Scalable Capital's Agentic Investing, Explained

On 25 August, the Munich neobroker Scalable Capital did something no other European bank had done: it let customers point ChatGPT, Claude or Grok straight at their brokerage account and place real trades by typing a prompt. The feature is called Agentic Investing, and it turns ‘should I buy this’ into ‘buy this’ without the user ever opening the Scalable app.

Scalable is not a fringe outfit testing a toy. It runs more than 60 billion euros in client assets for over a million customers, mostly in Germany and Austria, with operations spreading into Italy, Spain, France and the Netherlands. When a broker that size wires general-purpose chatbots into live securities orders, it is worth understanding exactly what has been switched on, and what has not.

Table of Contents

How it works

The connection runs over the Model Context Protocol, the open standard for plugging AI assistants into outside tools and data. Crucially, Scalable is clear this is not a formal partnership with OpenAI or Anthropic. It is a standalone offering that uses the MCP integrations those assistants already support. At launch, a connected assistant can:

Analyse your existing portfolio.
Execute trades on your instruction.
Set up savings plans.
Manage watchlists and create price alerts.

Chief Product Officer Alexander Seipp gave the example of asking an assistant to find stocks that have fallen for several months running, monitor them, then prepare an order. The guardrails matter here: users must explicitly approve every trade or savings plan before it goes through, agents cannot make payments or withdraw money, the same strong customer authentication as the normal app applies, and you can switch agentic access off at any time.

Can the AI actually invest?

The marketing runs ahead of the evidence here, and the evidence is mixed. A study from Elm Wealth ran a ‘Crystal Ball Challenge’, feeding Claude, ChatGPT, Gemini and Grok historical Wall Street Journal front pages with market-moving news while hiding the actual outcomes, then asking them to trade.

On picking direction, some models were genuinely strong. Across roughly 200 sessions, Claude beat human players in 76% of them and ChatGPT in 63%. Gemini managed 43% and Grok 51%. So far, so flattering for the machines.

The problem was position sizing. The researchers split the decision into two parts: what to invest in, and how much. The models were decent at the first and poor at the second. They could recite risk-management ideas like the Kelly criterion in theory, then place simulated bets sized at 7 to 12 times a sensible level. Given the US market has moved more than 5% on 23 days and more than 9% on seven days since 2000, that kind of over-sizing is a fast route to a catastrophic loss, even when your hit rate is good.

The honest read

Scalable’s design is more careful than the headlines suggest. Keeping a human approval step on every order, blocking withdrawals, and leaning on MCP rather than a bespoke deal are all the right instincts. Seipp himself hedged, saying whether customers ‘end up finding the holy grail together with your AI assistant on high returns and low risks or not, I think that remains to be seen’.

None of this is investment advice, and it is worth being blunt about the trap. An assistant that sounds confident and gets direction right most of the time, while badly misjudging how much to stake, is arguably more dangerous than one that is obviously useless, because it earns trust it has not fully paid for. The approval button is doing a lot of work in that scenario, and it only works if the human pressing it actually reads the order rather than rubber-stamping the robot.

Access to compute and intelligence ‘literally in your pocket, 24/7’, as Seipp put it, is a real shift. Whether that makes retail investors richer or simply faster at making expensive mistakes is the open question Scalable has now handed a million people. (Not investment advice.)

Did you know: the Kelly criterion, the position-sizing formula the AIs struggled to apply, was devised in 1956 by physicist John Kelly Jr at Bell Labs, and was later used by the likes of Ed Thorp to size bets on both blackjack and the stock market.

Sources

The free stack. One email a week: the AI tools and moves that actually matter, hype filtered out. Subscribe free →