Now AI Can Pass the CFA Exam. Can it Replace Investment Analysis?

• AI can pass CFA exams and automate parts of the investment research process, from earnings analysis to DCF modeling, but technical knowledge does not necessarily translate into investment judgment.
• AI investment agents still struggle with coherence, reliability, and attention. They can process financial information quickly, but may miss inconsistent assumptions, industry-specific signals, and the management claims experienced analysts know to question.
• The investment analyst’s role may shift from execution to judgment and evaluation. As AI handles more repeatable research tasks, human value increasingly lies in knowing where to look, what to doubt, and whether AI-generated analysis is good enough to inform an investment decision.

Twenty years ago, I was an IT engineer at an asset manager, and I started to study for the CFA exam. The CFA curriculum is not static. Through its practice analysis, CFA Institute continuously surveys practitioners about what they actually do and updates the Candidate Body of Knowledge (CBOK) accordingly. The curriculum is our profession's continuous attempt to write down what an investment professional needs to know.

I passed all three levels. Did that make me a good analyst? Of course not. The curriculum gave me a foundation. Judgment came afterwards, from facing the market, making painful mistakes, and working with great investors.

I raise this because the same question now applies to machines. Patel et al. (2025) report that current large language models (LLMs) pass all three levels of CFA mock exams [1]. Whatever we have managed to write down, AI has now mastered. The critical part is what we have not written down.

AI is moving beyond answering questions to carrying out multi-step tasks autonomously. Advances in training and inference-time reasoning, combined with access to tools such as search, databases, and code execution, allow AI agents to plan their work, gather information, analyze it, and revise it. That makes them relevant to investment workflows, where the challenge is not simply producing an answer but producing work an analyst can use.

Mercer’s February 2026 survey of 131 asset managers found that 55% had integrated AI into at least one investment process. Efficiency gains were the most commonly reported benefit, cited by 69%, while improved returns and reduced risk were each cited by only 8% [2]. Broader research coverage could eventually translate into better investment outcomes, but only if the additional work preserves analytical quality and adds differentiated insight.

Deploying AI Agents Into a Live Investment Management Process

I lead AI and quant research at a fundamental asset manager, and we are exploring the boundary between human and AI by building AI agents in-house for our investment management process. The equity analyst’s workflow breaks into six steps. Two agents are live: an Earnings Memo Agent that drafts a summary within 30 minutes of a release for the analyst to review before the morning meeting, and a Post-Earnings Thesis Review agent that checks whether the analyst's investment thesis still holds and returns one of three assessments of: "validated," "concern," or "not validated."

What we learned from building them is mostly about where AI fails.

Coherence. FrontierFinance, a benchmark from S&P Global, Kensho and MIT, gave agents real discounted cash flow (DCF) modeling tasks that take a human expert 11 hours. The agents finished in less than 90 minutes but scored 60% to 75% against 85% to 89% for humans, and graders judged that some outputs would be faster to rebuild than to fix.

The authors identify the core failure as maintaining a coherent picture of how information flows through the model [3], and we are seeing a similar pattern as we test our DCF agent. An experienced analyst would ask how higher capex and shareholder distributions are being funded, and whether the assumptions remain consistent across the model. Our DCF agent does not yet make those checks reliably.

Reliability. Artificial Analysis's AA-AnalystAgent benchmark runs each quantitative analysis task five times. For Gemini 3.1 Pro, the share of tasks solved in at least one of five attempts, a measure of capability, was 81%, while the share solved in all five attempts, a measure of reliability, was 41% [4]. There is a huge gap between capability and reliability. An analyst agent can still be useful when its work requires review, but the practical question is whether checking and correcting its output takes less effort than doing the analysis ourselves.

Attention. In our own study presented at the Japanese Society for Artificial Intelligence Financial Informatics (JSAI SIG-FIN) workshop in March 2026, we had an LLM write earnings summaries for the TOPIX 1000 and compared them with 1,600 sell-side reports [5]. Looking at topic distribution, the AI wrote essentially the same memo for a bank and a pharmaceutical company. Humans shifted focus by industry. To mitigate this, we tried few-shot prompting(showing the model worked examples), GraphRAG (retrieving context from a knowledge graph), and LoRA fine-tuning (lightweight retraining of the model). GraphRAG matched human use of industry vocabulary. None of the methods matched the human distribution of topics.

Our conclusion: we could transfer vocabulary, but we could not transfer attention. AI can learn the jargon. It has not learned where to look in earnings releases, or which management claims to doubt. That is what an analyst acquires by covering a company through a business cycle, and nobody has written down that knowledge.

Our thesis-review agent, PETR, has never returned "not validated" on any of our theses. Sycophancy toward the analyst was our first suspect, but giving AI a role as a critical senior PM barely moved the results. The likelier explanation is the same one: it believes what management says because it does not know what to doubt.

Arvind Narayanan's keynote at the International Conference on Machine Learning 2026 frames knowledge work as three layers: decide, execute, deliver. AI compresses execution, while deciding and delivering remain human and may expand [6].

For our profession, I would put it this way. What we can write down, we can delegate to AI. The CBOK is the written-down part, and AI has already mastered it. The unwritten part, or where to look and what to doubt, is what makes an analyst. And the work ahead is to write down more: as agent skills, as evaluation rubrics, and as organizational knowledge.

We delegate the work to AI, not the responsibility. The recommendation still goes out under a human name, and the job that grows is evaluation, deciding what is good when there is no ground truth to check against. That may turn out to be the most valuable intellectual property a firm owns.

This article summarizes the author's keynote speech, "Analysts with Their Agents: How Agentic AI is Redefining Investment Management," delivered at IEEE CIFEr 2026 on 10 September 2026.