How to Analyze Earnings with AI

Before you point AI at a report, know what matters, so you can ask for the right things. An earnings report is mostly four signals, and the order of importance is not the order they appear.

• Revenue and EPS versus the estimate. The headline beat or miss is reported revenue and earnings per share against the consensus analysts expected. A beat on both is the baseline good outcome, but it only matters relative to expectations, not in isolation.
• Forward guidance. What management says about next quarter and the full year often moves the stock more than the quarter just reported. A strong quarter with weak guidance frequently sells off; a soft quarter with raised guidance can rally.
• Margins and the trend. Gross and operating margin show whether the business is getting more or less profitable. One quarter is noise; the direction over several quarters is the signal.
• The call tone and Q&A. The earnings call is where management frames the numbers and where analysts push on the soft spots. What gets questioned, and how evasively it gets answered, is information the press release will not give you.

AI can help with all four, but the first two require numbers it does not reliably know on its own (the consensus estimate), so you supply those. The next sections walk through the workflow.

Lead with the digest. AI is at its most reliable summarizing a long document you paste in, because the answer is grounded in your text rather than reconstructed from training data. Give it the full press release and call transcript and ask for the five things that actually moved.

The output is a fast read of the quarter, not a fact sheet of record. Use it to orient yourself, then drill into the numbers in the next step. ChatGPT, Claude, and tools like FinChat all handle this well; the difference is mostly whether you paste the transcript yourself or the tool supplies it.

The single most useful AI task on an earnings report is turning the prose into a clean, repeatable table. Ask for revenue, EPS, margins, and guidance with the source line for each, so you can verify fast.

Then compare to expectations. This is where people get burned: AI does not reliably know the consensus estimate, and if you ask it to compare without supplying the number, it can invent a plausible-sounding estimate that is wrong. Pull the real consensus from your broker, an estimates provider, or the financial press, paste it in, and ask AI to lay out reported versus expected with the variance. The model does the formatting; you provide the ground truth on the estimate.

The transcript is where the story lives. Past the headline numbers, the prepared remarks reveal what management wants you to focus on, and the analyst Q&A reveals what they are nervous about. Paste the transcript and ask AI to separate firm commitments from soft optimism, quote any guidance verbatim, and list every question an analyst pushed back on.

A useful prompt: “Here is the call transcript for [COMPANY], [QUARTER]. [PASTE]. List the hard guidance with exact wording, separate it from vague optimism, and tell me the three things analysts questioned most pointedly and how directly management answered.” The value is the contrast between confident claims and the questions underneath them. Because the answer is grounded in the transcript text, this is one of the lower-risk AI tasks, but still quote guidance verbatim so a mis-transcribed number gets caught.

A stock can beat on revenue and EPS and still fall, which confuses people new to earnings. The reaction is about the gap between results and what was already priced in, not the results alone. AI can explain the mechanics, why a beat with cautious guidance sells off, why a miss with a raised outlook rallies, why a stock that ran up into the print falls on a merely good quarter.

What AI cannot do is see the live reaction or predict it. The core model has no market feed, so it does not know today's move, and any forecast of the after-hours direction is a guess in confident clothing. Use AI to understand why a reaction happened once you can see it (pull the live price from your broker or a market source), not to bet on the move beforehand. Ask “the stock fell X% after a revenue and EPS beat, what in the report or guidance most plausibly explains that?” and you get a useful framing; ask “will it go up tomorrow?” and you get noise.

To be upfront, since this is our site: the gap a general chatbot leaves on earnings is that it never sees what you own. You can analyze a report perfectly and still not know whether the company is a 2% sliver or a 20% concentration of your portfolio, which is what actually determines how much the move matters to you. That is the specific gap Walnut closes. Walnut connects your existing brokerage through SnapTrade, a regulated aggregator, so after an earnings move you can see how the affected holding sits against the S&P 500 (outperforming, in line, or lagging) and where it stands in your portfolio, with live numbers rather than generic examples.

You can work through Claude, ChatGPT, or a built-in assistant, and because broker feeds rarely pass cost basis, returns are framed as window returns rather than realized profit and loss, and Walnut says so. It is read-only by default, and every trade is gated behind your explicit approval, so the AI can summarize a report and frame the position but never act on its own. Walnut is one option for the portfolio side, not the only one, and it does not replace the document analysis that plain ChatGPT or a tool like AlphaSense does well. Walnut is not an investment adviser.

From a connected account you can dig into a specific stock after it reports, an ETF that holds it, or a theme the result speaks to. For more on the AI research side, see how to use ChatGPT for stock picks or the best AI portfolio analyzer.

There is no single best tool; there are general assistants and purpose-built platforms, and they do different jobs. Here is the honest split.

• ChatGPT and Claude are general assistants. They are excellent at summarizing and explaining a report and transcript you paste in, and they are free or cheap. They do not know the consensus estimate, do not see your portfolio, and hallucinate numbers recalled from memory, so you supply the source and verify the figures.
• AlphaSense is an enterprise research platform that searches across filings, transcripts, and broker research, with AI summarization on top. It is powerful and priced for institutions, useful if you analyze many companies rather than one.
• Quartr aggregates earnings call transcripts, audio, and slide decks across thousands of public companies. It is a source of the raw material; pair it with an AI assistant to summarize or question a call.
• FinChat (now Fiscal.ai) combines historical financials and segment data with an AI chat interface, so you can ask about a company's numbers without pasting a release each time. Useful for the number-extraction and trend steps.
• Walnut is narrower: it connects your real brokerage so an AI can frame an earnings move against what you actually hold, with live numbers, and place only trades you approve. It does not replace the document analysis the others do.

For one-report depth, a general assistant plus a transcript source covers most people. For screening across many companies, the purpose-built platforms earn their cost. For the portfolio impact, you need something connected to your account.

The failure modes on earnings are specific and worth memorizing, because each one looks exactly like a correct answer.

• Hallucinated numbers. AI can transpose a digit, mix GAAP with non-GAAP EPS, or state a revenue figure that is simply wrong, all phrased as confidently as a true fact. Always verify a number you act on against the actual press release, which is why every prompt above asks for the source line.
• Invented estimates. Ask AI to compare results to expectations without giving it the consensus, and it can fabricate an estimate that sounds plausible. The model does not reliably know the real consensus. Supply it yourself.
• Stale knowledge. The core model has a training cutoff and no market feed, so it does not know the live price, the actual reaction, or anything that happened after its cutoff. It cannot tell you today's move.
• Confident forecasting. Asked where the stock goes next, AI will answer, and the answer is a guess. It cannot predict the reaction. Treat any prediction as fiction.

None of this makes AI useless for earnings; it defines the lane. AI is a summarizing and explaining tool that works best on text you supply and numbers you verify. It is not a data source and not a forecaster.

The whole workflow condensed: what you do at each step, and where AI genuinely helps versus where you stay in control.

Notice the pattern: AI does the reading and formatting; you supply the source text and the estimates, verify the figures, and make the call. Keep the buy or sell verdict out of the prompt and in your own hands.

AI is a strong assistant for earnings analysis and a poor decision-maker. Use it to summarize the report and call, extract revenue, EPS, margins, and guidance into a table, compare them to estimates you supply, and read the tone and Q&A, with prompts that paste in the source and forbid a verdict. Do not trust it for the consensus estimate, the live reaction, or any number recalled from memory, because it hallucinates figures and cannot forecast the move. Verify everything quantitative against the actual release, and decide for yourself. To see how an earnings move affects what you actually hold, connect your broker through a tool like Walnut and keep every trade behind your approval.