AI Is the Answer for the Banking Industry. But It’s Also the Problem
By Patrick Van Deven, CEO at VaultSpeed
Simple Subscribe
Subscribe Now!
Every bank with a serious data team in place already knows that pointing a general-purpose AI agent at thousands of raw tables is a terrible idea. No seasoned technology leader would try it, and yet all banks are under significant pressure to do so, leading to paralysis because boards are demanding AI-driven insights, but the path from pilot to production keeps stalling. McKinsey, Gartner (and internal post-mortems) all tell the same story: the pilots work in a sandbox but collapse the moment they hit real enterprise data at scale. The question is why.
Key insight: The answer is not the AI: it’s the missing layer between the AI and the data. Enterprise data was never designed to be read by machines that guess. It was designed to be read by people who already know what “customer” means in their department, which source system wins when two tables disagree and why that one transformation pipeline was built the way it was in 2011. An AI agent has none of that context. Without it, every query is an expensive, unreliable, unrepeatable experiment.
The real question: Can you solve the AI-readiness problem with AI itself? Can you use agents to build the structured, machine-readable context that other agents need to operate, and do it fast enough and cheaply enough to justify the investment?
The answer is yes – but only with the right architecture underneath.
The Chat Experience Trap and the Cost of Freedom
Popular AI tools, which let you ask anything at any time, create unrealistic expectations for how AI should work in business environments. Unstructured AI queries against raw data become extremely expensive at scale. That’s because the AI has to rediscover the context with each request, with no memory of past decisions, no understanding of business definitions and no awareness of conflicts between source systems.
In banking, context is everything. The term “revenue” can mean very different things: an accounting definition for the CFO, a commission income figure for the retail division, a net interest margin for treasury or a recurring advised assets value for wealth management. Without a governed layer, each departmental AI agent may give a different answer, turning concepts like EBITDA into matters of opinion and making margin reconciliation a drawn-out exercise. The blame often falls on the data team, but the real issue is the lack of a clear, structured conceptual choice.
Case in point: One bank ran the experiment both ways. First, they pointed a generic AI agent at their raw data. The result: 7% accuracy. Bad enough on its own, but the dangerous number was the second one: 21% confidence. The AI was certain about answers that were deeply wrong. That’s the real risk: Not that AI fails, but that it fails while looking like it succeeded.
With structured context layered in, using the same data and the same AI models, accuracy jumped to 90%. That’s not a “deploy-and-walk-away” level of precision, to be sure. Outputs still require human review. The bank’s view was straightforward: 90% plus review isn’t AI perfection, but it is a tenth of the time analysts used to spend building the same output from scratch. Going from zero automation to 90% in four weeks changed the economics of the entire program.
Treat AI Agents as Employees, Not Oracles
AI agents should be treated more like new employees. Rather than handing them a complete file system and expecting them to “figure it out,” you’d give them structured onboarding, explain how things work and define what matters. AI agents need the same support.
In regulated industries like banking, this is fundamental. Banks need deterministic, auditable processes. When regulators ask how a number was calculated, the answer “AI figured it out” is not acceptable. There must be full traceability from the business definition to the source system, reproducible at all times.
This is where the approach turns recursive: AI builds the structured context that other AI agents need to operate reliably. One set of agents analyzes source systems, extracts business definitions, and maps relationships. That knowledge is captured in a governed, machine-readable layer. From there, deterministic code generation takes over for the data foundation itself, ensuring the same input always produces the same output. The AI does the heavy analytical and design work; automation guarantees code correctness. Humans review, approve and stay in control. Rather than AI replacing the data team, it gives them a 10x multiplier. And every downstream agent gets the context it needs to stop guessing.
The No-Regrets Move: Fixing the Foundation Before Monday
Just about financial institution faces a familiar challenge: decades of transformation logic encoded in pipelines, maintained by a small group of people with tacit knowledge and documented in fragmented (or sometimes nonexistent) ways. That may have been tolerable when data engineering was a back-office function, but it becomes unsustainable when the board expects AI-driven insights from that same data.
The good news is that this does not require a complete replacement or a five-year program. The structured data foundation already in place, or now being built, is exactly the context AI needs to work well. Investing now in governing and structuring data will make every AI initiative cheaper, faster and more accurate. Skipping that step means building on shaky ground.
This investment also acts as a hedge. If AI models change, token prices fluctuate or regulations tighten, a well-governed data foundation can adapt. It is independent of any one model, platform or vendor, which makes it a true no-regrets move.
Adopting this approach is critical. The cost of unleashing AI on unstructured banking data, the false promise of a consumer-style chat experience in the enterprise and the need to treat AI agents like employees who require structured context all point to the same conclusion. Banks must address the “dragon in the basement” of legacy transformation debt and recognize regulatory pressure as an accelerator for getting AI right. Ultimately, properly structured data is the no-regrets move that will allow AI investments to truly pay off.
