How to Use AI for Earnings Season: Previews, Calls and Model Updates
The earnings-week workflow with AI, step by step: the preview, the 8-K release, the call, guidance changes, the model update and the post-print note.
AI for earnings season works best as a crew that does the reading and the arithmetic while the analyst makes the calls: it drafts the preview, reads the release and the transcript the moment they land, lines up new guidance against last quarter’s words, updates the model with every changed cell cited, and drafts the post-print note. Kepler runs that whole week for a coverage list: agents start when a company’s 8-K lands, transcripts come from Quartr, our transcript partner, and code computes every number in the model and the note. We build Kepler. This guide walks the earnings-week workflow step by step, says what to automate and what to keep a person on, and shows how agents can do the first pass overnight.
The earnings-week timeline
| When | What arrives | The analyst’s job | What AI should do | What to check |
|---|---|---|---|---|
| A week out | Nothing new | Preview: what matters this print | Pull last quarter’s guidance, your estimates, peers’ read-across | Consensus source and date |
| Release | 8-K Item 2.02, Exhibit 99.1, sometimes a deck | First read: beat, miss, guide | Extract every figure, compare to estimates, show GAAP and adjusted | The non-GAAP reconciliation |
| Call | Webcast, then transcript | Listen for what changed | Quote guidance, check spoken figures against the release, compare tone | Quotes and speakers |
| That night | Model update | Roll actuals in, revise estimates, cite every changed cell | Adjusted vs reported, recasts | |
| Next morning | The note | Draft in house format, every figure traced | The view is yours | |
| Days or weeks later | 10-Q | Reconcile | Compare the 10-Q with the release, update segment and balance sheet detail | Any figure that changed |
Step 1: The preview
A good preview fits on a page and answers one question: what will decide this print? Build it from five things:
- Last quarter’s guidance, verbatim. Paraphrased guidance hides the hedges management chose.
- Your estimates against consensus for revenue, margin, EPS and the two or three KPIs that move the stock. Note the consensus source and its date.
- Read-across from peers that already reported. By the time a mid-season name reports, its suppliers, customers and competitors may have said plenty about demand, pricing or inventory.
- What management said to watch. Last call’s Q&A often tells you which number they’ll defend.
- Questions for the call. Three, ranked.
This is reading-heavy work, and AI does it well when it has read everything. “Every company in our semiconductor coverage that reported this month, what each said about inventory at customers, with the quote” is a five-minute question in Kepler and an afternoon by hand.
Step 2: The release (8-K Item 2.02 and Exhibit 99.1)
Most companies publish results first in a press release furnished on Form 8-K under Item 2.02, Results of Operations and Financial Condition, with the text attached as an exhibit, by convention Exhibit 99.1 (Form 8-K instructions). Item 2.02 information is furnished, not filed, unless the company says otherwise. Any non-GAAP measure in the release must come with a reconciliation to the nearest GAAP figure under Regulation G and Item 10(e) of Regulation S-K.
What to pull in the first ten minutes:
- Headline results against your estimates and consensus, GAAP and adjusted side by side.
- Segment and KPI tables.
- The guidance table, if there is one.
- The reconciliation: what was added back this quarter that wasn’t last quarter.
- Anything that changed definition: a KPI recalculated, a segment renamed, a metric dropped.
Kepler reads the release as it lands and builds the comparison against your estimates in code, so the first read is ready before the call starts.
Step 3: The call
Item 2.02 lets a company hold its call without a separate 8-K if the call follows the furnished release within 48 hours, was announced in advance by press release, is broadly accessible, and the financial information discussed is posted on the company’s website. Most companies do exactly that, usually on the day of the release.
What to get from the transcript:
- Guidance, in management’s words, from prepared remarks and from Q&A, where it often gets sharpened.
- Figures said aloud that aren’t in the release, checked against the release and later the 10-Q.
- Changes in language on the topics you track: a move from “stabilizing” last quarter to “improving” this one is a question for the Q&A.
- The questions analysts asked and the ones management deflected.
Quotes have to be exact and attributed. Kepler’s transcripts and IR decks come from Quartr, and each statement is tied to its speaker and passage; our post on how Quartr helps Kepler build citation-grade earnings research explains how a quote opens the transcript at the timestamp.
Step 4: Guidance changes
Lay the new sentence beside the old one, then classify: raised, lowered, narrowed, maintained or withdrawn. Then do the arithmetic.
A worked example, with made-up numbers. Last quarter, full-year revenue guidance was $10.0 to $10.4 billion. This quarter it’s $10.3 to $10.5 billion, and nine-month revenue is $7.6 billion.
- Old midpoint $10.2B, new midpoint $10.4B: a $200M raise.
- Implied Q4 at the old midpoint: $10.2B − $7.6B = $2.6B.
- Implied Q4 at the new midpoint: $10.4B − $7.6B = $2.8B.
If the quarter just reported beat by $150M, most of the raise is the beat flowing through, and the real question is whether Q4 consensus already sat near $2.8B. Check the basis every time, because a guide that switches basis can look like a raise while lowering the comparable figure: GAAP or adjusted, constant currency or reported, with or without a pending acquisition.
Step 5: The model update
Roll actuals in from the release, then refine when the 10-Q arrives with full statements and notes. Revise estimates for the lines guidance touched. Three rules:
- Keep adjusted and reported figures apart, each traced to its line in the reconciliation.
- Watch for recasts. If the company resegmented or reclassified, prior quarters in your model need the same treatment; see as filed vs as restated.
- Reconcile the 10-Q to the release. Release figures occasionally change by the time the 10-Q is filed, for example when a subsequent event moves an accrual.
This is where retyped numbers do the most damage, because a wrong input compounds through every forecast. In Kepler, no model types an input: actuals come from the filing in code, each changed cell cites its line, and the 10-Q reconciliation runs automatically when it lands. Peer multiples refresh the same way; see how to build trading comps with AI.
Step 6: The post-print note
The PM reads the headline and the first paragraph. Put the view there: what changed and whether it matters. Then results against you and consensus, the guidance change in management’s words, two or three call takeaways quoted with the speaker, the estimate changes old and new, and open questions. AI drafts this well in a saved house format. In Kepler, the note template is a playbook, and every figure in the draft links to the release, the call or the 10-Q it came from. The view stays with the analyst.
What to automate and what to keep
| Task | Automate | Keep a person on |
|---|---|---|
| Pulling last quarter’s guidance and Q&A | Yes | |
| Read-across from peers’ calls | Yes | Which read-across matters |
| Beat or miss table, GAAP and adjusted | Yes | Whether the quality of the beat holds up |
| Guidance comparison and implied figures | Yes | What management meant |
| Tone and language changes | First pass | The interpretation |
| Model actuals and cited inputs | Yes | Estimate changes |
| The note’s first draft | Yes | The headline and the call |
How the tools on the market approach earnings week
- Claude for Financial Services added an earnings analyses Agent Skill in its October 2025 update, which researches quarterly transcripts and financials to pull out key metrics, guidance changes and management commentary, along with an Aiera connector for real-time call transcripts.
- Aiera offers real-time AI transcripts of earnings calls alongside human-reviewed versions, delivered through enterprise APIs and MCP integrations.
- Daloopa sells an Excel add-in that updates models during earnings season by linking cells to its source-linked database of 6,000+ tickers.
Three things to test in any of them: whether a figure said on the call gets checked against the release, whether each changed cell in the model cites its line, and whether the work starts on its own when the 8-K lands.
Agents that do the first pass overnight
Earnings season is a scheduling problem: a dozen names in two days, many reporting after the close. Event-driven agents take the first pass as each document lands. When a company in the book furnishes its 8-K, an agent reads the release and builds the beat-or-miss table before the call; when the transcript lands, it compares guidance and checks spoken figures; when the 10-Q lands, it reconciles and updates segment detail. The digest is waiting before the morning meeting. Two requirements: each agent acts with exactly the permissions of the person who set it up, and nothing it reads in a release or transcript can change what it does. For the tools that cover each phase, see the best AI for earnings call analysis; for how funds fit this into the rest of the week, see how hedge funds use AI for research.
How Kepler does this
Your coverage and your model sit in your firm’s ontology. Our forward-deployed engineers build a living model of your firm: the names in your book, your models, the KPIs your PM watches for each one, and your definitions of adjusted metrics. Kepler also ships with an ontology of public companies built from SEC filings, so each company’s segments, KPIs and history are already resolved.
Every document is read in full as it lands. Releases, 10-Qs and 8-Ks arrive inside a corpus of 26M+ SEC filings, and call transcripts and IR decks from our partner Quartr join 40M+ IR documents and transcripts, for 14,000+ companies. Extraction models we train read each one at ingest, so every figure, quote and slide sits beside its passage, and “what did every peer say about pricing” returns a complete list.
Code computes every number. The model never writes a figure. Beats and misses, implied guidance, the model update and every number in the note run as code over cited inputs. Click a cell to see its source line; click a calculated figure to see the formula. When the call and the filing disagree, you see both, flagged.
Agents and playbooks run the week. Your preview, your earnings update and your note template are saved as playbooks. Agents trigger on each 8-K, transcript and 10-Q, overnight, acting as their owner and never taking instructions from what they read, and every run is logged. Results land in Excel, Word, Slack, the morning digest, or inside Claude and ChatGPT. Our engineers built this layer before at Palantir. They set up your coverage, models and templates on site, and they’re with your team for the first print.
Try it on last quarter’s prints at app.kepler.ai, or book a demo before your next earnings season.
Frequently asked questions
How can AI help with earnings season?
AI can take the reading and the arithmetic off an analyst's plate across the whole week. Kepler drafts each preview from last quarter's guidance and your model, starts work the moment a company's 8-K lands, compares new guidance with the prior quarter's wording, updates your model with every changed cell cited to the release, call or 10-Q, and drafts the post-print note in your house format. The analyst keeps the judgment.
Where does a company publish its quarterly results first?
Usually in a press release furnished on Form 8-K under Item 2.02, Results of Operations and Financial Condition, with the release attached as Exhibit 99.1. The 10-Q follows within 40 days of quarter-end for large accelerated and accelerated filers, or 45 days for others, and contains the full statements, notes and MD&A. Some companies file both on the same day.
What should an earnings preview include?
Last quarter's guidance in management's exact words, your estimates beside consensus for the lines and KPIs that drive the stock, what peers that already reported said about shared themes, the two or three numbers that will decide the print, and the questions you want answered on the call. Keep it to one page so the PM reads it.
How do you track guidance changes between quarters?
Put the new guidance sentence next to last quarter's, then classify the change: raised, lowered, narrowed, maintained or withdrawn. Compute the change at the midpoint, and the implied remaining-year figure when full-year guidance moves. Check that the basis is the same, such as GAAP or adjusted, constant currency, and whether acquisitions are included.
Can AI update my financial model after earnings?
Yes, and the inputs are the part to check. Excel assistants can now edit workbooks directly, but when a model types each historical figure, every cell inherits its error rate. The safer pattern is a tool that pulls each actual from the release or 10-Q in code, cites the cell to its line, keeps adjusted and reported figures apart, and flags where the call and the filing disagree.
What should a post-earnings note include?
A headline on what changed in your view, results against your estimates and consensus, the guidance change with management's words, two or three call takeaways quoted with the speaker, the estimate changes old and new, what it means for the thesis, and open questions. Every number should trace to the release, the call or the 10-Q.