Employee ideas drive Bank of America’s plans to double AI spending

Bank of America’s AI expansion combines employee assistance with lower staffing

Bank of America’s use of artificial intelligence (AI) shows how employee assistance and a smaller workforce can develop together. Its developers use coding agents, employees get answers from internal self-service tools, and headcount has fallen as the bank controls hiring while employees leave. Management decisions connect these changes: leaders choose how to use the capacity AI creates, which investments to fund and how much responsibility employees retain.

The financial question is how quickly that spending pays off. Speaking Wednesday at a BofA Securities conference, Bank of America Co-President Jim DeMare said, “The question was really getting at, are we going to see returns quickly enough?” DeMare represents a bank with a commercial interest in showing that its investment improves operations and encouraging employees to use the tools. His account combines measurable gains with the difficulties of putting AI into everyday work.

Those difficulties extend across banking as financial-services AI spending rises. Consulting firm Accenture found that only 20% of bank leaders see widespread, sustained value from their AI initiatives. Accenture’s consulting business also benefits from demand for help implementing AI, so its finding carries a commercial interest of its own. At BofA, the practical questions start with what work has changed and how the bank measures the result.

Employee-assistance tools coexist with lower headcount

Software development offers DeMare’s clearest example of measurable gains. He said technology functions such as development and coding have so far delivered the most measurable AI returns. Across most industries, he put the coding-productivity improvement at about 15% to 20%, and said BofA is seeing the same range as its 20,000 software developers use coding agents, AI tools that can carry out coding tasks with some independence. “That’s clearly identifiable,” he said.

A productivity gain gives managers more capacity to allocate. Depending on the organization’s needs, they can use it to take on more development work, change delivery schedules or adjust future hiring. The staffing outcome therefore depends on management’s decisions about how much work to undertake and how many people it requires.

Internal self-service changes another part of the workload. DeMare said the bank’s AI-powered virtual assistant, Erica, has considerably reduced help desk inquiries by handling employees’ internal requests. Employees who get answers through Erica can continue working without contacting the help desk, reducing demand on support staff. BofA also says Erica handles work equivalent to that of about 11,000 people, a comparison that describes the volume of work the assistant performs.

Beyond those operational measures, Bank of America CEO Brian Moynihan has described a financial benefit across the bank’s applications. At a Barclays conference last week, he said about 140 AI uses had been implemented, “at a cost of $400 million, generating a benefit of $800 million.” Like DeMare, Moynihan represents the bank whose spending and operating decisions these claims concern. His figures describe the combined economics he attributes to those uses.

These measures cover different parts of the business case. Coding productivity describes development efficiency, help desk inquiries describe demand for support, and Erica’s work-equivalence figure describes workload. Moynihan’s cost and benefit figures cover the applications collectively, so managers need to keep those categories distinct when deciding how an operational improvement affects expenses or capacity.

Staffing adds a separate management outcome. While discussing AI benefits, Moynihan said the Charlotte, North Carolina-based, $3.5 trillion-asset bank had moved from about 213,000 employees at the beginning of the year to 209,000. He also put attrition, the rate at which employees leave, at about 8.5%. He explained the decline in terms of how the bank fills vacancies.

Hiring decisions allow the bank to reduce staffing as employees depart. “We’re not laying off anybody. We don’t have to do that. All we do is just manage the hiring carefully,” Moynihan said. Attrition creates vacancies, and selective hiring determines how many are filled. Employees therefore encounter AI assistance within an organization that is also managing a smaller workforce, making their expectations about replacement an important deployment issue.

Fear of replacement makes employee participation an implementation issue

The possibility of replacement can affect employees’ willingness to use the tools. DeMare said Wednesday that “one of the biggest risks to implementation of AI is people being fearful of it and thinking that it’s going to replace them.” For the bank, employee adoption helps turn technology spending into useful work; for employees, it also raises questions about their jobs. Explaining what a tool does helps address those different interests.

DeMare placed that concern within the bank’s broader experience of introducing technology. “That’s not unique to AI,” he said. “It’s apparent every time we try to use new technology.” In his account, implementation means helping employees become willing and able to use a system in daily work alongside the technical task of making it available.

To build that familiarity, DeMare said BofA has given about 95% of the company access to AI tools. He said access has increased familiarity and comfort by giving employees opportunities to learn through direct use. As they get to know the tools, he said, their assessment shifts toward practical benefits: “Then it’s about general productivity, and how is it improving their life at work, and how is it improving workflows.”

That practical assessment also helps the bank identify further investments. Employees know where their work becomes difficult or time-consuming, so their suggestions can connect a proposed tool to a specific task. BofA asks employees for ideas that could make their jobs easier, then reviews them to identify where AI could help.

Employee workflow ideas are helping justify more AI investment

Reviewing employee ideas brings financial judgment into the selection process. Hari Gopalkrishnan, the bank’s chief technology and information officer, said, “ROI is very much a part of the decision-making process.” ROI means return on investment: the value an initiative produces relative to its cost. Gopalkrishnan’s account makes the bank’s case for expanding spending based on the revenue and expense opportunities employees identify.

As the bank assesses those opportunities, more are becoming candidates for AI investment. In an interview this week, Gopalkrishnan said, “as we tag them for AI, it’s turning out that our level of AI investment interest is actually more than twice what it was last year.” The process starts with a proposed improvement to work, then considers AI as a way to deliver it. This sequence ties technology selection to an expected business benefit.

The growing set of proposals is also shaping spending expectations. Moynihan said last week that BofA expects to double its expense budget for AI next year. Gopalkrishnan qualified his own expectation: “We will probably end up spending twice next year [what] we did this year because they’re just good, rich ideas that are now starting to create the return on investment for us.” Both executives linked prospective spending to the bank’s assessment of opportunities, with Gopalkrishnan explicitly pointing to ideas beginning to generate returns.

Those plans sit within a wider technology budget. BofA says it spends about $4 billion each year on new technology initiatives, including AI. The annual total covers the broader technology program, while Moynihan’s planned increase concerns the AI expense budget. Both spending plans are distinct from the costs and benefits he reported for applications already implemented.

Within that budget, Gopalkrishnan argues for funding changes according to their expected value. “Our view is, well, if they’ve got all these ideas with revenues and expenses, then we ought to be able to invest more to give them that value. It happens to be that AI is now increasingly a part of the solution,” he said. For a budget owner, the sequence runs from a work improvement to its expected effect on revenue or expenses, then to the technology and spending needed to deliver it.

Preparation for client conversations offers a concrete example. Gopalkrishnan said thousands of employees now use AI-powered customer relationship management tools, which organize information used to manage client relationships. Before a client conversation, an employee receives data and talking points to support the discussion. AI changes the preparation work, while the employee conducts the conversation and decides how to use the information.

Delivering such improvements also requires choices about who supplies the technology. DeMare said BofA is assessing third parties and multiple providers that could improve automation with AI, while considering more complex and custom possibilities. These choices determine how the bank builds or obtains a capability. A further decision concerns how much freedom the resulting system should have to act.

Human accountability limits how autonomously AI can operate

The question of freedom becomes more consequential as AI agents take on more tasks. Gopalkrishnan is considering how much autonomy, the ability to act independently, the bank can accept. He said improvements in model efficiency and effectiveness could support greater autonomy over time, but tied that expansion to controls on system behavior.

Gopalkrishnan stated the condition directly: “As these models get more efficient and effective, we will want to, over time, expand the autonomy, but we’ll want to do that in a way that the guardrails are omnipotent.” Guardrails are controls intended to keep a system’s behavior within acceptable limits. He added, “We’re not going to do anything until such a time that we have the appropriate guardrails.”

These controls make permission to act a separate decision from the case for investment. A system may promise valuable improvements while still requiring limits on what it can do independently. Under Gopalkrishnan’s approach, the bank must judge whether the controls are appropriate before expanding that independence.

Human review is one such control when AI produces information employees may use. Moynihan said employees remain accountable for that information, with review covering both common sense and factual correctness. “The risk, for us, was really the risk of letting it start giving answers without humans checking to make sure the answer was right, both just common sense, but also literally making sure it’s right,” he said last week.

The client-preparation tools put that responsibility into a commercial setting: information prepared for an employee can affect the next client conversation. “If you give a wrong answer to a client, the client’s going to walk out on you,” Moynihan said. “That will gate its application in some ways.” The cost of a wrong answer therefore constrains where the bank can use AI and when an employee must check its output.

Key highlights

• Separate productivity gains from staffing decisions: BofA reports coding productivity gains of 15%–20%, while headcount has fallen through attrition and selective hiring. Budget owners need distinct measures for workload, capacity and financial returns when deciding how to use AI gains.
• Build employee confidence through use: BofA has given about 95% of employees access to AI tools as it addresses fears of replacement. Deployment owners can build familiarity around specific workflow improvements while explaining how AI affects employees’ responsibilities.
• Tie AI funding to employee-identified opportunities: BofA expects to double its AI expense budget next year as employee suggestions generate investment candidates. Funding reviews need to connect each proposal to a defined task, expected revenue or expense benefit, and delivery cost.
• Make controls a condition of autonomy: BofA ties greater AI independence to appropriate safeguards and keeps employees accountable for checking answers. Technology and business owners need explicit review requirements wherever incorrect output could affect clients.