Half of Swiss Banks Have No AI Strategy. The Other Half Have a Project List.
23
Overview
Switzerland's regulator says the question is no longer whether banks use AI, but whether they control it. A real AI strategy is a map: where AI decides, where people stay in the loop, and which projects to stop.
Year
2026
Industry
Financial Services
Challenge
Switzerland's financial regulator has stopped asking whether banks use AI. On 3 November, FINMA President Marlene Amstad told the Swiss Bankers Association's Digital Finance Day that the question now is whether banks have it under control. FINMA's own survey shows why she's asking. Across roughly 400 institutions, the average firm runs about five AI applications and is building nine more. Only about half have an explicit AI strategy. Most rely on outside providers, and FINMA is clear that the institution stays responsible for every tool, bought or built. So half the market has no strategy. I'm more worried about the other half. In most banks I've seen, the "AI strategy" is a project list: use cases ranked by excitement, a budget line, a vendor shortlist. A list tells you what you're building. It doesn't tell you who's accountable when a tool gets it wrong, and it never tells you what to stop. That matters because the value isn't spread evenly. McKinsey's latest banking research, published 30 October, found that only 37% of organizations report any earnings impact from AI. It also found that a few opportunities hold 70 to 80% of the AI value in a typical bank. If nine projects are in the pipeline, most of them won't pay. A strategy is a map, not a list The method I use started at IBM, where I helped develop component business modelling. Twenty years ago it decided which parts of a business became shared services. I've rebuilt it for AI. Now it decides which parts of a business AI should run, and how far. You break one business line into 20 to 40 components: client onboarding, document intake, suitability checks, corporate actions, reporting. Each has a clear boundary, defined inputs and defined outputs. That's the shape of an AI agent, which is why components work better here than process maps. Then you place every component on one of three levels. The level sets how much the AI decides.
Impact
That table answers Amstad's question one component at a time. A tool that drafts suitability letters sits at Control, so a person approves every letter. A tool that sorts incoming documents sits at Run, so a person samples the output. You don't need a new risk framework. You need to know which level each tool sits at before it goes live. The part most strategies skip The second step is the one that saves money. Score each component on value (revenue, cost or risk) and on readiness (clean data, written rules, a stable process). Then act on the result. • High value, ready: deploy AI now. • High value, not ready: fix the operating model first. • Low value, ready: only if it's cheap and reusable elsewhere. • Low value, not ready: no AI spend. Standardize, outsource or leave it. The most valuable cell is usually "fix first". These are the components where a bank keeps trying AI and keeps getting pilots that stall, because the rules live in people's heads or the data arrives wrong. AI on a broken process just gives the wrong answer faster. At one bank I worked with, we broke a legal document review into steps and written rules before any AI touched it. Only then did the tool read each agreement and flag gaps, with lawyers checking every flag. The rules were the hard part. The model was the easy part. The bottom row is the kill list. Most banks have projects sitting in it. Stopping them early frees budget and people for the few components where the value actually is. Five questions before you fund the next AI project 1. Which component does it serve, and is that component high value? 2. Which level is it at: Decide, Control or Run? Does the oversight match the level? 3. Are the rules written down, or do they live in someone's head? 4. Who checks the output, how often, and can they stop it? 5. If it's low value or not ready, why are we funding it now? If a project can't answer all five, it isn't part of a strategy. It's part of a list. FINMA wants banks to show they control their AI. A component map lets you show it. It also tells the COO what not to build.