AI Agents in Banking: Does Your Bank Know What “Revenue” Means?
A senior manager at a Swiss private bank said to me this week: “Honestly, I missed what is going on with AI. Everybody has an opinion, but what can we really do?”
That question stuck with me. It is the most honest one I have heard in months.
AI Agents Are Exposing an Awkward Data Problem
We spent roughly 15 years moving away from centralised semantic models. The goal was speed. Analysts should connect data, create calculations and build dashboards without waiting for an architecture board.
Self-service won.
Freedom won.
Then AI agents arrived. They do not politely ignore inconsistent definitions the way a human analyst sometimes does.
They ask the awkward question out loud:
- What does Revenue actually mean in this bank?
- Which customer table is authoritative?
- Is this the fiscal calendar or the calendar calendar?
- Do we exclude internal transactions or not?
Suddenly the old idea of shared, reusable business meaning started looking smarter than many of us admitted.
AI Is No Longer the Bottleneck. Meaning Is.
The technology is moving fast. Christian Sewing, CEO at Deutsche Bank, is right when he says AI is already accelerating processes and creating real client value. Platforms are solving the “one copy of data” problem. Agentic systems are beginning to support advice preparation, credit processes, portfolio monitoring and documentation.
Yet three stubborn realities remain:
- Inconsistent metric definitions still limit the ability to scale AI beyond pilots
- Shadow AI is spreading quietly across teams
- In the highest-value work of private banking and wealth management, succession planning, family governance and multi-generational wealth structuring, human judgment and trust still decide the outcome
An agent can prepare a brilliant briefing. It cannot yet replace the relationship manager who sits with a family and navigates the emotional and legal complexity of the next generation.
So What Can Banks Actually Do to Scale AI?
The answer is not “buy more models” or “launch another pilot.” It is to deliberately design the middle ground between pure central control and uncontrolled freedom.
Here is the practical operating model that is starting to work:
- Let the business own the use cases. The relationship managers, product owners and domain experts decide where AI creates real value. They remain accountable for the outcome.
- Let IT provide the governed platform, security and standards. IT does not try to own every model. It becomes the reliable platform, identity, security and monitoring layer that makes safe scaling possible.
- Put a living semantic and knowledge layer on top of the data platform. This is the missing piece most organisations still underestimate. Metrics, relationships, business rules and domain knowledge need to be explicit, versioned and reusable. Divergence should be visible rather than hidden. Inheritance should be preferred over silent copying.
- Keep human judgment accountable for the final advice. Especially in wealth management, AI should augment preparation, consistency and coverage, never replace the human who carries the client relationship and the regulatory responsibility.
When these four elements work together, governance stops being the brake and becomes the accelerator.
Teams can move faster because the guardrails are already in place. Trust increases because the definitions are clear and the lineage is visible.
Why the Semantic Layer Matters for AI in Banking
We did not eliminate the need for shared meaning. We temporarily moved it into every individual workbook, notebook and local calculation. AI agents are simply forcing us to put it back where it belongs: as reusable, governed infrastructure. This time the consumers are both humans and agents. Both expect the definitions to be consistent, versioned and explainable.
Closing the Gap Between AI Opinion and Action
That Swiss private banker’s question remains the right one: Everybody has an opinion about AI. But very few institutions have yet answered the practical question: What does the menu look like in our bank, and who is responsible for keeping it accurate and up to date?
The banks that put the menu back on the table, clearly, governed and usable, will be the ones that turn AI from an interesting experiment into a reliable source of client value and operating leverage. The rest will keep producing confident answers that no one fully trusts.
If you are asking what this could look like in your bank, book a 30-minute exploration call with us. We can look at where you are today, which AI use cases have the most potential, and what semantic and governance foundation you need to scale them with confidence.