OpenAI is embedding engineers directly inside Customers Bank to automate core banking workflows including lending decisions, deposit onboarding, and payments. While the partnership offers significant efficiency gains (cutting commercial loan closing from 30 45 days to 7), the…
Sam Sidhu, the CEO of Customers Bank, used an AI clone to deliver his prepared remarks on the bank's last earnings call. A week later, he announced that OpenAI engineers would be embedded inside his $26 billion-asset bank to automate the lending decisions at the core of its business.
The announcement landed as a partnership story. But the accountability question it raises is not a mystery — it is a deliberate gap. The bank bears the regulatory risk. OpenAI's engineers build the system inside the bank's environment. The bank signs off on every deployment. That means if something goes wrong, the liability lands on the bank's balance sheet, not OpenAI's.
What OpenAI gets is more durable than a licensing fee: real-world training data from a regulated institution, and a reference case it can show to every other bank on the fence about AI adoption.
OpenAI's engineers will work on-site at Customers Bank building custom AI for lending, deposit onboarding and payments, the company said in a press release. The work involves end-to-end workflow orchestration, with AI handling document collection, credit memoranda, legal documents and post-closing collateral monitoring. Bankers will focus, Sidhu said, on structuring deals and servicing clients.
When you have an autonomous agent, you're essentially creating a digital worker that can work around the clock, he told CNBC.
The regulatory framework has not caught up to the deal structure. The OCC issued updated model risk guidance on April 17, 2026, and it explicitly carves out generative AI and agentic AI models — they are not within the scope of this guidance. The OCC, Fed and FDIC plan to issue a future request for information on AI, but for now, banks deploying embedded AI vendors in live lending workflows are operating without public regulatory guidance on what the liability structure means in practice.
Banks are subject to fair lending laws, anti-discrimination statutes and a web of state and federal oversight. When a human loan officer denies an application, the bank's compliance team can trace the decision, audit the reasoning and defend it in court. When an AI system built by a third-party vendor makes or materially assists that same decision, the liability allocation the deal specifies between bank and vendor is clear — but the OCC has not said what it means for the bank's regulatory obligations.
The numbers are Sidhu's, and they are real. The bank told CNBC it expects to cut the time to close a commercial loan from thirty to forty-five days to seven, and to open complex commercial accounts from more than a day to under twenty minutes. The efficiency ratio, currently about 49 percent, will move to the low 40s, Sidhu said. Those are projections from a CEO with an obvious interest in the story he tells about his own bank.
The bank has already saved 28,000 hours of work using AI, equivalent to not hiring about fifteen full-time employees. AI has written roughly half the bank's software code, and 75 percent of the bank's team members use OpenAI-powered tools, Banking Dive reported.
Six US banks — JPMorgan, Citi, BofA, Goldman, Morgan Stanley and Wells Fargo — shed 15,000 employees in the first quarter, ResultSense reported, as AI productivity claims begin landing in earnings results.
The comparison Sidhu leans on is scale. JPMorgan Chase has $4.9 trillion in assets. Customers Bank has $26 billion. The big banks have sprawling global operations, complex regulatory exposure and layers of internal governance that slow AI adoption. Smaller banks, Sidhu argues, can move faster because they face lower regulatory expectations.
The bank expects to deploy AI agents across lending, deposits and payments within six to twelve months. Whether those timelines hold is testable. The loan closing compression from forty-five days to seven is the most public benchmark, and it is the one competitors will point to if it works or comes up short.
OpenAI's chief revenue officer, Denise Dresser, said the company was proud to help Customers Bank build a more intelligent operating model. The bank's press release described the engagement as co-creation. Sidhu told CNBC the goal is end-to-end automated agentic workflow for lending, deposits and payments, and that the bank and OpenAI will co-create enterprise solutions OpenAI could eventually sell to other banks.
If that happens, the first customer is also the proof of concept. If it fails, it becomes the cautionary tale that other banks' compliance teams point to when their boards ask whether AI adoption is worth the risk.
The bank said it wants to keep customer relationships human. It is automating the work that surrounds those relationships, not the relationships themselves. Whether that distinction holds when an AI-assisted decision costs a small business owner a loan she qualified for is a question neither the bank nor OpenAI has answered.