Banks Are Wrapping AI Agents Around 40-Year-Old Cores, Not Replacing Them

Sutherland, a business-transformation firm, and FlowX.AI, an agent-orchestration vendor, announced on October 19, 2026 that Sutherland will deploy and run FlowX.AI's library of more than 220 prebuilt AI agents on top of existing bank core systems across North America and Europe. The target workflows are specific and unglamorous: customer and commercial onboarding, account opening, lending, mortgage underwriting, KYC and anti-money-laundering checks, financial-crime risk assessment, and customer retention. FlowX.AI's own figures claim one European lender cut manual credit handoffs by 80 percent, and another cut lending operating costs by 40 percent, using the same approach. Neither number required touching the core banking platform underneath.

Why the core stays untouched

Many retail banks still run core systems that are 30 to 40 years old, built on COBOL or similarly dated stacks that predate the engineers now maintaining them. A full core replacement is a multi-year, multi-hundred-million-dollar undertaking with a well-documented failure rate, and regulators treat it as a systemic risk event in its own right. Sutherland and FlowX.AI are explicitly selling around that reality: the agents sit at the workflow layer, call into the core through existing interfaces, and handle the manual steps a human used to do between core transactions, without anyone needing boardroom sign-off on a core migration.

This is an integration decision, not an automation decision

It is tempting to read this as "banks are adding chatbots." The actual engineering decision is narrower and harder: which 220 agents, wired into which systems of record, with what guardrails, actually survive contact with a regulated workflow where a wrong answer has compliance consequences. That is an integration and orchestration problem before it is an AI problem. Teams that treat it as "plug in an agent vendor and go live" tend to discover, mid-rollout, that the hard part was never the model, it was mapping agent outputs onto audit trails, exception handling, and the existing core's data model. A genAI integration engagement that starts with that mapping, rather than with model selection, is the difference between a pilot that survives an audit and one that gets quietly shelved after the first compliance review.

• Scope the agent layer around a handful of high-volume, well-defined workflows first — onboarding and KYC checks, not the full back office — and prove the audit trail before expanding.
• Treat every agent action against a regulated workflow as something that needs a human-reviewable log, not just a confidence score; examiners will ask for it.
• Keep the integration layer loosely coupled to the specific agent vendor. The core is staying put for years; the agent tooling market is not nearly that stable yet.
• Measure the metric the business actually cares about, like handoffs eliminated or cycle time, not "agents deployed" — FlowX.AI's own 80 percent and 40 percent figures are process metrics, not model metrics.

Nobody signed up to migrate a 40-year-old core. They signed up to stop retyping the same KYC form by hand.

The broader pattern

This deal is one data point in a trend that has been building all year: regulated industries are choosing to wrap legacy infrastructure with new intelligence layers rather than replace it, because the replacement is slower, riskier, and harder to justify to a board than a workflow-level win they can point to within a quarter. For engineering leaders at banks, insurers, and other regulated shops watching this, the question worth asking is not "should we modernize the core," it is "which manual workflow, if automated correctly today, pays for the integration work by itself." Sutherland and FlowX.AI are betting that the answer, at scale, is a very long list.