Invest1 distinct publisher3 min readPublished
A Munich broker has connected mainstream AI assistants to a regulated execution rail. The one independent test cited alongside it says the models pick well and size badly.
The Investor · Invest desk
Compiled by The InvestorSomething wrong?How this is made
Alexander Siepp's own description of the design is the part to read closely: the client journey starts inside an AI assistant and finishes inside Scalable's regulated banking infrastructure [4]. That splits a trade across two custodians of responsibility. The analysis, the screening and the wording of the instruction happen in a chat product the bank does not run. The order lands on a rail that European supervision actually reaches. The seam between them is one tap.
Scalable's guardrails are real as far as they go: the customer approves each trade and savings plan, assistants cannot make payments or pull money out, and the connection uses the same authentication as the existing apps [6]. But approval is a weak instrument when the thing being approved is already written. Siepp's example has the assistant find stocks that have fallen for consecutive months, watch them, then prepare the order from the user's instructions [13]. What the human contributes at that point is the absence of an objection, not an independent calculation of size.
Size is exactly where the Elm Wealth work says the models fail. Bell, Haghani and White found the systems relatively good at deciding what to invest in and poor at deciding how much, comfortable with the Kelly criterion and the Merton share as concepts and unable to apply them under simulated pressure [9]. Their measured average position sizing was 7x to 12x, against a US market that has moved more than 9 percent on seven days since 2000 [10]. Run that arithmetic: a 9 percent adverse move at 7x exposure costs 63 percent of capital, and at 12x it costs 108 percent [1]. The strong headline number and the dangerous one come from the same study.
Scale matters for who absorbs that. More than 60 billion euros of client assets spread across more than a million customers, mostly in Germany and Austria [7], implies an average account of roughly 60,000 euros [2]. These are not desks with independent risk limits reviewing the ticket. Siepp concedes the rollout will not reach all client segments at the same speed [14].
Two things Fortune's account leaves open. Scalable said in July it would offer more than 1.8 million derivatives from seven issuers [12], and the piece does not establish whether those instruments are reachable through the assistant connection, which is the difference between a prompt that buys an index fund and one that buys leverage. And OpenAI, Anthropic, Google and xAI did not respond to questions about their assistants being used to trade [11], which leaves the bank as the only party on the record for the workflow. Siepp calls the result a level playing field, with information, compute and intelligence available in your pocket around the clock [5]. What is being distributed at that scale is a capability the only cited test rates strong on selection and unreliable on the sizing decision that determines whether an account survives a bad week.
Ranked by verification strength, evidence, and original report placement.
Scalable Capital is opening its investment platform to AI assistants including ChatGPT, Claude and Grok, giving European investors the ability to analyze portfolios, set up savings plans and place trades through prompts.
The new service is called Agentic Investing and allows customers to connect their Scalable accounts to supported AI agents through the Model Context Protocol (MCP).
Scalable Capital Chief Product Officer Alexander Siepp told Fortune the company sees the integration as a way for investors to begin their financial client journey inside an AI assistant and complete it through Scalable's regulated banking infrastructure.
Siepp said the integration "certainly creates a level-playing field" and that access to information, compute and intelligence is now available literally in your pocket, 24/7.
Scalable said users must approve trades and savings plans before they are executed, its current system does not allow AI assistants to make payments or withdraw money from Scalable accounts, and the AI connection follows the same core security protocols as existing applications, including strong customer authentication.
Scalable Capital, founded in 2014, told Reuters it has more than 60 billion euros in client assets and more than 1 million customers, primarily in Germany and Austria, with operations also expanding across Italy, Spain, France and the Netherlands.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Launch is documented, effectiveness is thinly evidenced
The product's existence, mechanism (MCP), assistant list and stated guardrails are on the record from the company via a named executive, and the platform's scale is company-disclosed. Against that, everything about efficacy rests on one secondhand summary of an external study with no methodology detail, the primacy claim is unverified, and there is only one publisher in the cluster with no independent technical or regulatory review.
Shipped into a large installed base, zero disclosed usage
This is a real deployment on a platform with over 1 million customers and over €60 billion in assets, which raises the ceiling. But no figure describes use of the agent channel itself — no connected accounts, prompts, or orders — and the company guides to uneven segment-by-segment uptake. Adoption is therefore availability, not demonstrated usage.
Headline outruns the product and the study
The framing that a bank 'let Claude and ChatGPT trade for customers' overstates a channel where the human must approve every trade and where agents cannot move money, and the 76% figure is the best of four models in a WSJ-front-page game — Grok, which Scalable also supports, scored 51%. The same study's core finding is that the models size positions dangerously, and no model provider would comment on trading use. The underlying protocol story is, if anything, under-told relative to the trading-superiority framing.
Promotional sourcing on both sides of the story
The product narrative comes from the vendor's own chief product officer, with company-supplied scale figures and an unverified first-in-Europe claim; the counter-evidence comes from Elm Wealth, an asset manager whose research argues for disciplined position sizing, which is also its commercial stance. The four model providers named declined to comment, so no independent institutional voice balances either side.
Facts are clear, significance is not
Confidence in the basic facts is reasonable — a named executive on the record, specific mechanism, specific guardrails. Confidence in what it means is low: one publisher, no regulatory or security analysis, no adoption data, a benchmark whose leveraged-game setup may not transfer to a retail order ticket, and silence from the model providers whose policies could change the picture.
product
A school agenda shipped with "Vitoiis" and a planet named Marc, and no one read it first1 distinct publisher
product
Incogni ranks 13 AI assistants by privacy risk: bigger is worse, except ChatGPT1 distinct publisher
build
MCP's roadmap fast-tracks five priorities and quietly queues everything else1 distinct publisher
product
Pew finds Americans trust chatbots with their lab results, not their loneliness1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.