How to use AI for investing without handing it the steering wheel
Jack Buffet shows how to use AI to organize filings, challenge an investment thesis, check scenarios, and document decisions without treating a chatbot as an adviser.
For: An individual investor who wants AI to improve the research process without outsourcing judgment, verification, privacy, or trade execution.
Treat an AI assistant like a fast junior researcher who never gets tired and occasionally invents a door in a wall.
Useful? Absolutely.
Unsupervised? Not with the family money.
AI can help organize a filing, explain an unfamiliar term, compare two business models, test arithmetic, expose missing questions, and turn scattered notes into a readable investment memo. It can also use stale data, misunderstand accounting, omit a decisive fact, fabricate a source, repeat market manipulation, or deliver an answer with much more confidence than evidence.
That leaves AI with a valuable but limited job: improve the research process without becoming the decision-maker.
The short version: Begin with primary documents you obtained yourself. Ask AI to extract, organize, calculate, compare, and challenge—not to predict a price or name a winner. Require a source beside every material claim, reopen those sources yourself, recompute important numbers outside the chat, protect private information, and write the final thesis in your own words. Never give an unverified AI service brokerage credentials or permission to trade.
Put AI in the right seat
The easiest way to misuse AI is to ask a decision question too early:
“Should I buy this stock?”
That prompt invites a polished conclusion without requiring the work that makes a conclusion useful. The tool may not know your goal, time horizon, other holdings, tax position, liquidity needs, or ability to absorb a loss. It may also be working from incomplete or outdated information.
The SEC’s Investor.gov AI-fraud alert warns investors not to rely solely on AI-generated information. Even when the input appears accurate, the output can be faulty or fabricated. AI-generated conversations can also encourage emotional or impulsive decisions.
So divide the work into three zones:
| Zone | Appropriate AI jobs | The human responsibility |
|---|---|---|
| Green: organize and explain | Build a filing index, define terms, summarize a section, produce a comparison outline, format notes | Supply the authoritative document and verify the summary against it |
| Yellow: analyze and challenge | Calculate ratios, create scenarios, compare periods, list risks, argue the opposite side | Check formulas, assumptions, source dates, units, and omitted facts |
| Red: decide and execute | AI may help draft a decision memo | A human decides position size and timing; no unverified tool receives credentials or trading authority |
AI belongs near the evidence and far from the button.
Step 1: define the assignment before opening the chatbot
Write the research question in a form that can be answered with evidence.
Weak assignment:
Find me a stock that will double.
Better assignment:
Help me evaluate how this company earns money, what could impair those earnings, how management allocates cash, what expectations appear necessary to justify the current valuation, and what evidence would invalidate the thesis.
Then record the constraints the AI cannot infer:
- the goal this money serves;
- the earliest date it may be needed;
- whether the investment is a core holding or an optional position;
- the maximum position size under consideration;
- existing exposure to the same company, industry, customers, or economic risks;
- the required return or margin of safety, if you use one; and
- the conditions that would make the idea unsuitable before research begins.
This is not personalization magic. It is a guardrail against letting an interesting company silently become an inappropriate investment.
Step 2: collect the evidence yourself
Do not ask AI to “find everything important” and assume it found the right company, the newest filing, or the complete document.
For a U.S. public company, begin with the SEC’s free EDGAR filing search. A basic research packet might contain:
| Document | What it can answer |
|---|---|
| 10-K annual report | Business model, audited financial statements, risks, competition, legal matters, executive discussion, and a full-year baseline |
| 10-Q quarterly report | More recent financial statements, operating changes, liquidity, and updated risks |
| 8-K current reports | Material events such as acquisitions, leadership changes, financings, or significant agreements |
| DEF 14A proxy statement | Executive pay, board structure, ownership, related-party matters, and shareholder proposals |
| Earnings release and presentation | Management’s selected framing and non-GAAP measures, which should be reconciled to the filings |
| Fund prospectus and shareholder report | Objectives, strategy, risks, costs, holdings, and results for a mutual fund or ETF |
The SEC’s guide to reading a 10-K explains that the filing gives a detailed picture of the business, risks, and operating and financial results. It also notes that investors must decide how much weight to give non-GAAP measures.
Save the document date, reporting period, form type, and source URL. A correct number from the wrong quarter is still the wrong number.
Step 3: ask for extraction before interpretation
Make the first pass boring. Boring is where the evidence lives.
Use a prompt like this with a filing you have supplied:
Using only the attached filing, create a table of revenue, operating income, net income, operating cash flow, capital expenditures, cash, and debt for every period shown. Preserve the units. For each value, cite the filing section, table, and page or document location. If a value is unavailable or ambiguous, write “not established” instead of estimating it.
Then verify the table against the filing.
Follow with focused extraction prompts:
- “List every customer-concentration disclosure and cite the exact section.”
- “Separate management’s reported GAAP measures from adjusted measures.”
- “Identify every material change in risk-factor language from the previous annual report.”
- “List debt maturities and the entity where each obligation sits.”
- “Extract acquisitions, divestitures, repurchases, dividends, and stock-based compensation.”
- “Create a list of statements that describe expectations rather than completed results.”
The phrase using only the attached filing narrows the task. It does not guarantee accuracy. The citations give you a route back to the document; they are not proof until you follow them.
Step 4: make AI show its work on calculations
AI can be useful for arithmetic because it can create a repeatable structure quickly. It can also transpose a number, mix millions with billions, confuse a balance-sheet point in time with a cash-flow period, or calculate a ratio from incompatible definitions.
For every material calculation, require four fields:
| Field | What to record |
|---|---|
| Formula | The exact mathematical relationship |
| Inputs | Each number, unit, reporting period, and source location |
| Result | The calculated value with sensible precision |
| Limitation | What the calculation omits or cannot establish |
Example prompt:
Calculate five-year revenue growth, operating margin by year, free cash flow under two clearly stated definitions, net debt, and share-count change. Show every formula and input. Do not fill missing values. Flag any change in accounting presentation that may make periods incomparable.
Then rebuild the important calculations in a spreadsheet or calculator. If the thesis depends on a number, the number deserves an independent second life outside the chat.
Do not ask AI for a single “fair value” and accept the decimal point as authority. Ask for scenarios:
- What assumptions produce the low, middle, and high cases?
- Which input has the greatest effect on the result?
- What happens if margins normalize, growth slows, or dilution continues?
- What outcome appears embedded in the current price?
- Which assumptions are evidence and which are guesses?
The value of a scenario is not the output. It is the argument about the inputs.
Step 5: use AI to attack the thesis
Confirmation bias can make a research folder look busy while every document is being recruited for the same conclusion.
Give AI the job of opposing you:
Act as a skeptical investment-committee member. Using only the supplied sources, identify the five strongest reasons this thesis could fail. Separate business risk, balance-sheet risk, management risk, valuation risk, and portfolio-fit risk. For each objection, cite the evidence and name the additional information needed to resolve it.
Run a second prompt from the other direction:
What facts in these documents weaken the bear case? Do not repeat management’s claims without supporting evidence.
Then ask the uncomfortable omission question:
Which conclusion in my draft memo has the weakest support? List every sentence that depends on an uncited claim, forecast, or assumption.
AI is often more useful as a disagreement generator than as a recommendation generator. A tool does not need to know the future to point out that your logic skipped a bridge.
Step 6: compare alternatives, not just the company
A stock does not compete only with cash. It competes with every other reasonable use of the money.
Ask AI to construct a comparison framework, then populate it from verified sources:
| Question | Candidate company | Broad-market fund | Closest business competitor | Do nothing yet |
|---|---|---|---|---|
| What job would it perform? | ___ | ___ | ___ | Preserve optionality |
| What are the main return drivers? | ___ | ___ | ___ | Current holding’s return |
| What can cause permanent loss? | ___ | ___ | ___ | Inflation or missed opportunity |
| What overlaps existing holdings? | ___ | ___ | ___ | None added |
| What evidence requires monitoring? | ___ | ___ | ___ | Conditions for revisiting |
| What are the fees, taxes, and trading costs? | ___ | ___ | ___ | Usually no transaction |
“Do nothing yet” deserves a column. AI makes producing another report nearly free, which can create pressure to turn research activity into portfolio activity. Waiting for clearer evidence or a better price is a valid result.
Step 7: keep a verification ledger
An AI-generated answer is not a source. Build a ledger that forces every consequential claim to touch evidence.
| Claim | Primary source | Period/date | Verified by human? | Status |
|---|---|---|---|---|
| Revenue grew by ___ | 10-K, financial statements | Fiscal years ___ | Yes / No | Established / unresolved |
| Management changed guidance | 8-K or earnings release | Date ___ | Yes / No | Established / unresolved |
| Largest risk is ___ | Filing plus your analysis | Date ___ | Yes / No | Judgment, not fact |
| Current valuation implies ___ | Price source plus model | Date/time ___ | Yes / No | Scenario, not fact |
Use three labels consistently:
- Fact: directly supported by an identified source.
- Calculation: reproducible from identified inputs.
- Judgment: an interpretation or forecast that may be reasonable but is not established fact.
AI likes to make all three sound equally smooth. Your ledger should make them look different.
Step 8: write a decision memo before placing a trade
Turn the verified work into a one-page memo:
Investment memo
| Field | Your answer |
|---|---|
| Security and date reviewed | ___ |
| Job in the portfolio | ___ |
| Time horizon | ___ |
| Maximum position size | ___ |
| Business in one plain sentence | ___ |
| Three evidence-backed strengths | ___ |
| Three ways the thesis can fail | ___ |
| Valuation range and assumptions | ___ |
| Existing portfolio overlap | ___ |
| Tax and liquidity considerations | ___ |
| Evidence that would prevent purchase | ___ |
| Evidence that would trigger review or sale | ___ |
| Next filing or review date | ___ |
| Unresolved questions | ___ |
AI can format the memo and challenge its gaps. You should write the final thesis in language you understand well enough to defend without reopening the chatbot.
If you cannot explain why the investment belongs, what could break it, and why its size is survivable, another prompt is not the missing ingredient.
Keep private information out of the research loop
The safest investment-research prompt usually does not need your account credentials, Social Security number, tax return, complete brokerage statement, bank numbers, private employer documents, or personal identity documents.
Before uploading anything, ask:
- Is this information already public?
- Does the task require the private detail, or can it be replaced with a category or rounded hypothetical amount?
- What does the provider retain, use for training, share, or allow administrators to review?
- Can the same task be completed with a redacted copy?
- Would exposure create financial, identity, employment, legal, or insider-information risk?
Never paste passwords, authentication codes, recovery keys, API keys, or brokerage login credentials into a chatbot. Do not upload material nonpublic company information. If a tool asks for account access, stop and verify the provider, registration status, permissions, data practices, and revocation process independently.
FINRA has warned about unregistered auto-trading services that invoke AI while seeking brokerage access or promoting unsupported profitability. The concern is not merely a bad prediction: an outside service with trading authority can create losses before the account owner realizes transactions occurred.
Research assistance and custody should remain strangers.
Know the red flags AI cannot polish away
The CFTC’s advisory, AI Won’t Turn Trading Bots into Money Machines, warns that AI cannot predict the future or sudden market changes. Promises of guaranteed returns, enormous win rates, secret algorithms, or effortless automated income remain red flags with newer vocabulary.
Walk away from a pitch that combines AI with:
- guaranteed or nearly guaranteed returns;
- urgency or a supposedly closing window;
- requests for crypto, wire transfers, or credentials;
- an unregistered person or platform;
- screenshots instead of independently verifiable records;
- celebrity, executive, or family videos that cannot be authenticated;
- a strategy that cannot explain losses, fees, spreads, taxes, and drawdowns; or
- instructions to keep the opportunity secret.
AI can generate a confident voice, a professional headshot, a polished report, invented customer praise, and a chart that rises from left to right. None of those items proves that a person exists, a firm is registered, money is invested, or returns occurred.
A practical 30-minute AI research session
When you want a compact routine, use this order:
- Five minutes: define the question. Write the investment’s proposed job, time horizon, and maximum size.
- Five minutes: collect sources. Open the latest filing yourself and record its date and URL.
- Five minutes: extract. Ask AI for a cited table of the key operating and balance-sheet facts.
- Five minutes: challenge. Ask for the strongest bear case and the claims with the weakest evidence.
- Five minutes: verify. Reopen the cited sections and recompute the two numbers that matter most.
- Five minutes: document. Update the verification ledger and memo. End with “research more,” “reject,” “watch,” or “eligible for further consideration”—not an automatic order.
Thirty minutes will not complete serious due diligence. It can prevent an interesting headline from becoming an unexamined trade.
The bottom line
Use AI to make the work more organized, more skeptical, and easier to repeat.
Let it index the filing. Let it build the first table. Let it explain a term, draft a formula, argue against your thesis, and expose the blank cells in your memo.
Do not let it manufacture missing facts, substitute fluency for evidence, define your risk tolerance, hold your private credentials, or place the trade.
AI should help you ask better questions. The primary sources answer the factual ones. The financial plan defines the boundaries. You remain responsible for the decision.
If your question is instead how to invest in the AI trend, read Jack’s separate guide, AI is a theme, not a portfolio.
This article provides general educational information, not individualized investment, tax, legal, cybersecurity, or financial advice. AI systems, provider terms, market data, company facts, regulations, and security threats can change. Verify consequential claims against current primary sources, protect sensitive information, and consider qualified professionals who can evaluate your complete circumstances.