Customers Bank signs OpenAI deal to automate lending, deposits and payments using its own operational data
Customers Bank, a $25.9 billion regional lender, signed a multiyear partnership with OpenAI in April 2026 to automate lending, deposits and payments, targeting an efficiency ratio move from about 49% to the low 40s. A co-development clause lets the pair build tools OpenAI could sell to other banks, handing a regulated-finance training corpus across the table in exchange for tool access.
Where the rake sits
The rake is owned by Customers Bank. The efficiency gain lands on its own lending, deposit and payments book, and OpenAI is paid as a vendor rather than sharing in the upside: the bank does the enrichment, using the models as tooling. Embedded OpenAI engineers do not make this partner-enriched, because a model vendor is substitutable and the automation persists if the vendor changes. The exposure is elsewhere. The co-development clause hands a regulated-finance corpus across the boundary in exchange for tool access, so the question worth asking is not who holds the rake on the efficiency gain but what was given away to get it.
What happened
- $25.9 billion asset regional lender signs multiyear partnership with OpenAI (announced April 2026); OpenAI engineers embedded inside the bank.
- Stated financial target: efficiency ratio improves from ~49% to low 40s, with higher returns starting 2027.
- Commercial loan cycle compressed from 30–45 days to ~7 days; complex commercial account opening from >1 day to <20 minutes.
- AI already writes ~50% of the bank's software code, saving 28,000 hours (≈15 FTEs not hired)
- Co-development clause: 'We're going to be co-creating enterprise solutions they could potentially sell to other banks' — OpenAI gets real-world use cases inside a regulated FI.
- Initial relationship dates to 2023; CEO Sam Sidhu held a small personal VC stake in OpenAI prior to the commercial deal.
Who is involved
US super-community bank (subsidiary of Customers Bancorp) that holds the internal transaction, deposit and lending data being fed into OpenAI models to automate its own banking workflows.
Parent Customers Bancorp trades on NYSE as CUBI; ~$27 billion in assets and ~1,150 employees (2026).
San Francisco AI lab supplying the foundation models and tooling used to automate Customers Bank's lending, deposits and payments workflows; it is the vendor of the model layer, not the data owner.
Post-money valuation of ~US$500 billion in a March 2026 funding round; ~4,500 employees (2026).
The reading
The enrichment work happens inside Customers Bank: OpenAI's technical teams are embedded on-site in West Reading to build custom AI tools on the bank's own processes, data and institutional knowledge, with the bank running the resulting agents across lending, deposits and payments.
The substrate - the bank's own commercial lending files, deposit onboarding documents and payments records - is worth something because it lets agents make the bank's own operational calls: whether to advance a commercial loan (currently 30-45 days), whether to open a complex commercial account (currently more than a day). The payer of value here is the bank itself, consuming internal efficiency.
On the facts reported, the bank's data stays within the bank; OpenAI engineers work on-site on the bank's processes and data. Sidhu did say they would be 'co-creating enterprise solutions they could potentially sell to other banks in the future' - a potential future boundary question around co-developed tooling, but no such sale is announced.
Not determinable. This reads as an improve case from Customers Bank's seat (internal operational uplift, no external buyer of the data), so the under-capture question - which concerns value left on the table in external deals - does not cleanly apply to this transaction as reported.
Why it matters
This case looks like a productivity story but is actually an improve-route play with a wrap option attached: the bank's operational data is the substrate, the efficiency-ratio delta is the buyer-side P&L proof, and the co-development clause quietly hands a regulated-finance training corpus to OpenAI in exchange for tool access. The rights architecture question — who owns the resulting models and what can ever be sold to competitors — is the decision that should have been made deliberately before signing, and the public reporting suggests it was made by default. Worth pairing with the rake_model discussion: Customers Bank is taking the improve gain now and giving up the sell upside on its own behavioural signal.
The argument this deal tests: The rake is decided before the negotiation begins
Related deals
- PitchBook becomes data partner for OpenAI's ChatGPT for Financial Services, delivering firmographic data via Essential connector2026-09-10
- Google wins $10M bankruptcy auction for Spirit Airlines' enterprise data and code to train AI models2026-08-14
- nference and Rush University System for Health announce long-term partnership for Rush-hosted clinical analytics platform2026-09-02
Sources
- cnbc.com — primary
- en.wikipedia.org — party background
- customersbank.com — party background
- en.wikipedia.org — party background
- en.wikipedia.org — party background
Announced 2026-04-27 (reported)