THE BANK OF THE FUTURE

Five Field Notes from Frankfurt

When people ask me what the bank of the future looks like, I resist the temptation to show a glossy demo. The answer is not a demo. It is a way of working. It is a mindset that combines purpose, craft and patience. These five notes are the spine of my talk at Kinfos in Frankfurt this week. They are not theories. They are the patterns I see when AI actually lands in teams, processes and customers’ hands.

1) Human at the steering wheel, then human in the loop

We love to say human in the loop. It sounds safe. It signals control. But the loop is a control layer. It is not the beginning. The beginning is human at the steering wheel.

The steering wheel is purpose. It is the human who asks the uncomfortable question: why does this process exist at all, and what outcome do we want for customers, for risk, for cost, for time to value? If we skip that step, we end up with a supervised version of the same old process. Faster, maybe. Cheaper, sometimes. Better, rarely.

Without people who steer, human in the loop becomes a checkbox. With them, it becomes a discipline. These are the people who select the right problems, shape the guardrails, and decide when AI should give an opinion, when it should make a decision, and when it should stay quiet.

Make it real

• Name your steerers. They are usually product owners, experienced operators, auditors with a system view, or analysts who understand both data and consequences.

• Give them permission to redraw the map. Not just to validate outputs, but to rewrite steps, SLAs and acceptance criteria.

• Tie their goals to outcomes, not to model scores. Customers do not feel a higher ROUGE or BLEU. They feel faster resolution, fewer errors, simpler journeys.

Get the steering wheel right and the loop stops being a compliance ritual. It becomes a learning engine that improves the system every day.

2) AI lifts the carpet, and it does not blink

I spent years as an internal auditor. My job was lifting carpets to see what lived underneath. AI does the same thing, only faster, deeper and with fewer blind spots. That is why it makes people nervous. It will expose our routines. It will show how many steps exist because someone once said so. It will surface decisions that travel from inbox to inbox without adding any value. It will ask whether a control prevents anything or simply documents that we tried.

The best teams welcome that discomfort. They treat AI as a mirror, then as a blank page. First the mirror: what patterns, exceptions and bottlenecks does the system find in our logs, tickets, emails and contracts? Then the blank page: if we designed this process today, with the customer in mind and the data we now have, how would it look?

Make it real

• Run a humility audit. Pick one process and write down everything you do because that is how it has always been done. Challenge every line.

• Use AI as a probe, not as a patch. Let it read the corpus, cluster cases, detect drift, and point to the 3 places where 80 percent of the pain lives.

• When you find something ugly, resist the urge to hide it with automation. Fix the design. Then automate.

AI does not blink. Let it show you the room, dust and all. Then clean with intent.

3) AI becomes a commodity. Your moat is creativity, context and determinism

The model itself will not be your edge for long. Most organizations will access models that are more than good enough. The compounding advantage will come from three things you fully control.

Creativity. The courage to reframe the problem. The discipline to imagine workflows we have never tried. Creativity is not waiting for a eureka. It is a daily practice: reframing questions, proposing two alternative flows, deleting a step, moving a decision earlier in the journey. This links directly to the blank page in point 2.

Context. LLMs are powerful pattern machines, but they do not guess your world. Feed them rich, coherent, integrated context and they return useful, specific answers. Starve them and you get generic prose. Context is not just data volume. It is curation: the right docs, the latest rules, the approved templates, the allowed actions, the forbidden ones, the preference order when trade offs appear.

Determinism. LLMs are probabilistic. Business needs repeatability. The bridge is decomposition. Break the big problem into many small units with narrow scope and clear acceptance tests. Make those units deterministic with structured prompts, schemas, tools and constraints. Then compose them into a reliable flow.

Make it real

• Build context pipes. Create living bundles for each domain: policies, glossaries, snippets, reference decisions, calculators and tools. Version them. Govern them.

• Teach decomposition. Write user stories for agents the same way you do for services. Single responsibility. Clear inputs and outputs. When a block is too fuzzy, split it again.

• Reward creative reframing. Add a field in every ticket: alternative approach attempted. Measure it.

  
  

4) A benign virus beats a central command

Large, central AI programs look impressive on slides. They struggle in the wild. AI is not a single platform you deploy. It is a set of skills and habits that spread. Think of it as a benign virus. Your job is to find the carriers and make contagion easy.

Carriers are the people who enjoy tinkering. The ones who cross the hallway to talk to risk. The ones who share a notebook and a lesson learned without being asked. Give them a small budget, a safe sandbox, and a way to publish micro wins. Make their artifacts lightweight and shareable: a prompt that saved an hour, a context pack that fixed 20 percent of errors, a tiny agent that closes a ticket end to end for a narrow case.

Central teams still matter. They set standards, provide platforms, secure the data, and keep you out of trouble. But culture travels through stories and tools that people can touch. Infection beats instruction.

Make it real

• Map champions across functions. Do not cluster them in one tower. Finance, operations, compliance, legal, branches, contact centers. Everywhere.

• Publish in small, edible bites. A one pager. A 3 minute video. A starter kit with five example prompts and a checklist.

• Celebrate cross team wins publicly. Not the model release, but the moment a branch used a micro agent from operations to solve a customer problem in 30 seconds.

When enough people share, borrow and adapt, the organization changes shape. Silos soften. The dumb idea that data equals power loses oxygen. Capability spreads.

5) Build like a collector: patient, selective, compounding

We want the big leap. Real change arrives as a collection. A collector knows two things. First, value accrues to the set, not to the single item. Second, a great collection takes time.

Treat AI the same way. Do not try to flip a switch. Accumulate. One well designed micro agent that does a narrow job flawlessly. One context bundle that becomes the source of truth for a domain. One prompt pattern that turns a vague request into a structured action plan. Curate them. Improve them. Connect them.

Prompting is not a trick you learn in a morning. It is like learning a language. You need hours, feedback and exposure to better examples. Context does not clean itself. You need custodians who prune, version and deprecate. Processes do not become deterministic by decree. You get there by shrinking the unit of work and writing tests.

Make it real

• Schedule practice like you would for a language. Two sessions per week. One to create. One to review and refine with peers.

• Track compounding value. Each micro asset should have a tiny metric: minutes saved, errors avoided, NPS uplift, risk flags caught. Watch the stack add up.

• Sunset aggressively. A collector removes items that no longer fit the set. You should, too. Kill prompts, agents and bundles that have been superseded.

You will wake up one day and realize you did transform. Not with a single project, but with a patient accumulation of capabilities that now feel obvious.

How the five connect

Put a human at the steering wheel and you get purpose. Use AI to lift the carpet and you get the truth. Accept that the model is a commodity and you shift effort to creativity, context and determinism. Spread capability like a benign virus and your culture starts to move. Build like a collector and the change sticks.

Purpose gives direction to the mirror. The mirror invites creativity. Creativity needs context and deterministic blocks to become reliable systems. Those systems spread through carriers who share simple assets. The assets compound into a collection that quietly changes how you work. That collection, in turn, gives your steerers more powerful tools to redraw the map again. The loop closes. Not a control loop. A learning loop.