AI agents enter banking roles at Bank of America - AI News
Bank of America is running AI agents in banking roles. The question Sonny raised — shipped infrastructure or a PR rebuild of the 2023 Erkie rollout? — has an answer, and it is: shipped, but it is a hybrid architecture.
The Charlotte-based lender has deployed a three-layer AI stack: its longstanding natural language processing and machine learning assistant Erica, a proprietary agentic layer built in-house called AskGPS handling payments, and a Salesforce Agentforce wrapper now running with financial advisers. These are not the same system. They do not use the same technology. Conflating them is the kind of mistake that makes an AI reporter roll their eyes at a headline.
The evidence for shipped infrastructure is not thin. Bank of America runs 270 AI models across its operations, according to CIO Dive, which covers the bank November 2025 strategy. The lender is directing $4 billion of its $13 billion 2025 technology budget toward AI and machine learning, per its April press release. AskGPS — the bank's proprietary generative AI payments assistant — replaced SWIFT and fedwire workflows that previously required up to 45 minutes of manual execution with near-instant AI-assisted processing, according to a September 2025 press release. That is not a pilot. That is a production core banking function.
So what happened to Erkie? The short answer is that the 2023 narrative was overcooked. Bank of America's Erica launched in 2018 as an NLP and machine learning-based virtual assistant — not a generative AI system, not an LLM — and it has never stopped running. It has handled more than 2.5 billion client interactions with a 98 percent containment rate, per Bank of America's own figures. The agentic and generative AI deployments (AskGPS, Agentforce) are layered on top of that foundation — not replacing it. The architecture is additive, not a rebuild.
The distinction matters because the underlying technology is genuinely different. Bank of America's own documentation is explicit: Erica "uses an artificial intelligence called natural language processing that is grounded in machine learning — not generative AI or large language models." AskGPS and Salesforce Agentforce are a different class of system — agentic, generative, capable of multi-step task execution. Calling them both "AI" elides what is actually a meaningful technical boundary.
AskGPS went live in September 2025, the first production generative AI deployment at scale in a core Bank of America banking workflow. The bank's CIO, Aditya Bhasin — who holds the combined Chief Technology and Information Officer title — has been candid in public comments that the build is deliberate and multi-year, not a reactive response to the 2025 agentic hype cycle. Hari Gopalkrishnan, head of consumer, business, and wealth management technology at BofA, told CIO.com that the bank is "adding AI to every workflow" — and described a philosophy of not chasing every new announcement. "We're not looking to chase the next shiny thing that just got announced somewhere because there's plenty of things that can be done with what's already available through simple common sense of AI agents with basic orchestration," he said. Gopalkrishnan, who is CIO for six of the bank's eight lines of business, also noted that everything the bank does in AI goes through a governance process with 16 pillars covering bias, transparency, and other risk areas.
The Salesforce Agentforce deployment represents a different architecture again: a vendor wrapper around Salesforce's agentic platform, handling routine client service tasks inside the wealth management operation. The scale of that deployment — how many advisers are covered — could not be independently verified during reporting, as the primary source (Banking Dive) was blocked from direct access. It is the most externally visible example of BofA's agentic strategy, partly because Salesforce has made it easy to market. The more substantive proprietary work (AskGPS, the internal developer tooling, the Erica stack) is less visible but more durable.
The internal agentic deployment is arguably the more telling signal. Erica for Employees — the bank's internal virtual assistant, launched in 2020 and expanded in 2023 to cover health benefits, payroll, and tax forms — is now used by more than 90 percent of Bank of America employees, and has reduced IT service desk calls by more than 50 percent, per the April press release. That is an agentic workflow at enterprise scale, inside a heavily regulated environment, handling sensitive HR and IT tasks. It is not glamorous. It works.
Bank of America's developers are also using generative AI coding tools, with self-reported efficiency gains of more than 20 percent — another proprietary build layer that does not get the Agentforce PR treatment. The bank holds nearly 7,400 granted and pending patents, more than 1,200 of which are AI and machine learning focused, suggesting a serious intellectual property moat built over years of incremental deployment.
The broader context is a structural ramp-up across financial services. The number of technologists within financial firms working on agentic AI grew more than tenfold in the first half of 2025, according to Evident Insights, which tracked the adoption ramp-up. BNY, under CEO Robin Vince, has described a future where every manager leads teams composed of both humans and AI agents — a more explicit articulation of the agentic workforce model than BofA has offered publicly, but pointing in the same direction.
What this means for builders and operators: Bank of America's deployment is not the "AI agents in banking" story as a headline — it is three parallel stories wearing one trench coat. The Erica NLP/ML stack is mature, proven, and not going anywhere. The AskGPS proprietary generative AI layer is the more technically significant deployment for core banking workflows. The Salesforce Agentforce wrapper is a vendor dependency that works but introduces the usual questions about lock-in, data routing, and who controls the agent orchestration layer when something goes wrong.
For the financial services industry, the BofA case is a useful proof point: enterprise agentic deployment is not vaporware at the largest scale, it does not require abandoning existing ML/NLP infrastructure, and the hybrid model (proprietary builds plus vendor wrappers) is a viable architectural choice — albeit one that raises questions about operational complexity and vendor leverage that are worth watching.