AI and Risk Management: Why Deutsche Bank Insists on a Human in the Loop

Quick Summary

Artificial intelligence is a clear enabler across KYC, sanctions screening, transaction monitoring and credit risk, says Deutsche Bank''s Tsvetanka Nankova. But banking is built on trust, regulators and clients have zero tolerance for error, and her prescription is responsible velocity with a human in the loop.

Where is AI adding most value in risk management?

Nankova sees value across several risk disciplines at once. In KYC, AI can structure the client information gathered at onboarding and during regular reviews, so the bank does a better job of assessing the risk attached to each relationship. In transaction monitoring and fraud detection, it helps review specific alerts, a capability that matters more as payments move to real time and the fraud window narrows accordingly. In credit risk, it supports faster decisions, including real-time approval of credit applications, always within the bank''s risk appetite. Speed, she notes, is what clients consistently ask for: inject as much of it as possible into everything the bank does.

Where should banks still be cautious?

The caution is about model quality, not model ambition. Nankova recounts a colleague in financial crime risk management describing two team members running effectively the same prompt and receiving different responses. One produced satisfactory results; the other did not. In a business where one mistake can take you out of business or carry serious repercussions, that variance is the whole problem. The requirements that follow are concrete:

• Models must be extremely well tested and quality assured for the exact use they are put to.
• The bank must be able to explain to regulators, A to Z, how a specific conclusion was reached.
• Deployment must respect data standards, data privacy, and ethical and moral standards.

What does responsible velocity mean in practice?

Nankova''s phrase for the balance is responsible velocity: move at the speed clients are asking for, but keep a human in the loop and never let the pace of adoption outrun the testing regime. It is a deliberately unglamorous position, and it reflects where accountability actually sits when an automated decision goes wrong, which is with the bank rather than with the model.

FF NEWS TAKE:

The same-prompt, different-answer anecdote is the most useful thing in this interview, because it is the failure mode most AI governance frameworks are least equipped to catch. Reproducibility, not accuracy, is the standard a regulator will end up testing against, and very few institutions can currently evidence it.