The order of operations is the argument. Firms in this category normally ship the assistant inside their own product and decide later, if ever, whether an outside model gets to touch the pipes [4]. Reversing that only makes sense if you have concluded the icon on the home screen is a depreciating asset, and that a decade of spend defending it stops compounding the moment customers start financial tasks somewhere else [13]. The commercial version of the same thought: the broker plugged into the chatbot captures the flow, and the one still building its own does not [14].
Run the arithmetic and the hurry is easier to read. More than a million clients against over €60bn under management works out at roughly €60,000 an account [3][16]. This is not a sandbox cohort of test balances being pointed at a model vendor.
What actually crossed a line here is execution. Read-only portfolio access is a convenience; an interface that can place an order is a different class of thing, and the gap between them is measured in what happens when the model misunderstands an instruction, as TNW puts it [7]. Prompt injection is the part with no clean answer, because an assistant that reads a portfolio also reads whatever text it is pointed at, and one that can trade is a more rewarding target than one that can only summarise [8]. Scalable says security measures protect customer accounts and has not described them; how a trade is confirmed, what limits apply, and who pays when an assistant acts on an ambiguous request are all unpublished [9].
The regulatory seam is thinner than the security one. A German broker sits under MiFID II suitability and appropriateness obligations, and those rules were not drafted for a world where the interface talking to a retail client belongs to OpenAI or Anthropic rather than to the firm holding the money [11]. TNW reads Scalable's position as execution rather than advice, which is defensible and also narrow: when a model summarises a portfolio and the user acts on the summary, the line between information and recommendation carries a lot of weight quietly [12]. Every comparable European effort so far kept the model inside the institution's own perimeter, bunq's assistant included, while Citi's Jane Fraser has been arguing that two AI races decide banking's future [10].
Which leaves the justification. Podzuweit's stated hypothesis is that AI use could improve average returns, and he says that has yet to be demonstrated [6]. He also called the launch a first step and allowed that plenty of people are not ready to let ChatGPT manage their portfolio [5]. So the case for going outside first is not performance and is not customer demand. It is channel position, taken early, with the control questions still open.
