Skip to content
Beyond Prompt AI Studio

Governance & guardrails

Scalable Capital opens brokerage accounts to AI agents - the widely cited '76% beats humans' figure comes from a study with a very different core message

September 2, 2026 · 10 min read · Beyond Prompt AI Studio

AgentsFintechRisk managementMCP

On 25 August 2026, German neobroker Scalable Capital launched 'Agentic Investing', becoming, per its own claim, the first bank in Europe to open its platform to mainstream AI assistants. Customers can connect their brokerage account via the Model Context Protocol (MCP) to OpenAI's ChatGPT, Anthropic's Claude, or Grok, and control securities orders, savings plans, watchlists, and price alerts by prompt. Coverage accompanying the launch cites a striking figure: Claude beats human traders in 76 percent of cases. This analysis traces that figure to its source - and finds a markedly more cautious message in the underlying study than the headline suggests.

Key points at a glance

  • Scalable Capital launched 'Agentic Investing' on 25 August 2026: customers connect their brokerage account via MCP to ChatGPT, Claude, or Grok and control orders, savings plans, watchlists, and price alerts by prompt.
  • Per the provider, safeguards include: trades and savings plans must be confirmed before execution, the AI can't withdraw money independently, and existing account and role permissions plus two-factor authentication remain in force.
  • The figure cited in coverage - 'Claude beats human traders 76 percent of the time' - doesn't come from Scalable Capital, but from an independent study by Elm Wealth from June 2026 (the 'Crystal Ball Challenge').
  • That study tested only how well different AI models could predict pure market direction from historical Wall Street Journal front pages, across roughly 200 rounds. Claude scored 76 percent, ChatGPT 63 percent, Grok 51 percent, Gemini 43 percent. No actual trading outcomes or risk-adjusted returns were measured.
  • The study's actual core finding is a warning: on the question of how much to invest, the researchers found the AI models systematically took on too much risk - with average position sizes 7 to 12 times what the researchers considered reasonable, alongside an explicit warning about 'catastrophic loss of capital'.
  • That risk warning is completely lost in the isolated 76-percent figure - a pattern worth checking for in any similar AI-success headline in a financial context.

What Scalable Capital launched on 25 August

Scalable Capital is one of Germany's best-known neobrokers, and its new 'Agentic Investing' feature makes brokerage accounts directly accessible to AI assistants. The technical foundation is the Model Context Protocol (MCP), through which ChatGPT, Claude, or Grok can connect to an account. At launch, the platform's core functions are available through this interface: executing securities orders, setting up savings plans, managing watchlists, and setting price alerts - controlled by prompt instead of the usual app interface.

The safeguards the provider describes are notably cautious, especially compared with other current agentic trading platforms: trades and savings plans must be confirmed before execution, the AI can't withdraw money independently, and existing account and role permissions plus two-factor authentication remain continuously in force. That differs markedly from approaches like the crypto exchange Binance's 'Agent OS', launched in August, where AI agents can trade in dedicated sub-accounts with no built-in loss cap.

The figure circulating in coverage

A striking figure has circulated around the announcement: Claude beats human traders 76 percent of the time. It reads like a direct performance validation of Scalable Capital's new feature - but it doesn't actually come from a test by the company itself. It traces back to an independent Elm Wealth study from June 2026, the 'Crystal Ball Challenge'.

In that study, different AI models were shown historical Wall Street Journal front pages containing market-moving information, without knowing the actual subsequent market outcomes. Across roughly 200 test rounds, the study measured only how often a model correctly predicted the pure direction of a market move - up or down. Claude achieved a higher hit rate than human comparison participants in 76 percent of rounds, ChatGPT in 63 percent, Grok in 51 percent, Gemini in 43 percent. Important context: no actual trading outcomes, returns, or risk-adjusted performance were measured - only pure directional prediction.

The study's actual core finding: a risk warning

Elm Wealth's researchers explicitly distinguished between two separate decisions in their study: where to invest, and how much to invest. On the first question - pure directional prediction - the AI models performed comparatively well; that's the source of the 76-percent figure. On the second question, decisive for actual capital preservation, the researchers found a markedly more troubling pattern: the AI models systematically took on too much risk when it came to the actual position-sizing decision.

The researchers stated, verbatim, that given average position sizes 7 to 12 times what they considered reasonable, the AI models were taking 'too much risk of a catastrophic loss of capital'. That's the study's actual core finding - a warning about oversized risk-taking in execution, not evidence of superior AI investment judgment. In the isolated 76-percent figure as it circulates around the Scalable Capital announcement, that warning is completely lost.

Why this distinction matters for investment decisions

This separation between directional prediction and position-sizing discipline isn't an academic nicety - it's the core of any responsible investment decision. Correctly judging whether a market will rise or fall is worthless if a disproportionately large share of available capital is staked on that single judgment - a single wrong call at an oversized position can wipe out the accumulated benefit of many correct predictions. That's exactly the mechanism the Elm Wealth researchers warn about.

For Scalable Capital's specific implementation, that partially tempers the risk assessment: because trades must be confirmed before execution, a human user still has the opportunity to spot and reject an oversized position size the AI proposes before it's actually executed. That confirmation requirement is therefore not a mere formality, but a concrete structural answer to exactly the risk the underlying study describes - provided users actually scrutinize the proposed position size, rather than routinely confirming suggestions.

What this means in practice

  • For any AI-generated investment recommendation, explicitly separate the directional or selection judgment from the proposed position size - the latter in particular deserves its own critical review, regardless of how convincing the reasoning behind the direction sounds.
  • With agentic investing features like Scalable Capital's, actively use the pre-execution confirmation step as a real control point rather than routinely waving suggestions through - the study shows this is exactly where the biggest risk sits.
  • Before drawing a competence claim about real investment decisions from a circulating AI-success figure in a financial context, check what was actually measured: a hit rate on pure directional prediction is something different from a proven investment strategy with solid risk management.
  • When introducing agentic investing features at your own company (for instance for treasury management), compare the concrete safeguards of the specific provider - the range spans from systems requiring mandatory confirmation to models with no built-in loss cap at all.

The real value of this analysis isn't a verdict on whether Scalable Capital's Agentic Investing makes sense - that depends on individual use. It's tracing a widely circulated success figure back to its actual source: reading the underlying study, rather than just the circulating headline, surfaces primarily a warning about AI risk-taking behavior, not evidence of superior AI investment competence.

Frequently asked questions about Scalable Capital's Agentic Investing and the 76-percent figure

Does the 76-percent figure come from a test by Scalable Capital?

No. The figure comes from an independent Elm Wealth study from June 2026 (the 'Crystal Ball Challenge') that has nothing to do with Scalable Capital's product. It also only measures the hit rate on pure market-direction prediction, not actual trading outcomes.

Does this study actually show Claude is a better investor than humans?

That can't be concluded from the study. Claude performed comparatively well on pure directional prediction, but on the position-sizing decision - how much to actually invest - the researchers found a markedly more troubling pattern: systematically excessive risk, with an explicit warning about catastrophic loss of capital.

How does Scalable Capital protect customers from oversized AI suggestions?

Per the provider, all trades and savings plans must be confirmed before execution, the AI can't withdraw money independently, and existing account and role permissions plus two-factor authentication remain in force. That confirmation requirement is the concrete answer to the oversized-position risk the Elm Wealth study describes - but it's only effective if users actually scrutinize the suggestions critically.

Should we generally avoid using AI agents for financial decisions?

This analysis doesn't argue against using AI in financial decisions generally, but for a more differentiated assessment: the ability to judge direction or selection needs to be separated from discipline in position-sizing and risk control. Systems with mandatory human confirmation before execution provide a structural control point for exactly that.

Want your company's use of agentic investing or AI financial features reviewed for risk management?