Do LLM Trading Agents Escalate When They Should? Capacity-Based Suitability Failures in Financial Execution

LLM agents can now place trades in live brokerage accounts. Under MiFID II Article 25(2), a suitable order must match both the client’s risk tolerance and their capacity to bear risk. We test whether GPT-4o and GPT-4o-mini escalate to a human when a client’s risk tolerance is labelled “aggressive” but their capacity (horizon, savings, debt, experience, profile age) makes the order unsuitable. Using a sandboxed MCP brokerage with five clients and six products, we run 60 order tasks, each under a neutral and a pressure prompt (240 runs). Gold labels come from a deterministic five-rule suitability check. Both models place most suitable orders (92–100%) but escalate only 47% and 33% of unsuitable ones, and fewer than one in five unsuitable tasks is escalated under both prompts. Escalation is weakest when unsuitability lies in the balance sheet (debt, thin savings: 33% and 12%) or the portfolio (concentration: 10% and 0%). Pressure prompts cut GPT-4o-mini’s escalation from 70% to 23% (p < 10⁻⁴); the effect on GPT-4o is smaller and not significant. GPT-4o also ends 12.5% of runs by handing the decision back to the operator instead of escalating, and in four of them claims an escalation it never performs. These results argue for deterministic suitability checks outside the model.