It is easy to forget, watching the AI patent race, that the heaviest filers are not all chipmakers and model labs. In the week ending 25 May 2026, every one of the ten newly published U.S. patent applications carrying the Bank of America (BofA) name was filed by the bank itself — and the AI-flavored ones tell a coherent story about where a large retail and commercial bank is pointing its research dollars. A published application is not a product and not a grant; it is roughly an 18-month-delayed snapshot of what a company was funding. Read that way, this batch is a forward signal that BofA is investing to put generative and agentic AI directly inside the machinery that moves customer money and data.

The most forward-looking filing in the set is US20260140771A1, an application for directing resource allocations using an artificial-intelligence agent. The abstract describes something more ambitious than a chatbot:

In particular, the system may generate a virtualized community of users based on anonymized and/or virtualized versions of users based on historical user data.— System and method for directing resource allocations using an artificial intelligence agent, US20260140771A1

The application goes on to describe each virtualized user as an AI agent that interacts with a target customer to give feedback on their preferences and goals, assembling a personalized group of agents tailored to that customer. Strip the patentese and it describes synthetic, AI-driven stand-ins built from historical data that a customer can query and that nudge how resources get allocated. Whatever ships from it, the filing is evidence that BofA was funding agentic AI — autonomous software agents acting on a goal — as a customer-facing concept, not just an internal efficiency tool.

Generative AI moves from the demo to the back office

The agentic filing does not stand alone. US20260140974A1 covers data enrichment using AI-based interactive query generation: machine-learning models scan customer data records for gaps, and generative AI then writes user-specific questions to fill them, iterating until the missing data is captured. That is generative AI pointed squarely at one of banking's least glamorous and most expensive problems — incomplete and stale customer records that drive everything from compliance to credit decisions. The same two inventors, Katherine Dintenfass and Charles Phillip Valentine, appear on both the data-enrichment and the AI-agent filings, which marks this as a single research thread rather than scattered ideas.

It is worth pausing on what kind of company files this. Bank of America is a buyer of AI, not a model lab — it does not sell a foundation model, and its competitive pressure comes from JPMorgan, Wells Fargo, and the fintechs, all of whom are making the same build-versus-buy choice. A bank that merely licensed a vendor's chatbot would have little reason to file its own patents on how agents are constructed from customer data. Filing on the construction itself suggests BofA wants defensible ownership of the specific way generative AI touches its workflows, which is a different posture from renting capability off the shelf. The filings do not prove that strategy succeeds; they document that the bank was investing as though it intended to own the integration rather than outsource it.

Two more applications, US20260141743A1 and US20260141741A1, share an inventor team and cover the document-processing layer of payments. The first runs an optical-character-recognition data file through at least two machine-learning engines operating in parallel, generates a confidence score, and processes the result over a real-time settlement rail; the second uses an AI system to auto-format that image data into XML and route it to a compatible settlement rail based on the confidence score. Together they describe AI deciding, in real time, whether a scanned payment document is trustworthy enough to settle and how to route it. And US20260142887A1 applies an AI engine to network-event monitoring and remediation, forecasting an event and generating fixes — the operational-resilience side of the same investment.

What the cluster signals, and the caveats

The directional read is straightforward and worth stating without overreach. A bank's patent estate has historically skewed toward authentication, fraud, and transaction plumbing; this week's AI cluster suggests BofA is extending that estate into generative and agentic AI applied to its own operations — customer interaction, record-keeping, payments document handling, and network resilience. The filings point to AI being treated as infrastructure to embed, not a feature to bolt on. That is a different posture from a model vendor's; it is a buyer and integrator of AI capability staking claims on the specific banking workflows it wants to automate.

The caveats cut the way they always do with publications. These are applications, not granted patents — the claims can narrow before issue, or never issue at all, and a filed concept is not a deployed system. The roughly 18-month publication lag means this work reflects spending decisions made closer to 2024 than a live readout of 2026 operations. Patent activity also says nothing about adoption: a bank can file extensively on agentic AI and still deploy it cautiously under regulatory and model-risk constraints that a tech vendor does not face. What the records establish as fact is narrower and still meaningful — that in this single week, the entirety of BofA's published patent activity was its own, and that the AI portion of it concentrated on generative agents, generative data enrichment, and AI-gated payments processing rather than on any one narrow tool. For a question a lot of people are asking — are the big banks building real AI or just buying it? — the filings are one concrete, dated piece of evidence about where the money was going.