Agentic AI in Banking: The Inbox Was Only the Warm-Up
CCO Managing Partner
The first wave of AI in banks lived in the inbox. A better email. A drafted note. A summary of a meeting nobody wanted to write up. That wave was easy to like and easy to ignore. Nothing fundamental in the control framework had to move. Nobody had to change who signs.
The second wave is different. It is moving into onboarding, Source of Wealth, KYC, AML and periodic reviews. These processes combine high volumes, heavy documentation and, at the end, human judgment. That combination is exactly why they are likely to be industrialised early. It is also why they will go wrong quickly if the bank only buys speed.
Deutsche Bank Private Bank’s Source of Wealth agent in Singapore and Hong Kong is a good example of where this is heading. Research, assembly, consistency and documentation can move. A person still reviews the outcome and remains accountable. The bank has made the important distinction explicit: tasks can be automated, accountability stays with people.
That sentence should sit above every Agentic AI project in a regulated institution.
Why These Files, Not the Newsletter
A quarterly letter can be wrong and the client may still stay. A Source of Wealth file that does not add up can become a problem for years. The same is true for KYC that asks for the same passport three times, or an AML narrative that looks complete but is not.
In my banking years, those files were never one document. A company sold in one country. Dividends booked in another. A trust mentioned late. Three versions of the same life sitting in three systems. The relationship manager collected evidence for a week. Compliance asked for parts of it again the week after. The client only felt the delay.
That is not a chatbot use case. It is a workflow use case.
An agent that can pull approved public records, match them to what is already in the file, draft one narrative and flag where dates or amounts do not line up is doing useful work. It turns the banker from a document hunter back into the person who still has to look the client in the eye.
What Speed Does Not Buy
Speed without one owned narrative simply industrialises a weak file.
I would not judge a pilot on the salaried profile. That file will look good in the first demo. I would take the hard one: several entities, several jurisdictions, wealth that did not arrive neatly as a payslip.
Then I would ask four questions:
- Did the agent produce one consistent story, or three fluent ones?
- Did it surface the contradiction a human might have missed, or bury it in polished language?
- Can the banker explain the conclusion without the prompt history?
- Would the client recognise their own story in the file?
If the answers are yes, this is no longer just an experiment. If they are not, you have a nicer pilot and the same old biography, only written faster.
The Operating Model Hidden Inside the Workflow
Regulated work has always had three layers. Banks mixed them together for years because experienced people could smooth over the gaps: facts that can be checked, a narrative that has to make sense and a judgment that cannot travel with the prompt.
Agentic AI can be very good at the first two when the sources are clear and the definitions are shared. It becomes dangerous at the third if the human signature is only formal.
Onboarding is still one of the first proofs of whether an institution is organised around the client or around its own departments. Source of Wealth is perhaps the sharpest version of that test. Get the assembly right and the relationship begins with competence. Get it wrong and the damage remains in the file.
The same logic applies to periodic reviews. If the annual update becomes a new novel every year, you never had a knowledge layer. You had a writing contest.
What I Would Tell a Swiss Incumbent
Do not start with a copilot in Outlook and call it transformation. Start with one regulated workflow where the pain is real and the owner can be named.
Put the agent on research, matching, drafting and exception flags. Keep the conclusion with a person who can be asked how they reached it. And make the narrative inherit. The story that leaves onboarding should be the story that arrives at the next review.
If three departments still keep three versions of the same client, the agent will not save you. It will scale the disagreement and give it a confident tone.
The inbox made AI look friendly. The regulated workflow will decide whether it is actually usable. Tasks can move. The file still needs an owner. Speed is only progress when the same story survives the first serious review.
If you are looking at where Agentic AI could take real work out of KYC, Source of Wealth, AML or another regulated banking process, book a 30-minute exploration call with us. We can look at one concrete workflow, identify what can safely move to an agent and where human judgment and accountability still need to remain explicit.