Banks have been talking about AI assistants for years. Most of those conversations end the same way: a pilot that works in demo, a rollout that stalls in compliance, and a quiet sunset six months later. The pattern is so common that “AI in banking” has become shorthand for “expensive experiment.”
This piece is for the operations leader or technology executive at a mid-market financial services firm who has watched larger competitors announce AI initiatives and wondered whether the math actually works—or whether the press release is the product.
The shift worth watching: The banks succeeding with AI assistants are not buying packaged solutions or waiting for perfect data. They are building narrow, internal tools that solve one workflow at a time—and they are doing it with small teams in weeks, not enterprise programs over years.
What Changed in the Last Twelve Months
A recent case from Singular Bank, covered by OpenAI, reflects a pattern we are seeing across financial services. The bank built an internal assistant called Singularity using ChatGPT and Codex. The tool handles meeting prep, portfolio analysis, and follow-up tasks. Bankers report saving 60 to 90 minutes daily.
The numbers matter less than the approach. This was not a multi-year digital transformation. It was a focused build targeting three specific workflows where bankers were losing time to repetitive synthesis—summarizing client portfolios, preparing for meetings, and drafting follow-up communications.
The underlying shift: foundation models have crossed a threshold where internal tooling teams can build useful assistants without training custom models or waiting for vendor roadmaps. The capability gap between “what we could demo” and “what we can deploy” has narrowed significantly.
Why Most AI Banking Projects Still Fail
The failure rate for AI initiatives in financial services runs between 70 and 85 percent, depending on how you count pilots that never scale. The reasons are predictable:
- Compliance review cycles that outlast the project’s executive sponsor
- Data integration requirements that balloon scope before any value is delivered
- Vendors promising industry-specific solutions that require 18 months of customization
- Horizontal rollouts that try to serve every department before proving value in one
The projects that succeed share a different shape. They pick a single workflow—not a department, not a business unit—and build a tool that makes that workflow faster. They treat compliance as a constraint from day one, not a phase-gate at the end. And they staff the project with people who will use the tool, not just people who will build it.
The Compliance Trap
Mid-market banks often assume they need the same AI governance apparatus as a top-ten institution. They do not. What they need is a clear answer to three questions: what data touches the model, what decisions does the output inform, and who reviews the output before action. A 90-day compliance review for a meeting-prep assistant is a sign that the governance framework was built for a different class of risk.
The Integration Trap
The instinct to connect the AI assistant to every internal system—CRM, core banking, portfolio management, email—adds six to twelve months and often kills the project. The successful builds start with a narrower integration surface: one or two data sources, read-only access, human review of outputs. Expansion comes after adoption, not before.
What Mid-Market Leaders Should Assess Now
If you are evaluating whether to pursue an internal AI assistant in the next 12 months, the decision hinges on three factors:
- Workflow density: Do your client-facing staff spend more than 90 minutes daily on synthesis and preparation tasks that follow predictable patterns? If not, the time savings may not justify the build.
- Internal build capacity: Do you have two to four engineers who can ship a production internal tool in eight to twelve weeks? If your IT organization’s minimum project size is six months, this approach will not fit your operating model.
- Data readiness: Can you expose the necessary client and portfolio data through APIs without a multi-quarter data platform project? If the data lives in green-screen systems with no integration layer, you are looking at a different kind of investment.
The organizations moving fastest already had internal tooling teams building small productivity applications. AI assistants are an extension of that capability, not a new capability entirely.
The Cost Equation
A focused internal assistant—one that handles meeting prep and portfolio synthesis for a team of 50 to 100 bankers—typically costs $150,000 to $400,000 to build and deploy, including API costs for the first year. That range assumes you have the internal engineering capacity; if you are hiring contractors or a consulting firm, double it.
The ROI math depends entirely on what you value an hour of banker time at. At 60 minutes saved per day across 75 bankers, you are recovering roughly 18,000 hours annually. If you value that time at $100 per hour—a conservative figure for client-facing financial services staff—the annual value is $1.8 million against a first-year cost of $300,000 to $600,000.
The counterargument: bankers rarely convert saved time directly into revenue-generating activity. Some of that recovered time goes to email, some to internal meetings, some to nothing at all. The more honest framing is that you are buying capacity and reducing friction, not directly buying revenue. The projects that justify themselves are the ones where that capacity translates into measurable output—more client meetings, faster response times, higher-quality preparation.
The window for experimentation is open, but it will not stay open indefinitely. Within two to three years, the vendors will have packaged solutions that are good enough for most use cases, and the competitive advantage will shift from building these tools to adopting them. The banks building now are not just solving a workflow problem—they are developing the internal muscle to evaluate, deploy, and iterate on AI tooling before that capability becomes table stakes.
The question is not whether AI assistants will work in banking. The question is whether your organization can build one in weeks instead of years, and whether you will learn by doing or by watching.