How banks are expanding AI while restricting outside agents

Banks are expanding AI use while controlling access

A bank can deploy AI for employees and customers while restricting access by outside AI agents. JPMorgan Chase is doing both, separating its decision to use the technology from its decision to admit external systems. For banking technology leaders, that separation raises three practical questions: which uses deserve permission, which models justify their cost and which work the bank should do itself.

Banks addressed those questions in earnings discussions about agentic AI, systems that can carry out tasks, and AI’s operating benefits and costs. JPMorgan Chase and Wells Fargo discussed their approaches on Tuesday’s Q3 2026 calls, followed by Bank of America on Wednesday. Their decisions cover access for outside agents, model routing, the assignment of a query to a particular AI model, and in-sourcing, which means bringing work inside the organization. Each changes a different part of how a bank deploys AI.

Outside agents face separate access decisions

Access comes first because an outside agent needs permission to operate within a bank’s systems. JPMorgan CEO Jamie Dimon described the bank as neither “against agents or for agents,” pointing to its adoption for employees and customers. Dimon has a commercial interest in gaining AI’s benefits while protecting customers and the business from its risks. His position separates the value of using agents from the exposure created by admitting them.

Dimon explained the security concern: “because you don’t want rogue agents, and there could be thousands of them.” Thousands describes the possible scale of rogue-agent activity. As the number of agents grows, the bank could face many sources of unwanted activity, making control over entry an important part of deployment.

JPMorgan is therefore developing tools to protect customers while restricting access to its systems. “We’re not going to allow just any outside agent to come inside the company because that’s just too risky,” Dimon said. A provider whose agent is refused entry cannot run it inside the bank, even as JPMorgan expands its own AI use.

Wells Fargo described how it would assess third-party technology by weighing its benefits against its risks. CEO Charlie Scharf said the bank saw opportunities for itself and its customers while “taking a disciplined approach to managing risk.” Scharf also has a commercial interest in finding useful capabilities for Wells Fargo’s customers while limiting the bank’s exposure.

Scharf set out what that assessment would cover: “We will also evaluate how third-party models and agents can provide value-added capabilities to our clients, while at the same time considering the potential risk these tools can introduce for both our clients and Wells Fargo,” he said. A capability can help a client while creating exposure for the institution providing it, so the assessment covers both parties. Wells Fargo described an evaluation process; JPMorgan stated a restriction on entry. The statements establish different approaches to access.

Model routing makes cost control part of deployment

After a bank approves an AI use, it still has to decide which model should handle the work. JPMorgan is pursuing a model-agnostic strategy, meaning it aims to avoid dependence on a single model, to preserve its ability to switch providers and manage spending. Dimon pointed to routing tools that direct queries toward the quickest, cheapest models. By choosing a model as work is assigned, the bank can manage costs query by query.

Routing works when the model fits the request, because different queries make different demands on a system. “You don’t need the most expensive thing for every query you have and very often people just see the one they’re comfortable with,” Dimon said. Familiarity can lead an employee to choose a more expensive model even when a cheaper option could do the job. So model selection affects spending alongside the prices negotiated with providers.

Moving beyond habit still requires judgment about what each task needs. The quickest or cheapest model is suitable when it meets those requirements, so cost control depends on assigning the work well. For engineering leaders, the practical step is to define the task’s requirements before choosing a model, then use routing to apply that choice whenever a query is submitted.

Bringing work in-house changes what banks buy

Model selection controls one part of AI spending; deciding who does the work changes another. On Citi’s third-quarter earnings call Tuesday, CEO Jane Fraser said AI was helping the bank launch products faster and at lower cost, with an average 30% to 40% increase in product-launch speed. Fraser has a commercial interest in presenting Citi’s technology investment as improving its operations. Her launch-speed figure measures delivery performance, while her account of lower vendor spending concerns a separate sourcing decision.

Fraser tied the purchasing change directly to work moving inside Citi. “We’ve been in-sourcing work,” she said, followed by: “We’ve reduced our tech vendor spend by 30% year-over-year as a consequence, and that’s still got a long way to run.” She attributed the reduction to in-sourcing and presented it as an ongoing change. Technology-vendor spending is a distinct measure from the cost of launching a product or processing a model query.

Citi’s engineering-tool adoption shows how work can move inside a bank while its teams continue to use outside technology. Citi reports that it has 14,000 engineers using Devin, its main agentic coding tool, a system that carries out software-development tasks, to make internal work more efficient. Buying a coding tool and outsourcing execution assign responsibility differently: the tool supports the bank’s own teams as they do the work. Devin’s use is therefore consistent with Citi retaining more responsibility for execution.

Beyond software development, Citi reports that it is reengineering 100 manual processes using AI. Reengineering changes how work is done, extending adoption into recurring operations. Because the effort is underway, the process count describes how much work is being redesigned. It measures a different stage of change from the reported improvements in launch speed and vendor spending.

These measures answer different questions about progress: engineer adoption measures participation, process redesign measures the scope of operating change, and launch speed measures delivery performance. Fraser’s average speed increase describes an overall result across the launches measured; it does not imply an equal improvement for every product or an equivalent percentage reduction in development time. Keeping the measures separate also preserves her explanation for the vendor savings: Citi is bringing work in-house, with coding tools and process redesign supporting its broader internal operations.

Operating benefits take different forms across banks

Citi’s delivery and purchasing measures differ from the workforce changes described by Bank of America. On Wednesday’s Q3 2026 call, CEO Brian Moynihan presented AI as supporting growth across the business, a claim that serves the bank’s commercial interest in demonstrating the value of its investment. Executives said automation and simpler workflows were reducing the need to hire replacements when roles became vacant. This concerns replacement hiring as work changes; it is a different measure from a quantified reduction in headcount or a layoff program.

Alongside that discussion of operating change, Bank of America faced a separate compliance event on Tuesday. The bank received a $1.2 billion fine for violations of anti-money laundering and sanctions laws, described as the third-largest fine ever imposed by the Treasury Department. The fine concerns compliance with those laws and is unrelated to the AI benefits executives discussed.

Morgan Stanley’s executives described AI’s operating contribution more broadly than Bank of America’s account of replacement hiring. CEO Ted Pick characterized its contribution to efficiency, productivity and resiliency as incontrovertible, while CFO Sharon Yeshaya said the technology was “fundamentally changing what our teams do,” emphasizing changes in team activity. Both executives represent a bank with a commercial interest in demonstrating gains from AI investment. Their qualitative descriptions identify changes in how teams operate; combining them with another bank’s hiring, launch-speed or spending measures into a common measure of financial return on AI spending would erase the differences between the reported outcomes.

Key takeaways for decision-makers

• Govern outside-agent access separately: JPMorgan is expanding internal AI use while restricting outside agents. Bank security and platform owners need explicit entry controls and assessments of the risks third-party agents introduce for customers and the institution.
• Route queries by task requirements: JPMorgan is using model routing to manage speed and cost while preserving provider choice. Engineering leaders can define performance requirements for each task and route queries to lower-cost models that meet them.
• Review which work belongs in-house: Citi attributed a 30% year-over-year reduction in technology-vendor spending to in-sourcing, alongside wider use of AI coding tools. Technology and sourcing leaders can assess where those tools support internal execution and change the need for outsourced work.
• Keep AI outcome measures distinct: Citi reported faster product launches, while Bank of America described reduced replacement hiring and Morgan Stanley cited broader operating improvements. Finance and technology leaders need separate measures for adoption, delivery speed, spending and workforce effects to assess each claim accurately.