McKinsey predicts that while agent AI could reduce unit costs for banks by 15-20%, the global profit pool could be eroded by up to $170 billion by 2030 if banks fail to adapt their business models.
getty
While the public imagination has been captivated by the conversational abilities of chatbots, a new report from McKinsey & Company suggests that the global banking industry is quietly approaching a more profound transformation: the era of agent-based AI.
According to the report, the banking sector is moving from an era of widespread experimentation to a paradigm shift defined by autonomous agents that can plan and execute multi-step workflows and use tools with minimal human intervention. This transition represents a shift from “hype” to “accuracy.” It’s no longer a matter of the novelty of a machine that can write poetry. It’s about the utility of a system that can autonomously reconcile ledgers and transfer mortgages.
However, this technological leap comes with a stark caveat. McKinsey predicts that while agent AI could reduce unit costs for banks by 15-20%, the global profit pool could be eroded by up to $170 billion by 2030 if banks fail to adapt their business models.
For more articles like this from Forbes, What is Agentic AI and what does it mean for financial services?
To understand how financial institutions are bridging the gap between theoretical potential and production-grade implementation, we spoke with Jonathan Pelosi, head of financial services at Anthropic, Scott Mullins, managing director of financial services at Amazon Web Services, and Steve Suarez, CEO of HorizonX, senior advisor at McKinsey, former global head of innovation at HSBC, and former GF.
Video: Jonathan Pelosi, Head of Financial Services at Anthropic
2026 Trust Horizon
For many years, the adoption of AI in banking has been hampered by a trust gap. In a regulated industry, models that hallucinate facts are a serious liability. Pelosi argued that this gap is rapidly closing as evaluation frameworks evolve.
“At that time a year ago, [researchers] “If you look at the facts, you might have 8 out of 10 facts that are correct. Now you’ll get about 99 out of 100,” Pelosi said.
Pelosi has identified 2026 as the year the industry reaches a psychological and statistical tipping point. He draws parallels with the introduction of self-driving cars. Just as passengers need data to trust self-driving cars, bankers need data to trust their agents.
“Unless you’re a bank, 80% to 99% accuracy is amazing,” Suarez said, adding, “Even with a 1% error, the system will incorrectly report 100 balances out of 10,000. AI in finance should aim for near zero mistakes.”
Video: Scott Mullins, Managing Director, Financial Services, Amazon Web Services
Beyond AI tourism
As technology matures, the industry’s approach to implementation matures with it. Mullins observes that banks are moving away from AI tourism and conducting pilots just to make a case for innovation.
“If what you’re trying to achieve is just, ‘I want to experiment with artificial intelligence,’ that’s not a real business outcome,” Mullins says. “People see the most value when they have a very specific business outcome in mind.”
This shift from “wow” to “how” is pushing banks toward what Pelosi calls “unsexy.” The most impactful immediate use case is not fancy chatbots, but major mid- and back-office operational improvements.
For more articles like this from Forbes, Legacy bank quietly building the future of finance.
“Unsexy” revolution
One of the most important applications of agent AI is modernizing aging infrastructure in industry. Many financial institutions still rely on COBOL-based systems created decades ago.
“It turns out these institutions are built on legacy code that’s 40, 30 years old, and honestly, people don’t even know how to code anymore,” Pelosi said.
He said Anthropic’s model is now successfully modernizing this legacy code, effectively reading and refactoring millions of lines of old programming into modern languages.
Similarly, compliance workflows such as KYC are moving from human-intensive processes to agent-driven automation. Mullins points to compliance reporting and risk management as areas where agents can significantly reduce manual intervention while improving accuracy.
However, integrating these agents requires overcoming the reality of “fixing the flight mid-flight,” as Mullins described it. Banks cannot upgrade by shutting down their core systems. AI agents need to be integrated into live, mission-critical environments.
Disruptive Threat: Shopping Agent
While banks focus on internal efficiency, the McKinsey report highlights a significant external threat: the rise of shopping agents.
Historically, banks have profited from customer inertia. It was very difficult for consumers to constantly switch accounts to find the best yield. Agentic AI is ready to remove that friction. McKinsey predicts that consumer-facing AI agents will soon be able to autonomously monitor interest rates and move deposits into top-tier accounts.
If these agents shifted just 5 to 10 percent of checking account balances into high-yield accounts, the industry’s savings profits could decline by more than 20 percent. This trend forces banks to not only compete with other banks, but also with the algorithms that manage customers’ financial lives.
Governance: Human Stakeholders
To avoid some of the risks, both Pelosi and Mullins emphasize the need for human-involved governance. The goal is not to replace bankers, but to sandwich AI agents between layers of human oversight.
“Machines can do 80 to 90 percent of the heavy lifting, while humans still have the added benefit of making sure the checks and balances are in place,” Pelosi said.
Mullins advises CIOs to take a “golf bag” approach to this technology, leveraging different models for different tasks rather than relying on a single vendor. This allows banks to choose the most secure and accurate tools for their specific workflows, ensuring that governance evolves with technology.
Key points for bank executives
1. Target “unsexy” people to make an impact
Stop chasing novelty. Invest directly in “problematic” mid-office bottlenecks, such as legacy code modernization and automated compliance reporting. These areas offer the clearest path to the 15-20% cost savings predicted by McKinsey.
2. Prepare for algorithmic competition
Recognize that customer inertia is ending. As shopping agents begin to automate the transition, banks will need to move from broad segmentation to “one segment.” Use internal AI to proactively deliver hyper-personalized value to your customers before external agents move their funds elsewhere.
3. Operationalization of “sandwich” governance
Do not deploy agents without upgrading your monitoring procedures. Implement a workflow where a human defines the goal, validates the output, and an agent handles the execution. As Mullins warned, simply throwing agents into a workflow without coordinating human oversight will fail.
