AI Won't Embed Banking. Banking Will Embed AI.

Why the institution that holds your data, your license, and your right to move money will own the AI layer, and what the Plaid-OpenAI deal really revealed.

On May 15, 2026, OpenAI flipped a switch that most of the financial-services press read as the beginning of the end for bank-built personal finance. ChatGPT Pro subscribers in the United States can now connect a bank account through Plaid and ask the assistant about their balances, transactions, investments, and liabilities. The integration reaches more than 12,000 institutions, including Chase and Schwab. The headlines wrote themselves: ChatGPT is now your bank app.

It is not, and the announcement actually proves the opposite. What ChatGPT received that morning is a flattened, read-only slice of consumer finance, the same slice every aggregator-fed PFM tool has been working with for fifteen years. The data that matters most for what AI is about to do (originate credit, move money, plan a household balance sheet, defend a customer against fraud) never left the bank. It is structurally incapable of leaving the bank under the current rules, and the regulatory floor that was supposed to force banks to be data utilities for external AI has just been kicked into the next administration.

AI will not embed banking. Banking will embed AI.

The institution that holds the deposit relationship (whether that institution is JPMorgan or Chime) owns a data position, a regulatory license, and an action surface that no external AI can replicate from the outside.

The structural ceiling on any external-AI approach

The Plaid-OpenAI integration delivers four data categories into ChatGPT: balances, transactions, investments, and liabilities. Read-only. No full account numbers, no ability to move money, place trades, or pay bills. That is the launch surface, and it is the product working exactly as designed.

The deeper point (the one the headlines missed) is not about any single integration. It is about the architecture of the data-sharing layer itself. Open banking in the United States was built by banks, governed by a coalition of industry groups, but strongly influenced by a bank-led standard body (FDX), and shaped by a regulatory framework that explicitly carves out the data banks are entitled to keep. Every external party that wants to reach a consumer's bank data, whether that party is an AI assistant, a budgeting app, a lender, or a brokerage, connects through that same shared layer. The connector itself is upstream of the question. The constraint is what banks have agreed to send through any pipe, and what regulation says they do not have to send through any pipe.

Inside that "data sharing" layer, the leading aggregators do an impressive amount of work. They translate disparate bank APIs into clean schemas, enrich raw transaction strings into structured merchant data, detect recurring patterns, surface identity and income, expose holdings down to the security identifier, and ship statement PDFs. The breadth is real. But the structured fields that any external party receives reflect what banks have chosen to expose, not the full set of fields a bank holds internally about the customer.

Consider the gap as a fact about what the data-sharing layer carries versus what lives in the bank's core. The Social Security number on file at the bank does not cross the wire to any aggregator under any current product. A Zelle debit shows up as a transaction, but the recipient's phone number or email (the structured counterparty identity) is not a field U.S. banks expose. Wire transfers arrive as a description string; the structured beneficiary name, beneficiary bank, SWIFT code, and originator-to-beneficiary memo are not in the schema. Bill-pay payee directories, with stored account numbers and routing instructions, do not move through the pipe. Scheduled and future-dated transfers do not appear. Joint account permissions, authorized users, beneficiary designations, and TOD/POD instructions stay internal. The CFPB's Section 1033 final rule explicitly exempted four categories of data from any mandatory sharing regime: confidential commercial information including risk-scoring algorithms, anti-fraud and AML data, information protected by other law (which captures most KYC artifacts), and information not retrievable in the ordinary course. None of the major bank-aggregator API contracts in the market today carry materially more than that floor.

The bank's own operational telemetry sits another layer behind that. The device fingerprint, the login geolocation, the behavioral biometrics, the dispute history, the card controls, the internal fraud score, the BSA/AML flag, the customer service transcripts, the loan officer's CRM notes: none of this is in the data-sharing layer because it was never meant to be, and the regulatory framework explicitly says it does not have to be.

The bank knows who the customer actually pays each month and how to reach them. The bank knows about the wire set up last quarter to the contractor in Florida. The bank knows whether the device asking for money right now is the one the customer has used for the past four years. An external AI plugged into the data-sharing layer, through any connector, today or in any plausible near-term scenario, sees "ZELLE TO J SMITH, $32.00, transfers."

The regulatory floor disappeared

If you were betting against the bank moat in 2024, you had a serious card to play: CFPB Section 1033. The final open banking rule, published in the Federal Register on November 18, 2024 under then-Director Rohit Chopra, would have forced banks to share a defined set of consumer data with authorized third parties through standardized APIs, with the largest data providers compliant by April 2026.

That card is gone. On October 29, 2025, Judge Danny Reeves of the Eastern District of Kentucky granted a preliminary injunction finding the rule likely exceeded the Bureau's statutory authority, including, critically, the prohibition on banks charging interface-access fees to aggregators. The Bureau under new leadership had already signaled in May 2025 that it intended to vacate the rule, and in August 2025 published an Advance Notice of Proposed Rulemaking to reconsider. As of late May 2026, the original rule is enjoined and unenforceable. A revised rule, if it ever appears, will almost certainly be narrower in scope and explicit in permitting access fees.

The court's ruling did not just delay open banking. It validated the monetization play. JPMorgan announced in September 2025 that Plaid would pay Chase for customer data access, the first time the money has flowed in that direction, after Chase documented 1.89 billion aggregator API calls in June 2025 alone, of which only thirteen percent traced back to active customer behavior. Yodlee, Morningstar, and Akoya signed similar deals by November 2025. Bank of America, Wells Fargo, and Citi are reportedly working through the same playbook. The economic structure of aggregation is being inverted. Banks have always been the data source. They are becoming the data toll booth.

This matters for the embedded-AI question because the strongest version of the "external AI wins" thesis assumed the regulator would force banks to expose data they would rather keep. That assumption has now been rebutted in federal court.

The biggest banks are moving at unprecedented speed

The lazy counter to a "bank wins AI" thesis is that banks are slow, bureaucratic, and culturally incapable of shipping AI products at the pace of an OpenAI or an Anthropic. That counter was correct in 2019. It is no longer correct.

Bank of America's virtual assistant Erica crossed thirty billion client interactions by March 2026, having taken a decade to reach its first three billion and then compressed the next twenty-seven billion into roughly eighteen months. Wells Fargo's Fargo assistant crossed 245 million interactions in 2024, running underneath on Google's Gemini Flash 2.0, with personally identifiable information deliberately stripped before any prompt reaches the language model. JPMorgan's LLM Suite, deployed across more than 200,000 employees, was named American Banker's 2025 Innovation of the Year and is explicitly model-agnostic, routing across OpenAI and Anthropic models underneath a JPMorgan-controlled application layer. Citi unveiled Citi Sky at Google Cloud Next on April 22, 2026, an AI agent built on Gemini Enterprise and DeepMind's real-time avatar technology rolling to Citigold wealth clients beginning this summer. Morgan Stanley's bespoke OpenAI partnership has reached ninety-eight percent adoption across its financial advisor force, with three live products including AskResearchGPT and the AI @ Morgan Stanley Debrief tool.

This is not slow. This is unprecedented for institutions of this size. And it is happening because attrition risk has finally become more painful than the cost and complexity of moving.

The resource base behind these moves is not appreciated outside the industry. JPMorgan Chase's 2025 technology budget was approximately eighteen billion dollars, with roughly $1.3 billion earmarked specifically for advancing AI capabilities; the 2026 budget is on the order of $19.8 billion. OpenAI's entire projected 2025 revenue was approximately thirteen billion dollars. One bank planned to spend more on technology last year than the leading external AI player generated in revenue. The aggregate technology spend of the top five U.S. banks comfortably exceeds the combined revenue of OpenAI and Anthropic.

OpenAI is growing extraordinarily fast, exiting 2025 at roughly twenty billion in annualized revenue and crossing twenty-five billion by March 2026, and the gap will close. But the banks have a structural advantage that cumulative revenue does not erase: a hundred and fifty years of accumulated infrastructure, regulatory licensing, and customer trust, deployed at the scale of national mass-market relationships.

The honest counter: the bank rents the model

The strongest objection to all of this is the one that gets published on LinkedIn every day: every flagship bank-built AI product is, at the model layer, a frontier-lab product. JPMorgan's LLM Suite runs GPT and Claude underneath. Citi Sky is Gemini. Wells Fargo's Fargo is Gemini Flash. Morgan Stanley's assistant is OpenAI. If the model is the source of value, the value accrues to the labs, not to the banks. The bank rents the brain. The lab owns it.

The response is that the model layer and the customer layer are two different businesses, and the AI labs will win one of them spectacularly without winning the other.

OpenAI, Anthropic, and Google are on track to build some of the most valuable infrastructure businesses ever assembled. Almost every bank-built AI in the United States runs on their models. Every neobank that ships AI features ships them on rented frontier model capacity. The model layer is going to be a multi-hundred-billion-dollar revenue tier, possibly the most valuable software business of the next twenty years, and the AI labs are positioned to own most of it. Nothing in this argument depends on disputing that. The labs win, and they win big.

The bank rents the brain. It owns the customer, the rails, the records, and the right to act.

The argument is about which layer owns the customer. The analogy that fits is cloud infrastructure. AWS, Azure, and Google Cloud became three of the largest businesses in the world by powering the applications that consumers actually use. Consumers do not have a relationship with AWS. Consumers have a relationship with Netflix, Airbnb, and Robinhood, all of which run on AWS. The model layer is shaping up the same way. OpenAI will power the assistants that Chase, BofA, Citi, Wells, Chime, and Robinhood put in front of their customers, and the customers will have a relationship with their bank, not with OpenAI. The labs become foundational infrastructure. The banks become the brand.

This pattern holds because the customer-facing layer in financial services is gated by things the model cannot supply on its own: a license to extend credit under ECOA, a license to hold deposits under the FDIC, a license to give fiduciary advice, the right to file a SAR, the right to debit under Reg E, the right to move money on a customer's behalf. ChatGPT, today, can show a customer their balance. It cannot originate a HELOC. It cannot execute a fiduciary trade. It cannot move money out of a wire account. It cannot underwrite a credit decision. It cannot collect a debit. The instant a customer wants to act on AI advice, the customer re-enters a bank's environment. The bank rents the brain; it owns the customer, the rails, the records, and the right to act. The labs cannot put that ownership into their roadmap because it is created by statute and by the consumer's pre-existing relationship, neither of which is available for purchase.

What about Mint, Credit Karma, and Rocket Money?

The historical analogy that gets thrown at any "banks win" thesis is the previous generation of PFM. Mint scaled to millions of users on aggregated bank data before Intuit acquired it in 2009. Credit Karma scaled past a hundred million members before Intuit paid $7.1 billion for it in 2020. Rocket Money was named the top personal finance app of 2026. None of these companies were banks. All of them were built on the same aggregator-fed data position that Plaid now supplies to ChatGPT. Surely, the objection goes, the same thing is about to happen to bank AI.

It is not, for two reasons.

The first is that those companies never won the action layer. They won the dashboard. They aggregated transactions, categorized spending, and surfaced recommendations. The actual movement of money, the actual extension of credit, the actual execution of trades, the actual closing of mortgages: none of it ever migrated to the aggregator-PFM layer. Banks remained the system of record. The PFM tools remained a window. AI changes this calculus precisely because AI's value compounds at the action layer, not the dashboard layer. Telling a customer "you spent $312 on subscriptions last month" is interesting. Canceling those subscriptions, refinancing the customer's mortgage, rebalancing the brokerage account, and rerouting direct deposits: that is what AI is about to do, and that is what cannot be done from outside an institution.

The second reason, and this is the one that does not get enough attention: the PFM challengers never won the most profitable banking customers. Mint, Credit Karma, and Rocket Money built mass-market businesses on the back of consumers who already had a fragmented financial life and limited assets to coordinate. Wealthy households did not move their banking relationship to Mint. They stayed at Chase Private Client, at Morgan Stanley, at Goldman Private Wealth, at the regional private banks. The customers worth the most to the financial system, by deposits held, by lending volume, by wealth assets under advisory, by lifetime revenue, stayed inside the institutions. Mint was eventually sunset by Intuit in 2024 because it could not monetize. The PFM aggregator model captured eyeballs and never captured the wallet.

In AI, the wallet is where the value will land. Generic personal finance chat is a commodity. The work that is worth doing, fiduciary advisory, household balance sheet planning, credit underwriting, fraud defense, estate and tax coordination, happens at the customer relationship inside an institution. That work is structurally protected by the same forces that kept wealth banking inside the banks for the past forty years.

What this means for bank and fintech leaders

For incumbents

The strategic implication for incumbents is that the window for embedding AI into the bank's own customer surface is open now, and it closes when ChatGPT, or whatever its 2028 equivalent is, earns the ability to act. Distribution remains a real risk. ChatGPT is approaching one billion weekly active users; the two largest U.S. banks combined have less than twenty percent of that reach. If the bank is content to remain an API behind ChatGPT's chat box, the bank becomes invisible to its own customer. The bank that wants to be the destination has to give the customer a reason to come to its app and stay, and that reason is the AI that does things the external chatbot literally cannot do.

For neobanks

For neobanks, the same logic applies with a different starting position. Chime, Robinhood, SoFi, and Dave already have an AI-native operating posture, native ledger control, and a customer base that came to them expecting digital-first delivery. Their advantage is not data volume; they have less of it than a top-five bank. Their advantage is pace and a willingness to integrate AI into the act of money management itself. Chime's recurring-spend detection runs natively on Chime's own ledger, not on borrowed Plaid data, because Chime is the bank of record for that customer. The deposit relationship is the data moat. Whoever holds it wins the AI layer.

For the AI labs

The implication for OpenAI, Anthropic, Google, and the rest of the AI lab tier is more nuanced than the financial press has framed it. The path to value in consumer finance runs through licensed financial institutions, not around them. That is a feature, not a bug, for the labs that recognize it. The Plaid integration is a useful demonstration product, a habit-builder, and a customer-acquisition channel for ChatGPT Pro. It is not a path to becoming the consumer's primary financial relationship, because the legal and economic structure of consumer finance does not allow that path to exist. The model layer, however, is being rented by every bank in the country and every neobank with an AI roadmap. That is a generational infrastructure business. The labs will be enormously valuable. They will simply not be the customer's bank.

The customer who can ask any chatbot any question about money is also the customer who will demand that the chatbot do something about it. The institution standing behind that "doing" is the institution that wins the relationship. The institution standing behind the model that powers the doing will win on a different scoreboard.

Banks will embed AI. AI will not embed banking.

Sources

• TechCrunch, "OpenAI launches ChatGPT for personal finance, will let you connect bank accounts," May 15, 2026. techcrunch.com
• Plaid, "What ChatGPT's new experience signals for digital finance," May 2026. plaid.com
• Plaid Transactions API documentation. plaid.com/docs
• Financial Data Exchange; FDX Spring 2025 API v6.4 release. financialdataexchange.org
• Plaid Identity API documentation. plaid.com/docs
• "Required Rulemaking on Personal Financial Data Rights," 89 Fed. Reg. 90838 (Nov. 18, 2024). federalregister.gov
• American Bankers Association Banking Journal, "Kentucky federal court enjoins CFPB from enforcing current 1033 final rule," November 2025. bankingjournal.aba.com
• "Personal Financial Data Rights Reconsideration," 90 Fed. Reg. 40909 (Aug. 22, 2025). federalregister.gov
• American Banker, "JPMorganChase reaches deal to charge Plaid for customer data," September 2025. americanbanker.com
• Bloomberg, "Plaid to Pay JPMorgan for Customer Data Amid Open Banking Feud," September 15, 2025. bloomberg.com
• Bank of America Newsroom, "BofA AI and digital innovations fuel 30 billion client interactions," March 2026. newsroom.bankofamerica.com
• VentureBeat, "Wells Fargo's AI assistant just crossed 245 million interactions." venturebeat.com
• JPMorgan Chase Technology Blog, "LLM Suite, American Banker Innovation of the Year." jpmorganchase.com
• Google Cloud Press Corner, "Citi Wealth Unveils Citi Sky," April 22, 2026. googlecloudpresscorner.com
• Morgan Stanley Press Releases, "Key milestone in innovation journey with OpenAI." morganstanley.com
• JPMorgan Chase 2025 Annual Report (SEC filing). sec.gov
• Yahoo Finance, "OpenAI tops $25 billion in annualized revenue." finance.yahoo.com