Quiz: AI Agents in Finance

Answer: (D) A copilot has three of the four components: it can reason (LLM), use tools (APIs), and remember context (memory). But without autonomy, it does not pursue goals independently -- it assists a human who directs each action. Adding autonomy is what makes it an agent.

Answer: (A) In each ReAct cycle, the agent first thinks about what information is needed, then acts by calling a tool (API, database, search), then observes the result. This cycle repeats until the agent has enough information to produce a final answer. The interleaved approach produces auditable reasoning traces.

Answer: (B) Episodic memory records what the agent did, what happened, and what it should avoid. For example: "Last time I recommended this stock based on a hallucinated EPS figure, the trade lost money -- next time, cross-check the number." This enables learning from experience.

Answer: (C) At Level 3, the agent acts autonomously within pre-set boundaries (e.g., trade size limits, approved asset classes) but a human can intervene and override at any time. Most financial agents today operate at Level 2--3. Level 5 remains aspirational and raises unresolved regulatory questions.

Answer: (A) When an LLM hallucinates, it produces a wrong answer that a human can spot. When an agent hallucinates, it may use the false information to call a tool, pass it to a calculator, and execute a trade -- turning a wrong answer into a wrong action with real financial consequences.

Answer: (B) Compliance is the most labor-intensive function in banking, with an estimated $270 billion spent globally per year and 95%+ false positive rates in transaction monitoring. Agents can reduce false positives by reading context, draft SARs automatically, and generate regulatory reports -- cutting analyst workload by 70--80% (illustrative).

Answer: (B) The EU AI Act (effective August 2025) classifies credit scoring, insurance pricing, fraud detection, and trading agents as high-risk. These systems must implement documented risk management, maintain human oversight capabilities, provide transparency about how decisions are made, and pass conformity assessments.

Answer: (C) The 7 percentage-point drift (67% vs. 60%) exceeds the 5% threshold. Operating at Level 3 autonomy, the agent checks market conditions (VIX at 18, no macro event), verifies risk limits and tax rules, then executes the rebalancing trade via broker API -- all without asking the client.

Answer: (D) The agent reads context that rule-based systems ignore: Has this client made similar transactions before? Is there relevant news about the client's business? Were previous alerts on this account dismissed? By combining pattern matching with contextual understanding, agents can significantly reduce false positives while maintaining detection sensitivity.

Answer: (C) The lecture presents a multi-agent earnings pipeline where a dedicated Review Agent checks for errors, hallucinations, and internal consistency. A simple cross-check ("Does this EPS match the figure in Bloomberg?") would have caught the $2.30 vs. $1.80 discrepancy before the trading agent acted.

Answer: (D) Savings = 1 - ($377 / $3,000) = 1 - 0.126 = 87.4%. The agent handles data gathering and drafting; humans spend 30 minutes per report on quality review. The cost reduction is compelling, but only if the agent's output quality is high enough to keep review time short.

Answer: (A) Multi-agent systems decompose complex financial workflows just like a trading desk: one specialist gathers data, another checks risk exposure, another verifies regulatory constraints, and the last executes trades. The Orchestrator acts as the desk head, coordinating the workflow -- but running in seconds, not hours.

Answer: (C) Markets are stable when participants have diverse views. If most agents share the same model, training data, and reasoning patterns, they will reach similar conclusions simultaneously -- creating a correlated sell-off (or buy-in) that amplifies market moves rather than dampening them. This is AI herding.

Answer: (B) The EU AI Act assigns primary responsibility to the deployer -- the institution that integrates the AI system into its operations. The rationale: the bank chose to deploy the agent, set its parameters, and should have implemented adequate risk management and human oversight. An agent has no legal personhood and cannot be sued.

Answer: (A) Generation 3 agents expand the information set from numbers to text + numbers + context. They can read an earnings call transcript, assess management tone, check macro indicators, and then decide whether to trade -- and explain their reasoning in English. Gen 2 systems respond only to numerical features.

Answer: (D) A traditional chatbot pattern-matches user input against a FAQ database. Klarna's agent has tools to query accounts, pull transaction data, and take actions (freeze cards, initiate refunds). It has memory of the client's history and autonomy to resolve issues without escalation -- unless the case exceeds its confidence threshold.

Answer: (A) The lecture identifies three blockers for Level 5: (1) liability -- no legal framework assigns responsibility for autonomous agent errors at this scale, (2) regulation -- the EU AI Act requires human oversight for high-risk financial systems, and (3) systemic risk -- a fully autonomous agent managing $500M could contribute to correlated market behavior during crises.

Answer: (D) Audit trails record what happened but do not prevent damage. Kill switches stop damage but only after detection. Risk limits cap damage but do not prevent rule violations. Approval gates prevent unauthorized actions but cannot handle all edge cases. Together, they form defense in depth -- removing any one layer creates a gap that another layer covers.

Answer: (C) This is a high-stakes judgment call that involves legal obligations (SAR filing requirements), ethical considerations (investor protection vs. client relationship), and market impact. The agent should flag the evidence and escalate -- it lacks the ethical judgment, legal training, and accountability to make this decision alone.

Answer: (B) The technology works at demo scale, but deployment requires answering: "Who pays when it fails?" (liability), "Does the regulator permit this?" (compliance), and "Has it proven reliable over time?" (trust). These institutional questions, not computing power or model quality, determine the pace of adoption.