Every artificial intelligence (AI) tool a bank adds to solve one risk problem gives it another vendor to vet, another platform to monitor and another set of decisions to explain to regulators. That\u2019s the paradox at the center of the next phase of AI in financial services. The technology meant to simplify risk work can make the institution harder to govern as a whole.<\/p>\n
Financial institutions have traditionally applied AI one task at a time, using separate tools to screen sanctions lists, detect suspicious transactions, review identity documents and flag fraud. Those applications improved individual tasks without fundamentally changing how work moved through the institution. Now, banks are being pitched a new generation of AI agents, often by a new generation of vendors, for every part of that work.<\/p>\n
\u201cWhat we\u2019ve heard from banks is they\u2019re getting inundated with people saying, \u2018I have this new AI widget,\u2019\u201d Bretton AI<\/a> CEO Will Lawrence<\/a> told PYMNTS CEO Karen Webster<\/a>. \u201cWe think a larger risk is onboarding dozens and dozens of new vendors and new platforms.\u201d<\/p>\n The bigger risk, in other words, isn\u2019t that banks will fail to find enough AI applications. It\u2019s that they\u2019ll create a fragmented AI estate before deciding how to govern it. Rather than selling banks dozens of specialized tools, Lawrence sees greater value in a common platform capable of managing multiple risk-based workflows while simplifying third-party oversight.<\/p>\n From Point Solutions to a Common Risk Platform<\/p>\n Bretton AI, formerly Greenlite AI, built its business automating financial crime investigations before expanding into lending, fraud, onboarding and broader operational workflows after raising $75 million. \u201cThe mandate grew,\u201d Lawrence said, explaining both the company\u2019s rebrand and its expanded strategy. Customers wanted the same AI capabilities applied beyond anti-money laundering and know-your-customer reviews to other areas where institutions must evaluate operational risk.<\/p>\n Lawrence argued that financial crime was the ideal proving ground because it demands high-volume decision-making inside one of banking\u2019s most heavily regulated environments. The Bretton AI platform has now supported 1.2 million investigations, experience the company believes transfers naturally into underwriting, fraud operations and complex customer onboarding.<\/p>\n \u201cWhat we found is agents are really good at doing those long and complex investigations,\u201d Lawrence said, pointing to institutional onboarding processes that can require analysts to spend an entire day reviewing ownership structures, documentation and jurisdictional records. \u201cWe\u2019re increasingly seeing people use us for use cases we never really imagined.\u201d<\/p>\n That expansion is what Lawrence said changes Bretton\u2019s story from an AI tool story into a platform story. Instead of answering one question or automating one task, agents can collect documents, conduct research, analyze risk and prepare recommendations across a workflow that previously required multiple employees and software systems. A common platform can carry the work from one step to the next under the same governance framework, rather than requiring the bank to add a new vendor and a new handoff for every task.<\/p>\n Why the Platform Matters<\/p>\n The third-party oversight argument is easy to overlook, but it may be the most practical reason banks resist accumulating AI point solutions. Each new vendor brings due diligence, integration work, data permissions, performance monitoring and an audit trail the bank must defend. Consolidating multiple workflows on a common platform doesn\u2019t remove that responsibility, but it can reduce the number of separate relationships and control systems the institution must manage.<\/p>\n It also raises the stakes for a platform like Bretton AI. Unlike a chatbot answering a single question, an agent can execute dozens of sequential tasks over several hours. Every intermediate judgment creates another opportunity for error, and one wrong conclusion can shape everything that follows. \u201cIf you screw up one part of the journey,\u201d Lawrence said, \u201cthe rest of it is going to be off.\u201d<\/p>\n Webster framed the issue directly: \u201cCan I really trust an agent to handle this from beginning to end, come up with a decision and execute and implement?\u201d The competitive question therefore becomes less about how much AI a bank can deploy and more about whether it can govern a platform performing increasingly consequential work across the institution.<\/p>\n Autonomy for the Work, Not the Decision<\/p>\n Lawrence separates autonomy into two categories: allowing AI to perform the work required to reach a conclusion and allowing AI to make the final decision itself. That distinction gives banks a practical way to capture the value of a common platform without handing it unlimited authority. \u201cWe\u2019re really big fans of using autonomy in the getting-to-an-answer phase,\u201d he said. \u201cBut rarely do you want an agent making a decision.\u201d<\/p>\n The greatest productivity gain may come from shrinking an eight-hour investigation into a one-hour review, with an experienced employee validating the agent\u2019s work instead of producing everything from scratch. \u201cYou don\u2019t want to go eight hours down to zero minutes,\u201d Lawrence said. \u201cSeven hours is great savings.\u201d The human remains accountable for the consequential decision, while the agent absorbs the collection, research and analysis that consumed most of the day.<\/p>\n That\u2019s more precise than simply declaring that a human will remain in the loop. The important questions are where the handoff occurs, what evidence the human receives, whether the agent\u2019s reasoning can be challenged and who is responsible when an earlier step sends the workflow in the wrong direction. A platform spanning multiple workflows must make those boundaries clearer, not blur them.<\/p>\n Governance Becomes the Product<\/p>\n As banks adopt more capable agents, governance increasingly becomes the limiting factor and, potentially, the platform\u2019s most valuable feature. Lawrence argues that institutions already possess much of the necessary framework through existing model risk management, audit controls and regulatory oversight. The challenge is adapting those disciplines to systems that execute multistep workflows continuously instead of producing isolated outputs.<\/p>\n \u201cThe main question we get from banks is how do I know it\u2019s safe to roll out AI?\u201d Lawrence said. Webster reduced the hesitation to its essence: \u201cThe stakes are high when it is wrong.\u201d That is especially true when one platform touches more workflows and an error can travel across them.<\/p>\n The common-platform argument makes governance concerns more important. If banks are going to put onboarding, lending, fraud, compliance and other risk work on shared infrastructure, they must be able to monitor performance across the entire journey, preserve an audit trail for every intermediate judgment and define exactly where human accountability begins.<\/p>\n The next phase of bank AI may therefore be won not by the institution that accumulates the most specialized tools, but by the one that can run more risk workflows on fewer, better-governed platforms.<\/p>\n Bretton has raised $75 million to make that case. It now has the opportunity to show that simplifying the vendor stack can also simplify the work, without simplifying away the human judgment that regulators, institutions and customers still need.<\/p>\n