What JPMorgan’s AI spending means for smaller banks

JPMorgan Chase spends about $2 billion a year on AI, and Jamie Dimon said in October 2025 that the spending is already paying for itself in cost savings, as Semafor reported. Most banks can’t spend at that scale. This post covers what a midsize bank, lender or wealth manager can take from it, and where I’d start.
The short version: you don’t need a giant budget to get real results from AI. You need the right foundation and a narrow focus.
What large banks are actually doing with AI
Most of what the big banks are doing with AI is practical work inside existing operations:
- Reading and summarizing documents: contracts, filings and internal communications that used to take hours of manual review.
- Assistants for developers, legal teams and customer-facing staff.
- Fraud and anomaly detection that learns from new patterns, in addition to fixed rules.
- Customer support tools that answer routine questions and hand the rest to a person.
- Code generation and internal tools that make their technology teams more productive.
None of this requires a bank’s own AI research lab. Most of it can be built today on the data and systems a midsize institution already has, if you know what to build and how to build it.
Why this matters for everyone else
Regional banks, wealth managers and fintech companies are watching this and asking a fair question: “How do we do our version of this?”
You don’t need thousands of engineers or your own AI model to start. You need to know three things:
- where AI creates a measurable gain in your business,
- how to apply it to the data and workflows you already have,
- how to run it under the controls your regulators expect.
Three places to start
1. Answering questions from your own documents
AI can now search, extract and summarize information from your own contracts, policies and filings, and answer with a citation to the source. Teams in compliance, legal, underwriting and operations get fast answers they can check. The method behind this is called retrieval: the system finds the relevant passages in your documents first, and the AI model writes its answer only from those passages.
2. Customer intelligence and segmentation
AI models can group customers by how they behave, in addition to who they are. They can predict lifetime value and retention, and flag the customers most likely to leave or to buy more. I wrote about the basics in how to retain and engage your customers.
3. Risk and anomaly detection
In regulated businesses, false alarms are expensive. Good risk models rank alerts so your team looks at the riskiest cases first. At Irth, Bennett built a model that ranks underground excavation tickets by damage risk, and half of all incidents fall in the riskiest 8% of tickets it flags. You can read how in the damage-risk model we built for Irth.
Why more institutions aren’t doing this yet
The reasons I hear most often:
- Internal teams are stretched thin.
- Old systems and data split across departments.
- Pilot projects that never reach production, so the return is never clear.
- Vendors that promise more than they deliver.
Each of these has a fix. I cover them in why AI projects fail to pay off. At Bennett, our engineers work inside the client’s team, build on the client’s live data, and record every result against the starting value. Bennett ONE sends any consequential action to a named person for approval, which matters a great deal in a regulated business.
Final thought
JPMorgan won’t be the only bank to make AI part of how it runs. The bigger change is happening in midsize banks, regional firms and fintech companies that use AI to serve customers faster, reduce overhead and get more from the data they already own.
If you’d like to talk about what that could look like at your institution, the first call with one of our senior engineers is free and takes 30 minutes.