what counts as a use, and then to be trusted on the count by a party with no view of the meter. If a model reads four articles and cites one, how many uses is that. If a summary paraphrases a paragraph without naming the outlet, does the counter fire. Every publisher in those talks is being asked to accept a measurement system built by the buyer, which is roughly the position card acquirers have held over merchants for about forty years. Publishers have not had to negotiate that shape of deal before. Acquirers could tell them what to ask for.

Anthropic would rather buy the efficiency than the hardware

Bloomberg reported on 13 August 2026, carried by Fortune, that Anthropic is in talks to acquire the startup Decart for about $6 billion, which would be its largest known acquisition, and that the talks are not final and could still fall through [5]. Decart builds software that makes AI chips run more efficiently and was valued at close to $4 billion after a funding round in May 2026, so the reported price carries a premium of roughly half again on a three-month-old mark [5]. Fortune notes Anthropic rarely makes large acquisitions and has been spending heavily on computing power ahead of an anticipated flotation [5].

Strip out the size and it is a decision any operator running on someone else's compute has made in miniature. You can buy more capacity, or you can buy a lower cost per unit of the capacity you already have. Only the second one keeps paying after the invoice clears. What is unusual here is the amount somebody is willing to spend once to move a recurring number, and the recurring number is the price of a token.

Read from the rails

All three of these are arguments about what a unit of machine work costs and who gets to count it. Findustry is selling a judgement about which units are worth spending, Apple is proposing to buy them one at a time rather than by the year, and Anthropic is reportedly paying $6 billion to bring the unit price down. The capability question is settled enough that nobody in any of the three is asking it.

Two things worth doing this week if any of this touches you. If you run disputes, find out what your current process declines and whether that decision is recorded anywhere a network could read. If you license content or data to a model, price the metered version before somebody offers it to you, because the first party to define the counting unit usually keeps it.

And on the Findustry integration, I would want one number before signing: the recovery rate on the cases the agent chooses not to fight. Nobody publishes that one.