Happy Monday, Fintech Takers!

I hope you had a relaxing weekend and are ready for a busy week.

I’ll be spending the remainder of this week in Austin, TX. My company (Workweek) is hosting its annual Upfronts meeting and we have some very exciting announcements, which I will be sure to share with you as soon as they’re live!

Plus, I understand that the weather in Austin in August is just about perfect. Not too hot, nice breeze. Ideal for a Montanan who spends most of the year surrounded by snow.

- Alex

P.S. — Speaking of Montana, if you are working on anything related to agentic AI and financial services and you are interested in spending a few days in the mountains of Montana talking to fintech operators, bankers, and policymakers about how we can get our industry ready for agentic AI, please consider applying to participate in the Agentic Readiness Summit.

Was this email forwarded to you?

Subscribe Here

Hands, stenciled at the Cave of the Hands.

3 FINTECH NEWS STORIES

#1: Agentic Commerce is Still in its Prehistory Phase

What happened?

Rain launched a coalition focused on agentic commerce:

Rain … today announced the launch of the Agentic Payments Alliance (APA), a coalition of organizations working together to help guide the development of agentic commerce. Founding members include Visa, Mastercard, Fiserv, Circle, Solana, and Remitly.

The Alliance itself is a working coalition, run collectively by its founding members rather than owned by any one company. Members will set its charter and mission together. Early work is expected to include shared research and frameworks, testing emerging standards for agent identity and authorization, and advocacy on the regulatory questions agentic commerce raises.

And CNBC reported that Synchrony and OpenAI have entered into a partnership:

By using OpenAI’s models, Synchrony hopes it can remain relevant in a future where AI agents help to research and purchase items. The partnership, which is in its early stages, is a step toward enabling Synchrony customers to have smoother online shopping experiences.

Separately, Synchrony said it is launching a ChatGPT plugin that lets consumers browse its marketplace deals, promotional financing and partner offers, and that it is deploying OpenAI’s latest models internally to speed up product development.

So what?

Interestingly, the CNBC report was substantially revised, after publication, because the information provided by a Synchrony executive reportedly mischaracterized the collaboration between Synchrony and OpenAI.

The updated version only talks about the two companies entering into a strategic partnership, with Synchrony using OpenAI’s models internally (they were doing this already, I’d imagine?) and launching a ChatGPT plugin to allow users to shop for deals and review offers from Synchrony’s marketplace.

Luckily for us, the original version of the story lives on in contemporaneous coverage of the CNBC story (here and here). For posterity’s sake, I’m going to republish the version reported by those other outlets here as well:

Synchrony Financial, the credit card issuer for brands including Amazon, Walmart and Lowe's, is working with OpenAI to allow shoppers to buy products directly inside ChatGPT using their store cards.

The deal is one of the first major moves by a U.S. consumer lender to bring financing, payments and rewards directly into an AI chatbot.

While agentic commerce has become a catchphrase for the next phase of online shopping, consumers who discover items in an AI agent are typically still routed to a brand's website to complete the purchase. To change that, OpenAI has signed deals with companies including Visa and Stripe to move toward in-chat purchases.

"What happens today is the transaction doesn't cleanly happen yet at the provider like OpenAI," said Maran Nalluswami, Synchrony's chief strategy officer, in an interview. "We want to ensure that if a transaction's going to happen in that ecosystem, our cards are loaded up in the right spots to ensure that that transaction finishes."

Still, there's work ahead before seamless agentic commerce becomes reality.

Nalluswami said doing the work to get general-purpose cards within ChatGPT will probably take six to 12 months, and possibly longer for private label store cards that only work at specific retailers, which takes additional coordination with the brands.

Consumers remain cautious about handing credit card information to AI or allowing an agent to complete a purchase. There are also questions about who pays for transactions completed inside ChatGPT. Nalluswami said the economics will need to be negotiated among retailers, Synchrony and OpenAI.

Synchrony is also talking with competing AI platforms, including Anthropic's Claude and Google's Gemini, on embedding their cards within those chatbots, Nalluswami said.

Wow! That’s very different from the updated version that CNBC has out there now. I’m honestly surprised that CNBC agreed to those changes, given that it had Synchrony’s Chief Strategy Officer on the record (he is not quoted at all in the updated story).

The updated story does not even use the term “agentic commerce,” let alone report that Synchrony is planning to enable customers to use its cards within ChatGPT. It doesn’t mention the need for cards to be “loaded up in the right spots to ensure the transaction finishes.” Nor does it give a timeline for how long that will take for general-purpose cards (6-12 months) and private label cards (longer due to additional coordination with the merchants).

I’m not sure if it was OpenAI that objected to the original CNBC story (it mentioned Synchrony working with competing AI platforms like Anthropic and Google on agentic commerce, and that detail got scrubbed). Or if it was Synchrony’s own comms team, feeling the need to clear up the overly candid and forward-looking comments made by their Chief Strategy Officer. Perhaps a combination of both.

Regardless, the most interesting part to me is the question in the original version about who pays for card transactions that happen inside ChatGPT.

Obviously, for general-purpose cards, the merchant pays the merchant discount fee (2-3%), which is then spread between the issuing bank, the acquiring bank, and the network. For private label — where the card can only be used with the merchant that offers it — the merchant doesn’t pay a merchant discount fee at all. Instead, the costs to run the card program are subtracted from the revenue it generates and the remaining profit is split between the merchant and the card issuer, usually with the merchant taking the lion’s share. In agentic commerce, the AI labs want to be paid for driving additional commerce to merchants (much in the same way that companies like Google and Meta get paid today for generating leads).This probably will happen, to some degree, for general-purpose cards, though merchants will grumble about it and do what they can to not get too disintermediated from the end customers. But for private label? The economics will have to look very different. In fact, to the extent that AI agents drive any spend on private label cards, the AI labs will probably need to get their cut from the issuer (Synchrony, in this case) rather than the merchant. Hence why the original article said private label integrations with ChatGPT will take longer than 6-12 months.

And that brings us to Rain’s Agentic Payments Alliance. To me, the most notable thing about the alliance is that its members hardly overlap at all with the companies that have been hard at work developing the various protocols — Agentic Commerce Protocol, Agent Payments Protocol, Universal Commerce Protocol, x402, Machine Payments Protocol — that agentic commerce transactions will, supposedly, run on someday.

No Stripe. No Google. No OpenAI. No PayPal, Coinbase, Meta, Amazon, Walmart, or Target.

Maybe that doesn’t matter. Maybe those companies are working on the protocols — the plumbing for how consumers talk to AI agents and AI agents talk to merchants and card issuers — and Rain and its fellow alliance members (which seem heavily concentrated in the issuer-processor and fraud/compliance markets) are working on the larger industry coordination and policy challenges. Perhaps everything will end up snapping together nicely. Rain’s alliance is certainly stocked with enough payments nerds (including Rain’s new head of payments, Sophia Goldberg!) to make that a reasonable bet.

That said, it feels, between these two stories, like we’re still in the very very early prehistory phase of agentic commerce.

It’s all cave art and stone tools at the moment.

#2: The States are Positioning Themselves for a Fair Lending Fight

What happened?

Illinois is stepping into the gap created by the Trump Administration:

Illinois has joined a growing number of states that are expanding fair lending obligations at the state level even as the federal government moves in the opposite direction. On July 31, 2026, SB 3777 became Public Act 104-0744, amending the Illinois Human Rights Act (Act) to prohibit not only intentional discrimination in lending and credit card issuance, but also the use of facially neutral underwriting criteria or methodologies that produce discriminatory effects. The Act applies to financial institutions, credit card issuers, employers and providers of public accommodations. The amendments to the Act will become effective on January 1, 2027.

Notably, the Illinois Human Rights Act does not define the term “financial institution.” As a result, questions may arise regarding the statute’s application to nonbank lenders, fintech companies, marketplace lenders, and other entities that extend consumer credit but are not traditional depository institutions. By contrast, the Act separately applies to any “person who offers credit cards to the public,” suggesting that the General Assembly intended the credit card provisions to have broad applicability.

For creditors operating in Illinois, this legislation marks a meaningful shift on the state level. It codifies a disparate impact framework for credit decisions—one that imposes a demanding burden on lenders to justify their practices and demonstrate the absence of less discriminatory alternatives.

So what?

I summarized the Trump Administration’s unusually long attention span when it comes to dismantling the theory of disparate impact in this newsletter, so I won’t rehash the full thing. Suffice it to say, there is virtually no federal agency that isn’t actively working to ameliorate the harmful effects of this radical woke legal theory that [double checks my notes] requires lenders to examine the effects of their decisions on protected classes, ensure that those effects are caused by practices that serve legitimate business interests, and search for less discriminatory alternatives (LDAs) that serve those same business interests.

(I overstate for humorous effect, obviously. Disparate impact is far from perfect, as I wrote about in this essay. However, I think the solution is to fix how the doctrine is applied, rather than throwing it away entirely.)

The states, in turn, are stepping up their individual efforts to keep disparate impact — and fair lending, more broadly — as supervision and enforcement priorities for lenders, and, in some cases, going further than the federal government had gone before. California, Colorado, Connecticut, Massachusetts, and Minnesota have all had robust disparate impact theories enshrined in law and regulation, and New Jersey, New York, and Maryland have each taken recent action in this area as well.

This sets up a really interesting clash between the states and the federal government.

To give one example, the FTC recently proposed a policy statement — Suppression of Accuracy in Artificial Intelligence Systems — that advances a novel theory: Steering a model away from accuracy (even to comply with a state law) can be a deceptive act if you don't disclose it. This policy statement (if it is finalized and made an enforcement priority) could clash, in spectacular fashion, with state-level fair lending requirements. The standard practice in fair lending — the thing Illinois now effectively requires — is a less-discriminatory-alternative search. You take your most accurate model, hunt for one with lower disparate impact that still serves the business purpose, and adopt it. An LDA is, by construction, a model tuned away from maximum accuracy and toward equity. That's the entire point.

So the same model adjustment is mandatory in Illinois and, under the FTC's theory, presumptively deceptive "accuracy suppression." Adopt the LDA (without publicly disclosing it) and you satisfy Illinois while handing the FTC a reason to investigate you. Ship the max-accuracy model and you satisfy the FTC while violating Illinois human rights law because a less discriminatory alternative existed and you didn't use it.

This exact fight may not end up happening. The FTC’s suppression of accuracy statement isn’t final, nor is it a formal rule or statute. However, it’s exactly the fight that Russ Vought and state legislators like Adriane Johnson (sponsor of the Illinois disparate impact bill) want to have. And it leaves lenders stuck, uncomfortably, in the middle.   

#3: Data Acquisitions

What happened?

Google is buying the data-rich husk of Spirit Airlines:

Strip an employee's name from a decade of HR threads and payroll records, and apparently that data is worth real money. Google just paid $10 million for exactly that kind of corporate ghost — the entire digital footprint of bankrupt Spirit Airlines, including roughly 100 million internal emails and 500 million Microsoft Teams items, acquired through a competitive bankruptcy auction and destined to train Google's AI models.

So what?

2026 is very weird.

You have a company (Google) that has operated an incredibly efficient, asset-light business model that threw off billions of dollars in free cash flow every year for more than two decades suddenly going free-cash-flow negative (-$5.85B) in Q2 of this year and raising more than $80 billion in new equity capital to support the same priority that sucked away all of its free cash flow in the first place.

That same company is also paying $10 million (not a huge sum, comparatively, but also not nothing when you’re operating with constrained cash flow) for the emails, team chats, customer service call recordings, and internal production code of a bankrupt business from an industry that Google has never operated in.

Weird!

The reason that this is happening, of course, is AI. The AI buildout has consumed all of Google’s free cash flow. And it has also transformed Google (and every other AI lab … and, eventually, every other company) into a voracious consumer of proprietary first-party datasets.

This, of course, makes me wonder what datasets in financial services might be attractive to Google and Meta and X and OpenAI and Anthropic, if they were to become available.

This is a tricky proposition in any industry (Google had to promise to de-identify the data that it’s getting from Spirit), but it’s especially tricky in financial services. In addition to the standard bankruptcy requirement to respect a company’s privacy policy in death (which sometimes generates controversy!), financial services providers are further bound by Gramm-Leach-Bliley and the Fair Credit Reporting Act, among other statutes and regulations.

Put simply, it’s very difficult to acquire useful datasets, post-mortem, when the dead body is a fintech company or a bank.

But it’s not impossible. And, more importantly, the definition of useful datasets, in the age of LLMs, is very different than it was a decade ago. Back then, you needed the PII because the PII is what had value, either for direct customer acquisition or for resale to marketers or data brokers. Today, LLMs are hungry for any data that can provide domain-specific texture (industry jargon, organizational communication patterns, customer service trends, etc.) to make model predictions more accurate.

So, in financial services, what we’re looking for is context-heavy data exhaust that clears this higher compliance bar. The best example of this — and one that I think will become a bigger target for AI and AI-adjacent companies over the next five years — is what I think of as the “reasoning layer,” the places where humans working for the bank or fintech company applied their own judgement and expertise to a genuinely uncertain situation.

This data has had no value to acquirers, historically. But today? It’s super valuable training data. Here are a few examples:

• Collections and loss-mitigation call transcripts, paired with the agent's notes. The asset is how a skilled collector or mortgage servicer reads a debtor's situation and navigates to a resolution: When to offer a hardship plan, how to de-escalate, how to work a loss-mitigation waterfall. All of it is bound by Reg F constraints the model has to learn. Strip out the name and account number and the negotiation logic survives intact.

• Dispute and chargeback adjudication files. A human investigator reviews evidence, applies the rule, and writes a resolution rationale (here's the evidence, here's why I found for or against the consumer, etc.). Evidence-to-rule-to-decision is the exact reasoning chain an AI model wants, and it detaches from identity cleanly.

• Fraud alert-triage decisions and case notes. The analyst's written rationale for why an alert got escalated or cleared. The investigative reasoning is the value, and it generalizes; the identifiers are incidental. However, once a fraud alert turns into a SAR, the data becomes confidential, so the data asset would need to be upstream of the formal regulatory reporting.

• Underwriting overrides and exception memos. The "second look" where a human underwriter overturns or upholds an automated decision and writes why. This would have tons of value to the AI labs and AI-native banking vendors, but the de-identification process is much harder because the underwriter’s reasoning is fused to the applicant's specific PII.

Sponsored by Spinwheel

Balances tell you what someone owes. They don't tell you what that debt actually costs them.

Spinwheel now pulls a consumer's current credit card APR alongside their balances, payoffs, and liabilities.

That means a more complete picture of consumer debt to sharpen savings calculations, payoff recommendations, refinancing offers, and financial guidance.

See how it works

2 READING RECOMMENDATIONS

#1:  WSJ Editorial Misreads the UWM Situation (by Doug Simons, The Accidental Financial System) 📚

I’ve really been digging Doug’s newsletter (you should subscribe if you haven’t already!) and this one, with the subtitle — An Obsession with Moral Hazard Distracts from the Need for Better Regulation — is especially good.

#2: Between Craft and Quality (by Keyana Sapp, Worse on Purpose) 📚

This article is very good, but the real reason that I’m sharing it with you is because the website it was published on (Worse on Purpose) is, far and away, the best new website I have discovered in years.

You’re welcome.

*Bonus: Know Your Agent: A New Problem for Old Defenses (by me, with Persona) 💻

Bots used to be our enemy, but not anymore. AI agents are already moving money for consumers, and someone's going to own the trust layer they all have to pass through. I sat down with experts from Persona, Lithic, and Glenbrook Partners to work out how and who. Watch here on demand.

*This rec is brought to you by one of our fantastic brand partners.

1 QUESTION FROM FINITY

There are a TON of interesting questions being asked in Finity (our digital community for fintech and banking nerds). I’ll share one question, sourced from the community, each week. However, if you’d like to join the conversation, please apply to join!

Please name one product that you used to love that you no longer love because it has become worse. And name one product that you still love because the craftsmanship is fucking amazing.

If you have any thoughts on this question, reply to this email or DM me in Finity!

Thanks for the read! Let me know what you thought by replying back to this email.

— Alex