What AI in Investment Research Actually Does, and Where It Fails

Two resource-sector investors, Sultan Amr Ali and Bill Powers, are running a roughly 1,000-word AI prompt to screen arbitrage spreads and a 2,500-word prompt to monitor their equity positions. Those functions used to require a junior analyst on the payroll.

That detail is a signal about where research economics are heading. The combination of freely available advanced language models and the document-heavy disclosure regimes that govern mining and energy has made it possible for an individual investor to run workflows that once demanded institutional infrastructure. The same combination has flooded the information environment with AI-generated misinformation, creating a verification problem that did not exist five years ago.

This is a practical map of that shift. Here is what is actually being done, how the prompts are structured, why checking AI output against primary documents is now a required discipline rather than a nicety, and what the competitive and regulatory stakes look like for anyone who gets it wrong.

What junior analysts used to do, and what AI now handles

Start with the specific tasks. The functions AI is now documented to handle in resource equity are the same ones that used to fill a junior analyst’s week.

• Parsing regulatory filings and pulling the material detail out of hundreds of pages
• Cross-referencing personnel across different corporate entities to map who sits where
• Assessing the historical reliability of the Qualified Persons who author technical reports (a Qualified Person is the credentialed expert who signs off on an NI 43-101 or equivalent technical study; NI 43-101 is Canada’s mineral disclosure standard, with JORC playing the same role in Australia)
• Feeding bankruptcy First Day Reports into a model to extract restructuring patterns in a given sector
• Framing the first pass of a financial model and flagging variables a human might overlook

Each of those tasks used to cost real hours. Mapping executive networks by hand meant working through filing after filing. Reading a distressed-company First Day Report and pulling out the restructuring shape was a day’s work for someone junior. That labour is now compressed into a prompt.

This is not a hypothetical workflow. Sultan Amr Ali and Bill Powers are the only named practitioners who have publicly described this specific function set in resource equity, and their disclosure is what makes the shift concrete rather than speculative.

The survey data confirms they are not outliers. Research from Christ, Kim, and Yip, published on SSRN on 30 September 2025, surveyed 190 sell-side equity analysts and found the tools are being embedded into first-pass work rather than final judgment.

58% of surveyed sell-side analysts use AI mainly to streamline existing processes such as summarising text and collecting data, according to Christ, Kim, and Yip. These are core junior-analyst functions.

The pattern holds beyond the sell side. Russell Investments, in its 10 December 2025 commentary, reports that quantitative teams use large language models to extract data and accelerate code development, while fundamental teams use them to summarise earnings calls and filings and to assist with first-pass valuation. Blankespoor and co-authors, in research dated 1 June 2026, argue that generative AI substitutes for traditional intermediaries by producing more timely, higher-quality summaries of disclosures.

Here is what that convergence tells you. When named practitioners and a survey of 190 analysts point in the same direction, the workflow has moved from experiment to baseline. The sophisticated investor who is not using these tools is now processing information more slowly than the people they are competing against for the same undervalued position.

How the prompts are actually structured for resource equity tasks

The signal is in the architecture. Ali and Powers have disclosed a two-stage prompt structure, and the design tells you more than any general advice about “using AI for research” ever could.

1. Stage 1, the screening prompt (approximately 1,000 words): A customised tracking prompt that filters for specific arbitrage spreads and liquidity thresholds. This is the wide net.
2. Stage 2, the monitoring prompt (approximately 2,500 words): A longer query that watches for fundamental shifts in the equity positions that survived the screen. This is the close watch.

Why does the monitoring prompt run more than twice the length of the screening prompt? Because length is how you encode domain knowledge. A longer prompt can carry more context about what to look for, how to weight competing signals, and what to escalate for human review.

That matters most when the underlying document is complex. Screening for a liquidity threshold is a relatively narrow instruction. Monitoring a position across a multi-hundred-page bankruptcy filing or a dense NI 43-101 report requires the model to behave like a trained analyst, and the only way to get that behaviour is to instruct it in detail. The more complex the document, the more the model needs before it stops summarising and starts analysing.

A structured technical assessment framework for mining exploration covers the same categories an AI monitoring prompt needs to encode: reserve methodology, capital cost assumptions, permitting status, and Qualified Person credentials, which is precisely why the monitoring prompt runs longer than the screening prompt.

This is the closest publicly available specification of how AI prompts are engineered for resource equity, and it is directly applicable to building your own.

Enterprise platforms for filing analysis and cross-document verification

The institutions have built the same logic into their tooling, with one feature the practitioner prompts handle manually: the audit trail. This is the critical differentiator between research-grade output and a general-purpose chat answer.

Bloomberg Document Insights lets analysts run natural-language queries across more than 200 million company documents, with contextual links back to the source material. FactSet Mercury is a generative AI knowledge agent for financial fundamentals with audit trails, meaning an output can be traced back to the specific data behind it. S&P Capital IQ’s ChatIQ Pro supports multi-document analysis with cross-filing consistency checks, useful for comparing one quarter’s numbers against another or two technical reports on the same project.

The common thread across all four is that they anchor outputs to source documents. That anchoring is not a convenience feature. It is the mechanism that makes AI output trustworthy enough to act on, which is exactly the problem the next section addresses.

Why AI-generated misinformation has made verification a required discipline

Here is the uncomfortable part. Every one of those workflows produces output that can be confidently wrong.

Ali and Powers make this a practitioner observation, not a regulator’s hypothetical: the spread of AI has already generated significant volumes of automated corporate misinformation, and that reality is why they treat manual verification of primary documents as a standing requirement in resource equity. This is a signal from inside the workflow, not a warning from outside it.

The risk sharpens in technical documents. A large language model is not designed to validate geological feasibility, mine-plan engineering, or reserve-volume accuracy. It can produce a fluent summary of an NI 43-101 report that misreads a discount-rate assumption or misclassifies a reserve, and a single distorted input can throw off an entire valuation screen.

JORC resource classification uses three confidence tiers (Inferred, Indicated, and Measured) that carry meaningfully different risk profiles; a model that conflates them will produce valuation inputs that are structurally wrong before any other assumption is applied.

Three categories of error are specific to resource documents and worth watching for:

The JORC Code 2012 resource classification standards define the framework that Australian and many international explorers must follow when reporting Inferred, Indicated, and Measured resource estimates, with independent competent person sign-off required before any resource figure can be disclosed to the market.

• Reserve classification errors, where the model conflates an Inferred resource with a higher-confidence category
• Cost assumption misreads, where a capital or operating cost figure is pulled from the wrong scenario or the wrong year
• Personnel or Qualified Person misattributions, where the model assigns a report or a track record to the wrong individual

The regulators reached the same conclusion from the outside. In remarks summarised by law firm Alston & Bird in July 2024, then SEC Chair Gary Gensler warned about AI’s failure modes directly.

AI models may produce “unpredictable and inaccurate outcomes (e.g., hallucinations)” that reinforce historical biases and pose risks to brokers, advisers, and market stability, according to Alston & Bird’s analysis of SEC Chair Gary Gensler’s remarks.

The concern is on the record elsewhere too. The SEC, NASAA, and FINRA issued a joint Investor Alert on AI and investment fraud in January 2024, and SEC Enforcement Director Gurbir Grewal warned in April 2024 that representations about AI use must not be materially false or misleading, flagging hallucinations and conflicts of interest in recommendations.

Now the detail that should change how you treat any unverified output. No quantitative data exists on how often AI fabricates resource-sector research details. There is no published prevalence rate for hallucinated filings or fabricated figures in stock-analysis documents.

That gap is itself the argument. With no benchmark to calibrate against, you have no basis for deciding an unverified output is probably fine. The only defensible default is to treat verification against the primary document as the rule, not the exception.

The competitive and regulatory stakes for investors who ignore this shift

So the tool is real and the risk is real. What does it cost to sit this out?

The competitive answer runs through thin coverage. Blankespoor and co-authors argue that generative AI substitutes for traditional intermediaries by delivering faster, higher-quality disclosure summaries. In heavily covered large-caps, that speed advantage gets competed away quickly. In junior mining and energy, where only a handful of analysts, or none, follow a given name, the same advantage compounds, because there are fewer people racing you to the same information.

34% of 1,074 investment professionals surveyed in PwC’s Global Investor Survey 2025 rely on generative AI “to a large or very large extent” when assessing how companies manage risks and opportunities.

Put that 34% figure against the thin-coverage reality of junior resource stocks and the picture is stark. The informational gap between AI-augmented and non-augmented investors is wider in this sector than in large-cap equities, and it is likely widening further.

The scaling is already visible at the top. BlackRock’s Aladdin Auto Commentary, a generative AI reporting tool launched in October 2025 and first implemented by Morgan Stanley Wealth Management, automates the kind of narrative monitoring reports junior analysts used to write. Citadel’s AI Assistant generates research reports and assembles portfolio-specific reading lists, though it is explicitly positioned as a copilot rather than an autonomous decision-maker.

That framing is the professional consensus, and it matters for how you build. Christ, Kim, and Yip find analysts use AI to streamline workflows, not to generate independent stock calls. Russell Investments frames it as summarisation and coding support, not final decisions. Certain judgments stay with humans:

• Technical feasibility assessment of a mine plan or engineering design
• Geological and reserves validation
• Long-term environmental and permitting risk

What “AI washing” enforcement means for investors evaluating AI-marketed products

There is a legal dimension too, and it cuts the other way. If you are evaluating a product that markets itself on AI capability, you now have precedent to hold it to.

SEC Release No. 2024-36, issued on 18 March 2024, charged two investment advisers, Delphia and Global Predictions, with false and misleading statements about their use of AI. The charges established “AI washing,” overstating AI capability to attract money, as an active enforcement category.

The practical read is straightforward. When any research service or investment product claims AI capability, that claim must now be accurate and verifiable, and there is regulatory precedent behind treating an inflated claim as a red flag rather than a selling point.

Building a verification-first AI workflow in resource equity

Enough context. Here is one workflow you could build this week, calibrated to the document-heavy, thin-coverage reality of resource equity.

1. Structure the prompt with domain context. Use the Ali and Powers two-stage model as the reference: a shorter screening prompt to filter, a longer monitoring prompt to watch surviving positions. Length is where you encode what to look for and what to escalate.
2. Run the AI pass and extract the flagged data points. Let the model do the volume work, pulling commodity price assumptions, discount rates, reserve figures, and personnel details out of the filing.
3. Cite-check every extracted point against the primary document. For each figure that matters, ask the model to name the specific passage it drew on, then open the original NI 43-101, JORC report, or filing and confirm it yourself.
4. Apply human judgment to model construction and the investment decision. The extracted parameters feed your model; you build and validate the model, and you make the call.

Transcript analysis for mining research applies the same document-volume logic the article describes for filings: the value is in extracting consistent signals across many calls or presentations rather than reading each one in full, and the same verification discipline applies when a model summarises what a CEO said versus what the text actually says.

The cite-check step is the individual investor’s version of an institutional audit trail. FactSet Mercury and Bloomberg Document Insights solve verification by anchoring outputs to source data automatically. You solve it manually, by demanding the citation and checking it.

Large language models are best used to summarise filings and assist with first-pass valuation, while human analysts handle deeper valuation work and the investment decision itself, according to Russell Investments.

Where the human-AI boundary sits in documented practitioner workflows

The financial modelling boundary deserves its own line, because it is where the temptation to over-delegate is strongest.

AI is documented as useful for outlining a modelling framework and surfacing variables you might have missed. It is not documented, in any named practitioner or institutional workflow, as the thing that builds and validates the model. JuniorMiningIntelligence’s tool extracts the parameters; the investor cross-checks them against the original technical report and applies the risk judgment.

Amundi’s scenario workflow is the clearest example of that division built into a process. Human experts articulate the scenarios, the AI generates structured narratives from them, and humans validate the resulting allocations before anything is acted on. Across every documented workflow, from a two-person team to BlackRock, AI handles volume and speed while humans hold judgment and verification. Invert that division and you are the one most exposed to hallucination.

What separates the investors who will gain an edge from those who will just add noise

The question was never whether to use AI. It is whether you have built the discipline to use it in a way that produces signal rather than importing hallucinations into your research stack.

The resource investor who builds a verification-first, prompt-structured workflow gains a genuine edge in a thin-coverage sector. The one who outsources judgment to unverified output is exposed to exactly the misinformation risk the practitioners and regulators have documented. Same tool, opposite outcomes.

The direction of travel supports the disciplined approach. The Q4 2025 Investment Artificial Intelligence Trends report, published 30 January 2026, notes that after two years of experimentation the industry has converged on insight summarisation, report drafting, and predictive intelligence as the highest-value uses, and the tools are maturing toward exactly those functions.

The AI investment cycle has moved from infrastructure build-out toward application-layer deployment, and the practitioner workflows described in this article sit at that application layer: the models exist, the document corpus exists, and the edge now belongs to whoever structures the prompts most precisely.

Junior mining and energy sit at the sharp end of this. The mix of AI adoption, thin analyst coverage, and complex technical disclosure creates both the largest potential advantage and the largest hallucination exposure of any equity sector.

The practitioners who are ahead, including the two named resource investors, share one practice. They treat AI output as a first draft that requires primary-source verification before it enters the investment process. Build the workflow deliberately, verify everything, and the document-heavy nature of resource equity becomes your advantage rather than your obstacle.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions.