Investors do not need AI to produce more opinions. Markets are already the world’s most efficient opinion factory. What investors need is research that becomes faster, cleaner, and easier to challenge. That last property is the part most AI investing content ignores.
That constraint shapes everything below. Language models are agreement machines: give one a thesis and it will furnish supporting evidence fluently, confidently, and endlessly. Used naively, AI does not improve an investor’s process. It industrializes confirmation bias. The workflow that follows is the minimum viable version I use in my own research and have helped investing-minded operators set up: four steps, no exotic tooling, and a structural defense against the agreement problem.
The four-step loop
Step 1: Define the thesis in one falsifiable paragraph. Before any AI touches the work, write what you believe and what would make it wrong. “Company X’s services segment is being mispriced because the market treats it as cyclical while its revenue base has shifted to multi-year contracts” is a thesis. “AI infrastructure is a big trend” is a mood. The falsifiability requirement is doing real work here: every later step gets evaluated against this paragraph, and a paragraph that cannot be wrong cannot anchor research.
Step 2: Gather the source material yourself. Filings, transcripts, industry data, pricing pages, competitor disclosures. You choose the sources; the model does not. This is deliberate. Source selection is where research quality is actually determined, and delegating it to a model means inheriting whatever its retrieval surfaced, weighted by whatever made it into training data. Primary-source reading stays human. The model is about to make everything downstream of this step faster, which is exactly why this step has to stay clean.
Step 3: Ask AI to structure the evidence, not to opine. This is where the leverage lives. Have the model summarize each filing against your thesis paragraph. Build comparison tables across competitors. Generate the risk checklist for the industry. Map scenarios: what does this position look like if the contract mix reverts, if the key customer concentration bites, if margins mean-revert? Extract every quantitative claim from a bull-case article so each one can be checked. The pattern is consistent: the model organizes and compresses; it does not conclude. Structuring prompts (“list the evidence in this 10-K that bears on this thesis, for and against”) produce checkable output. Opinion prompts (“is this a good investment?”) produce fluent noise.
Step 4: Write the counterargument before acting. Before position sizing, before entry, write the strongest case that the thesis is wrong. Here the model finally gets to argue, on the other side. Ask it to attack the thesis with the same source material. Ask what a skeptical short-seller would say. Then write the counterargument yourself, in your own words, and answer it in writing.
The counterargument is the control surface
Step 4 is the part people skip and the part that makes the loop safe. It works for a mechanical reason: it converts the model’s agreement bias into a feature. A system that agrees with anything will argue the bear case as fluently as it argued yours. Make it do both, then let the friction between the drafts show you where your evidence is thin.
The written answer requirement matters just as much. If you cannot answer the counterargument in writing, you do not have a position; you have an inclination with research-flavored decoration. The discipline compounds: a file of dated thesis-plus-counterargument pairs becomes an audit trail for your own judgment, which is worth more than any single trade. Founders use the identical mechanism when they pre-commit experiment thresholds before seeing data. I wrote that version up in a better cadence for founder experiments, and in both cases the point is the same: decide how you will be wrong before the evidence arrives.
Where AI belongs and where it does not
| Use AI for | Keep human |
|---|---|
| Summarizing filings and transcripts against a thesis | Choosing which sources count |
| Comparison tables across companies and periods | The thesis itself |
| Risk checklists and scenario maps | Position sizing and entry discipline |
| Extracting claims from bull/bear commentary for checking | Deciding when evidence has changed your mind |
| Drafting the attack on your own thesis | Answering that attack in writing |
The left column shares a property: outputs are verifiable against sources you already hold. The right column shares another: each one is a judgment where an error compounds silently. Position sizing deserves special mention because the temptation is real and growing. Models will happily suggest allocations, and the suggestion will sound calibrated. It is not. Sizing encodes your risk tolerance, your portfolio context, and your tolerance for being early, none of which the model knows and all of which it will confidently improvise.
A worked example, compressed
A recent loop on a mid-cap industrial: thesis written Monday (mispriced services transition, falsifiable on contract-mix disclosure). Sources gathered over two evenings: three years of filings, five transcripts, two competitor 10-Ks. The model then produced a contract-mix table across all periods and both competitors, a summary of every management statement about services margins with quarter citations, and a twelve-item risk checklist of which three were specific enough to investigate. Total structuring time: about an hour, versus the several evenings it used to take.
The counterargument round surfaced the position’s real weakness: the mix-shift disclosure the thesis depended on had a definition change two years back, which the model found because it was asked to attack, not summarize. Answering that in writing required one more evening with the older filings and shrank the position size. The workflow did not decide anything. It made the assumptions visible enough that the deciding was cleaner. That is the entire product: if the tool only makes you more confident, the workflow is incomplete.
This loop assumes you already have somewhere to put its outputs: a consistent structure for theses, watchlists, and review cadence. If you do not, build that first; the reasoning is in why research architecture comes before investing decisions.
Common mistakes
Asking the model for conclusions. “Should I buy this?” produces fluent noise with no accountability. Every prompt in the loop should produce output checkable against sources you chose.
Letting AI pick the sources. Model-retrieved evidence inherits invisible selection bias. Faster reading of sources you selected is the win; delegated source selection is the trap.
Skipping the written answer to the counterargument. Reading the bear case and nodding is not the same as answering it. The writing is where the thinking happens.
Using the time savings to take more positions. The loop frees hours; reinvest them in deeper work on fewer names, or the efficiency gain becomes an over-trading subsidy.
FAQ
Can AI actually improve investment research? Yes, in one specific way: it collapses the time between holding sources and having them structured into summaries, comparisons, checklists, and scenario maps. It does not improve judgment, and used without a challenge step it actively degrades judgment by manufacturing confirmation.
What is the biggest risk of using AI for investing? Confirmation at scale. Models agree with the framing they are given, so a bullish investor gets an inexhaustible supply of bullish synthesis. The structural fix is making the model argue the other side before you act.
Should AI decide position sizes? No. Sizing encodes risk tolerance and portfolio context the model does not have. It will still produce a confident-sounding number if asked, which is precisely why the boundary needs to be explicit.
What tools does this workflow require? Any capable general-purpose model plus your existing note system. The edge is not in the tooling. It is in the loop’s structure: falsifiable thesis, human-chosen sources, structuring-only prompts, and a written counterargument before action.