Earnings calls are among the richest public sources of company information: prepared remarks, guidance, and an unscripted question-and-answer session. They are also long, repetitive and released in clusters during reporting season. Few tasks seem better suited to large language models.
Where models perform well
Summarisation of prepared remarks is now routine and reliable when the model is constrained to the transcript. Extraction of explicitly stated figures — revenue, margins, guidance ranges — works well when outputs are checked against the source. Comparing the language of consecutive quarters to flag new topics or dropped phrases is another strength, because it relies on pattern detection rather than judgement.
Where caution is needed
- Numbers not in the text. Models may infer or invent figures that sound plausible. Every number should be traceable to a quoted passage — the classic hallucination risk.
- Tone and sentiment. Executives are trained communicators. Sentiment scores derived from word choice can reflect style more than substance.
- The Q&A session. Evasive or partial answers are often the most informative part of a call. Models tend to summarise what was said, not what was avoided.
The most informative part of an earnings call is often what management chose not to answer.
Designing a robust workflow
The strongest workflows we have seen share three features. First, retrieval: the model receives the actual transcript and cites passages. Second, structured output: figures go into defined fields that can be validated automatically against reported financials. Third, human review of anything that will inform a decision. Open-source projects such as FinGPT and FinRobot provide useful starting points for teams building their own pipelines.
Bottom line
Language models genuinely reduce the time needed to process earnings season. They are best treated as fast, well-read research assistants whose work must be checked — not as analysts whose conclusions can be traded on directly.
Evaluating a transcript tool
For readers considering a commercial or open-source product, a short evaluation exercise is worth more than any marketing material. Select five recent calls from companies you know well. Write your own one-paragraph summary and list of key figures for each. Then compare the tool's output against yours on three criteria: factual accuracy of every number, coverage of the points you considered most important, and whether it noticed anything you missed.
Track errors by type. A tool that occasionally misses a minor point is very different from one that occasionally states a figure that was never said. The second type of error is far more costly in practice, because it is harder to detect.
Context windows and long documents
Modern models can process an entire transcript at once, but performance is not uniform across a long document. Details buried in the middle of the Q&A are more likely to be missed than those in prepared remarks. Splitting transcripts into sections and processing each with explicit instructions often improves recall.
Comparisons across companies
Cross-company comparisons — which retailers mentioned inventory pressure this quarter, for instance — are a powerful use of language models. They depend heavily on consistent prompts and output formats, and benefit from spot-checking a sample of results by hand.