Earnings season is a time problem. In the space of three weeks, hundreds of companies release quarterly results — and for an analyst covering a broad universe, reading and processing every report before the week's trading begins is simply not possible if you do it manually.
The standard workaround is triage by prominence: read the reports from the largest positions closely, skim the rest, and accept that you'll miss things in the tail. It works until it doesn't — until a position you barely glanced at surprises you with guidance that changes the thesis.
Rachel was a buy-side analyst at a mid-size asset management firm. Her coverage universe had grown to 500 names. At 20–30 minutes of careful reading per report, processing all of them in earnings season was a 150–250 hour task. She had a week.
She started using AI batch processing to produce a first-pass structured summary of every report before she read any of them in full. The summaries told her where to spend her time.
What to extract versus what to summarise
There's a useful distinction to make before designing the prompt: extraction versus summarisation.
Extraction means pulling a specific piece of information that exists verbatim in the document — revenue, EPS, guidance figures. The answer is in the text; the model just needs to find it and present it cleanly.
Summarisation means interpreting and compressing — producing a two-sentence plain-English description of the narrative, not just the numbers. This requires reading comprehension and judgement.
Both are valid batch processing tasks, but they have different reliability profiles. Extraction is highly accurate — the number either appears in the document or it doesn't. Summarisation is more variable — the model's interpretation of management tone may differ from your own. Rachel used both, but verified extraction outputs differently from summarisation outputs.
Building the spreadsheet
She focused on the Management Discussion and Analysis (MD&A) section rather than the full report. MD&A is where management explains what happened and what they expect — it contains both the key metrics in narrative form and the forward-looking language that tells you how they feel about the next quarter. At typically 3,000–6,000 words, it's long enough to be meaningful but short enough to process efficiently.
| ticker | company_name | sector | report_period | mda_text |
|---|---|---|---|---|
| ACME | Acme Technologies plc | Software | Q1 2026 | [MD&A section text — typically 3,000–6,000 words] |
The sector column helped her sort the output by industry when reviewing — software businesses warrant different benchmarks for revenue growth than industrials. The report_period column prevented any confusion when she reused the same spreadsheet structure across quarters.
Writing the prompt
Prompt used:
You are a financial analyst reading the MD&A section of a quarterly earnings report. Extract the following information and produce a structured output. Be precise — report numbers as stated, do not estimate or round unless the report itself gives only rounded figures.
REVENUE: [reported revenue for the period, with currency and period — e.g. "$4.2bn Q1 2026" — or NOT_REPORTED if absent]
REVENUE_GROWTH: [YoY revenue growth as stated, or calculated if both periods are given — e.g. "+12% YoY" — or NOT_REPORTED]
EPS: [reported EPS, adjusted or GAAP as stated — note which — or NOT_REPORTED]
GUIDANCE: [next quarter or full-year guidance as stated — revenue range, EPS range, or key metric — or WITHDRAWN or NOT_PROVIDED]
GUIDANCE_CHANGE: [RAISED / MAINTAINED / LOWERED / NOT_APPLICABLE — relative to prior guidance if comparable language is present]
MGMT_TONE: [POSITIVE / CAUTIOUS / MIXED — based on language in the MD&A: forward-looking statements, risk disclosures, hedging language]
KEY_RISKS_MENTIONED: [up to three specific risks management named — e.g. "FX headwinds", "supply chain normalisation", "macro softness in Europe"]
SUMMARY: [two sentences maximum — the single most important thing that happened this quarter and what management is signalling about the next one]
Rules:
— Extract numbers verbatim where possible — do not interpret or restate unless necessary for clarity
— MGMT_TONE should reflect the actual language used, not your assessment of the numbers
— If the MD&A is ambiguous on a field, output UNCLEAR rather than guessing
— SUMMARY should be readable by someone who has not read the report — no unexplained jargon
The MGMT_TONE field is the most nuanced. Rachel found it genuinely useful — not as a trading signal in itself, but as a flag for which reports to read carefully. A company that beat estimates but had CAUTIOUS tone often had something in the MD&A worth understanding. POSITIVE tone on a miss suggested management was either confident in the recovery or managing expectations aggressively. Either way, worth reading.
Running the batch
500 rows on Claude Sonnet 4.6. MD&A sections vary significantly in length, density, and how carefully management buries or emphasises the important numbers. Sonnet's reading quality on this kind of structured but discursive financial text was noticeably better than Flash-tier models — it found figures that were stated once in a subordinate clause rather than in a table, and it caught hedging language that a less attentive read would miss.
The batch ran overnight. 500 MD&A sections, some over 6,000 words each, completed by morning. Total cost: £18.
What the output looked like
| ticker | AI responses (excerpt) |
|---|---|
| ACME | REVENUE: $1.84bn Q1 2026 REVENUE_GROWTH: +8% YoY EPS: $0.92 adjusted GUIDANCE: Full-year revenue $7.4–7.6bn; prior guidance $7.5–7.7bn GUIDANCE_CHANGE: LOWERED MGMT_TONE: CAUTIOUS KEY_RISKS_MENTIONED: Enterprise deal slippage into Q2; FX headwinds in EMEA; macro softness affecting mid-market segment SUMMARY: Acme reported in-line revenue but lowered full-year guidance citing delayed enterprise closures and currency pressure. Management's language around the mid-market was noticeably more cautious than prior quarters, suggesting demand softness beyond the deal timing explanation. |
How Rachel used the output
She sorted the 500 results by GUIDANCE_CHANGE first, then cross-referenced with MGMT_TONE. Her triage logic:
- GUIDANCE: LOWERED + MGMT_TONE: CAUTIOUS — read the full report today, update the model.
- GUIDANCE: RAISED + MGMT_TONE: POSITIVE — skim the summary, update the model if a position, otherwise flag for next week.
- GUIDANCE: NOT_PROVIDED or MAINTAINED + MGMT_TONE: MIXED — read the SUMMARY carefully, decide whether to escalate.
Of the 500 reports, she read roughly 80 in full over the course of the week — the ones the triage flagged as meaningful to her positions or her thesis. The remaining 420 she tracked through the structured output, noting any guidance changes or risk flags that warranted a note in her records.
What AI can and can't do here
The batch output is a reading aid, not a research substitute. A few things to be clear about:
Extraction is reliable; interpretation is not infallible. The revenue figure, EPS, and guidance numbers were accurate across the vast majority of reports Rachel checked. MGMT_TONE was more variable — the model's read on tone occasionally diverged from hers, particularly for companies with highly formulaic boilerplate language where the meaningful signal was in what management didn't say.
The SUMMARY field needs spot-checking. Most summaries were accurate and useful. Occasionally one omitted something she considered important, or emphasised a secondary point over the primary story. She read the full report for all positions regardless — the summary shaped what she was looking for, not whether she looked.
This doesn't replace sector knowledge. The model doesn't know that an 8% growth figure is disappointing for a company that guided 12% last quarter, or that "deal slippage" is a recurring theme for this management team. That context is the analyst's job. The batch gives you the facts efficiently; making sense of them still requires a human with the thesis.