Amy Avery, Managing Director, Analytics, Modeling and Insights, took her job at Bank of America because of a number. When she interviewed at Bank of America, she was told that the bank interfaced with, at the time, 67 million clients. “Gosh, that’s so much information,” she remembers thinking. “Think about what you could do with that.” She started in January 2020. Two months later, the pandemic made that abstraction very literal: the bank suddenly needed to know, in real time, how its customers were doing, thinking, and coping. Avery’s job was to figure out how to answer that.
Michelle Boston, Head of Data Management Technology & Enterprise Architecture, arrived by a different route entirely. She built her career in enterprise technology, rose to CIO of a startup that was eventually built and sold, and came to Bank of America first as a contractor to lead an information architecture practice. “Data has always kind of been in my blood,” she said. At Bank of America, she works at a scale few other organizations have and builds the platforms that serve as the enabling force for Avery’s work.
Despite a very different set of starting points, the two describe a partnership that has essentially erased the line between their jobs. “We probably know each other’s jobs better now than before generative AI showed up”, Avery said, because the pace of the last two years has forced her strategy team and Boston’s engineering team to make decisions in near lockstep.
Listen to the full episode to hear how Avery and Boston have built a shared language across the two functions, and how they’re stress-testing it against a technology cycle that seems to wait for no one.
Figure [FIGR] released its second-quarter earnings, showing that $1 billion in loan applications were flowing through its platform every week by early July. It’s a milestone that would have been hard to imagine a few years ago, when Figure was still largely viewed as a digital lender using blockchain to make HELOCs (Home Equity Lines of Credit) faster. The business looks different now.
In Q2’26, Consumer Loan Marketplace volume jumped 132% to $4.3 billion. Figure Connect, its blockchain-based marketplace for private credit, accounted for $2.8 billion of that volume. The company added 102 origination partners during the quarter, taking the total to 489, and expects $4.8 billion to $5.2 billion in marketplace volume in Q3.
Those numbers show how Figure has spent the last few years changing how it operates. It started by building a better way to originate loans. Then it built infrastructure to move those loans into capital markets. After that, it opened those rails to other originators. And now AI is becoming another important layer in that system.
In a nutshell, Figure is turning lending and capital markets into a structured system and is now making that system intelligent and programmable.
Figure was never really trying to stay a lender
Figure’s history makes it easy to put the company in a HELOC-shaped box. However, its lending business was increasingly becoming an entry point rather than an end goal.
The firm launched its own retail HELOC business in 2018 before moving toward a B2B model, allowing outside originators to use its technology. That eventually became Figure Connect, its marketplace for private credit. And the economics changed with it.
Two borrowers can have the same income, the same account balance, and even the same overdraft history, yet represent very different credit risks. Traditional models often struggle to tell them apart. Plaid believes foundation models can.
This year, the company introduced a two-layer foundation model architecture for financial data. The transaction model makes sense of individual financial events, while the sequential model looks at their order, cadence, and relationships, similar to how a large language model understands words in context rather than in isolation.
That richer understanding is improving lending, fraud detection, and payment risk across Plaid’s network of more than 12,000 financial institutions and 9,000 apps. Its transaction foundation model improved income classification accuracy by 48%, while its sequential foundation model reduced credit default risk by 13.6% at the same approval rate. This foundation can also feed into products such as LendScore, Plaid’s real-time cash flow-based credit risk score, which reduced lending risk by 41% compared to a traditional benchmark, without sacrificing approvals.
The approach earned Plaid the Data Innovation Award at Tearsheet’s AI Innovation Awards 2026. Tearsheet spoke with Suddu Seshadri, Plaid’s Head of Data & AI, about why Plaid moved beyond data connectivity, what drove its decision to build foundation models for finance, and how AI is changing what makes financial data useful.
Suddu Seshadri, Head of Data & AI at Plaid
Q: Plaid has evolved from a data connectivity company to an intelligence platform. What drove this change?
Suddu Seshadri, Plaid: Plaid sees financial activity across 12,000 institutions and 9,000 apps. As fintech took shape, the first challenge was bringing financial data online and enabling people to access it and connect it. Plaid powered this shift, providing the infrastructure to make access possible.
Now, with the digital financial ecosystem flourishing, the next phase will be intelligent finance, where AI reasons on top of this financial data. From our connectivity, we’ve built dimensionality around identity, connections, and transactions. This helped build an incredibly deep network and helped us answer critical questions around fraud, payment risk, and credit. The compounding effects of this network are fueling our next evolution.
Q: What convinced Plaid that financial data needed foundation models, not just better traditional risk models?
Suddu Seshadri, Plaid: The consumer demand for self-driving money is clear. In Spring 2026, we released consumer research with the Harris Poll that revealed 86% of adults in the US were already using AI to better understand and manage their money. And they’re not just using it; they want AI to do more for them. As a result, we want to ensure our customers have access to the richest data available as they build new intelligent finance experiences. While traditional models can perform very well on a defined task, each one often has to rebuild the same underlying understanding of transactions, income, cash flow, and behavior.
Financial data also has characteristics that make this difficult: transaction descriptions can be cryptic, labeled data is limited, and timing often changes the meaning of an event. Foundation models allow us to learn a reusable representation of financial activity once and adapt it across many tasks.
Q: What makes Plaid’s approach to building its foundation models fundamentally different?
Suddu Seshadri, Plaid: We build models to help improve how financial services work: our models are designed to make financial products more personalized, useful, and safe. From a decade-plus of powering digital finance, Plaid has built a powerful financial data network. Now, we’re using it to build models that can reveal and make sense of the transactions and patterns that make up a financial life.
We combine network data with models built specifically for financial activity. We are not asking a general-purpose model to guess from raw bank text. Our transaction model learns the economic meaning of individual events, while our sequential model learns order, timing, and changes in behavior over time.
That shared understanding can then improve product-specific systems across payments, fraud, and lending. For example, our transaction foundation work has improved income classification by 48% and loan payment detection by 14%.
Q: How does one shared AI foundation improve products as different as lending, fraud detection, and payments at the same time?
Suddu Seshadri, Plaid: Lending, fraud, and payments rely on many of the same underlying concepts: income, recurring obligations, cash-flow stability, account behavior, and changes over time.
The shared foundation turns those concepts into reusable representations. Each product then combines them with its own data, labels, thresholds, and controls.
That means an improvement in the underlying representation can benefit several products, without treating a lending decision, a fraud decision, and a payments decision as the same problem.
Q: Has AI changed what “good financial data” looks like? If so, how?
Suddu Seshadri, Plaid: The fundamentals have not changed. Good financial data still needs to be permissioned, accurate, current, reliable, and traceable. What has changed is how much value models can extract from context. A transaction means more when you understand the activity that came before it, its timing, its relationship to other events, and whether behavior is changing.
Sequence and relationships were always important. Modern models make it possible to use them at a much greater scale.
A lot of the conversations around artificial intelligence tend to circle back to one question: which institution has built the smartest, most capable model?
Bank of America argues the model quality isn’t the point; an AI feature is only as reliable as the data behind it. That reliability determines whether AI can be trusted to support decisions or should be kept out of high-stakes ones entirely.
Bank of America’s thinking on AI comes into focus through three perspectives: Matthew Davies, Head of Global Payments Solutions at Bank of America’s warning about fragmented data, EricaAssist’s role as an employee copilot grounded in years of client history, and CEO Brian Moynihan’s caution toward frontier models. All point to the same philosophy: AI should be treated as a supportive step in the process, not the one making high-stakes decisions.
Capital misallocation is the real risk
Bank of America has built its AI strategy to avoid the trap of treating AI as a shortcut to efficiency, an approach that often results in costly, misguided AI investments. Davies notes that the pressure every company now feels to adopt AI fast or risk falling behind is itself the real danger. “The biggest risk and challenge is misinvestment rather than underinvestment,” he says.
Bank of America has traced that misinvestment risk back to fragmented data scattered across ERP systems, treasury platforms, bank portals, and acquired businesses. Because these systems speak different languages, even the most advanced AI layered on top struggles to reconcile the underlying information. Davies emphasizes that without high-quality, standardized data, there’s no real foundation for automation, forecasting, or any AI solution built on top of it. Tearsheet’s recent reporting on Intuit Credit Karma echoes the same idea. Rather than building standalone AI assistants, the company built them on a shared view of a customer’s financial life, recognizing that better AI starts with better data and context.
For Bank of America, improving data quality delivers value long before any AI enters the picture. Standardized data alone reduces the manual reconciliation, duplicate entries, and reporting errors that consume hours of finance teams’ time every week. Davies outlined the bank’s order of operations: standardize the data first, automate repetitive tasks next, and only then pursue bigger AI initiatives.
EricaAssist: Inside Bank of America’s own data-first playbook
Employees at financial institutions are using AI tools their employers never approved for tasks that touch sensitive customer data. This unsanctioned use of AI, often called shadow AI, is not a toothless problem. In May 2026, an employee at Pennsylvania-based CB Financial Services, parent company of Community Bank, uploaded a file containing customer names, Social Security numbers, and dates of birth into an unauthorized AI application while preparing a presentation. The employee bypassed the bank’s approved AI tool for a personal account on a personal device. The bank caught the exposure quickly enough to get the data deleted before it could be used to train the vendor’s model. Notably, the bank already offered a sanctioned AI tool. The employee simply chose not to use it. The instinct inside most banks and credit unions is to treat this as a policy failure and respond with tighter restrictions, more monitoring, and firmer language in the acceptable-use policy.But that response gets the diagnosis wrong, according to Corey Gross, VP and Head of Data & AI at Q2 Holdings. “When employees bypass a sanctioned tool, they’re telling leadership teams that the approved option isn’t getting the job done.” In his view, the root cause sits upstream of governance entirely: institutions are buying AI tools without first understanding the workflows those tools are meant to support.Instead of writing stricter acceptable-use policies, Gross suggests the more useful question is why employees felt they needed to go around the tool they were given in the first place. “It’s rarely a governance or compliance issue,” he said.
The distinction sounds subtle, but it points to a different set of priorities for any bank trying to scale AI responsibly. It suggests that starting with the design of the work itself is more critical than the framing of the rules that surround it.
Workflow redesign has to come before the AI rollout
Intuit Credit Karma is building AI-driven optimization systems to take the guesswork out of recurring financial decisions by helping users determine their best next move.
With the launch of Credit Karma Intelligence, Debt Assistant, Refund Assistant, and Paycheck Assistant – all sitting on Intuit’s Consumer Platform – the company is creating a suite of AI-powered tools that tackle different financial decisions while drawing on a shared, evolving understanding of a customer’s financial life.
Gurpreet Singh, head of product at Intuit Credit Karma
“Financial decisions don’t happen in isolation,” says Gurpreet Singh, head of product at Intuit Credit Karma. “A recommendation for paying down debt, allocating a tax refund, or using an upcoming paycheck only makes sense when it’s informed by someone’s broader financial situation, including their income, cash flow, existing debt, spending patterns, connected accounts, and the goals they’ve told us they care about.”
Integrated AI-powered financial platform: Singh emphasizes that the firm didn’t build Debt Assistant, Refund Assistant, and Paycheck Assistant as standalone AI tools. Instead, they were built to support the financial decisions people make throughout the year as they manage debt, get paid, receive a tax refund, and make trade-offs between competing priorities. By grounding each experience in the same understanding of a member’s financial situation, goals, and history, the platform can deliver guidance that’s consistent across those moments instead of treating each one like a new, standalone problem.
“We’re not building a collection of individual AI tools, but an AI-powered financial platform that understands where someone stands, adapts as their situation changes, and delivers guidance that’s grounded in their financial data,” says Singh. “That’s what an integrated ecosystem makes possible.”
16,000 jobs are being lost per month due to AI, according to Goldman Sachs data.
Bearing the brunt of this augmentation-automation mindset are Gen Z, who are caught in an untenable “low hire, low fire environment”, according to Federal Reserve Chair Jerome Powell.
But looking at communication coming out of Wall Street, it seems like AI is only making bankers faster, better, and cooler at their jobs.
In today’s story, we look at murmurs and public addresses coming out of Wall Street to see exactly how the C-suite is planning to tackle AI-driven mass unemployment, also known as jobpocalypse.
Standard Chartered CEO says what all bank CEOs are thinking
In an ill-timed slip-up that became a PR nightmare, Standard Chartered CEO Bill Winters, a veteran of 11 years at the bank, described its efforts to streamline operations in a way that drew criticism: “It’s not cost-cutting. It’s replacing in some cases lower-value human capital with the financial capital and the investment capital we’re putting in.”
The clinical term hit a public nerve. Winters has since issued an internal memo as well as a public-facing apology for his choice words, while also stating that most outlets reporting on the story have taken his statement out of context. Here is the statement reproduced in full:
If AI agents can execute payments, what guarantees that they execute the right payments?
The move
American Express has launched Agentic Commerce Experiences (ACE) Developer Kit, which formalizes and verifies user intent before any transaction takes place. The firm has also introduced what it calls an “industry-first” protection against AI agent error, agreeing to cover eligible transactions when an agent executes an authorized but unintended purchase.
For example, a user asks an AI agent to book a “quiet hotel room under $250”; the agent finds a deal and completes the booking, but if it’s next to a busy street, the transaction is valid yet misaligned with intent.
How it works
The new ACE Kit shifts payments from simple authorization to intent-driven execution:
User intent is captured as a structured, verifiable, and enforceable input.
That intent is authenticated and tied to tokenized credentials before any transaction is initiated.
Agents can transact on behalf of card members only within clearly defined, authenticated intent and control layers.
Amex extends purchase protection into agent-executed transactions.
Instead of resolving disputes after the transaction, the system aims to reduce ambiguity before execution.
“The model includes card member enrollment and authentication and gives card members the ability to manage controls directly in the Amex app – using structured intent, spend limits, merchant preferences, and tokenized credentials so the agent can only act within clearly defined boundaries set by the card member,” noted Luke Gebb, EVP and Head of Global Innovation at American Express.
“The core of our approach is that intent is not treated as a loose instruction – it’s treated as a structured, verifiable representation of Card Member intent that the system can evaluate and enforce,” he added.
How is this different?
Traditional payment systems answer one question: Was this transaction authorized?
Agentic commerce introduces a harder one: Did this transaction reflect what the user actually meant?
That gap between execution and intent is emerging as one of the weakest links in AI-driven commerce.
McKinsey estimates that agentic and AI-driven commerce could generate trillions of dollars in economic impact by the end of the decade, but only if trust in automated execution scales alongside it.
Amex’s move directly targets that trust layer. Its closed-loop network provides end-to-end visibility across users, agents, credentials, and transactions, allowing it to link intent, execution, and liability within a single system.
Why it matters
By underwriting agent error, Amex is enabling AI-driven payments, but also pricing and absorbing a new category of risk.
That changes the equation for adoption. Agentic commerce won’t scale simply because agents can transact; it can scale when users trust that those transactions are executed correctly.
In the Chart: Amex’s take on securing the agentic commerce stack
The fintech industry has spent the better part of the last two years racing to add AI to its products. Chatbots have been bolted onto banking dashboards. Summaries have been appended to transaction histories. Assistants have materialized inside apps that users open, at best, once a week. Liran Zelkha, co-founder and Chief Technology Officer of Lili, a financial platform built for small business owners, thinks most of this activity is pointed in the wrong direction. Zelkha has spent years thinking carefully about the relationship between technology and small business owners: customers who are skilled at their craft, pressed for time, and rarely interested in learning a new interface. This customer profile has shaped the architecture of Lili from the beginning, and it informs Zelkha’s view of where AI in finance could genuinely move the needle: The goal is not a better in-app AI experience. The goal is to make a company’s financial capabilities available through whatever AI the customer has already chosen to trust. “The business owner shouldn’t have to learn our app to get value from us,” Zelkha says. “They should be able to ask their AI companion about their cash flow and get a real answer, backed by Lili.” …
Agentic commerce is minimizing the distance between execution, intent, and credit within payments. Functions that once operated in separate layers of the stack are becoming more connected and responsive to one another.
And as that happens, the rest of the stack can’t stay passive. Authorization needs to interpret, not just validate. Credit has to be adjusted per transaction, and infrastructure has to carry context, beyond just credentials.
Across the stack, different players are enabling this shift in different ways, using agentic AI to reduce the friction between how decisions are formed and how they are carried out.
Here’s how it’s playing out.
Stripe and the push to make agents transactable
Stripe is addressing the execution layer of this shift, enabling AI systems to complete transactions once a decision is made.
Its product direction, including work around agentic commerce in addition to stablecoin infrastructure, is aimed at making AI systems economically native. The Agentic Commerce Protocol (ACP), developed with OpenAI, is an attempt to define a shared language between merchants and AI agents so transactions can happen without bespoke integrations for every system.
The firm is trying to normalize the idea that agents will increasingly initiate commerce flows, and the system has to treat that as a normal input.
