The AI Edge: Modern Investing Tools

The industry sorts financial AI into four capability tiers, and knowing which tier a tool belongs to tells you how much to trust it:

1. Rule-based systems — a clerk following explicit instructions ("flag every stock with debt over 30% of market cap"). Deterministic and exactly repeatable.
2. Machine learning — an analyst finding patterns in historical data.
3. Generative AI — a writer synthesizing unstructured text.
4. Agentic AI — an investigator making autonomous multi-step decisions (the emerging frontier, and the one regulators are most nervous about).

Sharia screeners live in tier 1 — and their precision comes from the data plumbing underneath. When a US public company files its results with the SEC's EDGAR system, the numbers aren't submitted as prose: every figure is tagged in XBRL, a machine-readable standard where each data point (revenue, debt, cash) carries a unique tag under the accounting taxonomy. A screener queries these tags through an API and gets back exact, audited integers straight from the balance sheet — then runs ordinary arithmetic on them. No interpretation, no prediction, no "vibes." That is why a rules-based screener can be trusted with the AAOIFI math in a way a chatbot never can: one retrieves and calculates; the other guesses what a plausible answer looks like.

Beyond Items 1, 1A, and 7, two more narrative sections are good LLM material:

• Item 1B (Unresolved Staff Comments) — open disputes with SEC staff about the company's past filings. An AI can summarize what the regulator is unhappy about without you parsing dense legal correspondence.
• Item 3 (Legal Proceedings) — pending lawsuits: the core allegations, jurisdictions, and management's view of the potential impact.

A caveat from the research: generic "chat with your PDF" setups often fail on SEC filings, because page after page of near-identical tables confuses the retrieval step — systems can return irrelevant chunks or miss data buried in tables. Specialized multi-agent systems built for financial documents scored 56% on a notoriously hard financial-document benchmark versus 19% for a standard model — better, but still far from reliable. The practical lesson: paste the exact text you want analyzed into the prompt yourself rather than trusting the tool to go find it.

The professionals' experience matches this division of labor. Major quantitative funds like Man Group and Two Sigma weave generative AI into research for coding, data discovery, and summarization. But when a European equity fund piloted a GenAI-driven stock screener, it was abandoned: analysts couldn't explain why the model favored certain obscure small-caps, and trades that can't be explained can't be defended — to compliance, or to clients. Even institutions keep the writer away from the trigger.

The research catalogs how AI numbers go wrong — useful for spotting the damage:

Even formatting changes the model's behavior: 0.007, 7.0e-3, and 7×10⁻³ are the same number to you, but produce entirely different token sequences — and different predictions — inside the model.

The scariest failure is the most articulate one. Researchers testing models on deliberately flawed math problems (missing variables, impossible premises) found high-performing models often don't notice — instead they produce confident, verbose, well-structured answers justifying the broken premise. Applied to investing: ask an AI for a buy/sell verdict on garbage inputs, and it won't just compute wrong numbers — it will hallucinate a compelling narrative to defend them. Fluency is not accuracy.

The golden rules aren't just personal hygiene — regulators are actively policing AI hype. In the SEC's first-ever "AI washing" enforcement actions, two investment advisers were fined a combined $400,000: Delphia ($225,000) had claimed its algorithms put client data into "a predictive algorithmic model" to pick investments — false; Global Predictions ($175,000) had marketed itself as the "first regulated AI financial advisor" offering "expert AI-driven forecasts" — also false. Both violated the Investment Advisers Act. SEC Chair Gary Gensler's summary: "AI washing hurts investors."

The SEC, FINRA, and state regulators (NASAA) have also issued joint warnings that AI's popularity is being exploited to generate fake data, run scams, and even deep-fake impersonations of legitimate investment professionals. FINRA's Regulatory Notice 24-09 flags model accuracy, data privacy, embedded bias, and exploitation by bad actors as live concerns.

The takeaway for you as a retail investor: adopt the regulators' skepticism as your own. An LLM is not a fiduciary — it has no concept of your risk tolerance, no license, and no accountability. Treat any AI-generated number or definitive recommendation as inherently unverified until you've checked it against a primary source, a deterministic tool, or a licensed human.

Three upgrades to the CTST framework, straight from the research:

• Chain-of-Thought (CoT). For any multi-step reasoning task, explicitly tell the model to "think step by step" or "show your work before answering." Forcing intermediate steps stops the model from compressing complex logic into one leap — in benchmark tests, this raised accuracy on complex reasoning problems from 17.9% to 74.4%.
• Program-of-Thought (PoT). If a calculation genuinely must involve an AI, don't let it do the math in prose. Have it write code instead (a spreadsheet formula or small script containing the variables and formula), then run that code yourself in a real calculator/spreadsheet. The fragile text generator plans; a deterministic machine computes.
• Strict rule modifiers. When feeding filings to a model, add hard rules to the prompt: "Never fabricate or extrapolate financial data. Only reference the exact text provided. If data is missing, say 'not found' and move on." And paste the source text into the prompt yourself — don't trust the model's memory or browsing.

All of this rolls up into one operating principle the research calls the hybrid architecture, which enterprise compliance systems enforce and you should too: a quantitative layer (screeners, spreadsheets, APIs) that does every calculation, and a qualitative layer (LLMs) that reads, summarizes, and explains. The two layers never swap jobs. Think of the LLM as an entry-level analyst with an encyclopedic vocabulary and zero mathematical comprehension: enormously useful, supervised relentlessly, and never allowed near the trade button.

The research distills safe AI-assisted research on official filings (the SEC's EDGAR database) into a repeatable workflow:

1. Define your benchmark universe first, manually. Pick 8–12 companies genuinely comparable in industry, size, and business model before involving the AI — relevance in, relevance out.
2. Isolate the exact sections you need. Researching risk? Target Item 1A and the MD&A. Don't feed the whole 150-page filing and hope.
3. Paste the raw text into the prompt yourself. Don't ask the model to "look up" the filing — it may hallucinate content or pull from inaccurate third-party summaries. Ground it in the primary source.
4. Add hard rules to the prompt. "Never fabricate or extrapolate financial data. Only reference the text provided. If data is missing, say 'not found' and move on."
5. Verify before you rely. If the summary claims "a $400 million headwind in European markets," find that figure in the actual filing before it enters your notes.

One more current from the research worth knowing: regulators worldwide (the SEC and FINRA in the US, the EU's AI Act, the UK's FCA) are shifting their core question from "what did the algorithm do?" to "who is accountable when an AI makes a decision?" For your portfolio, the answer is always the same person: you.