An earnings call is mostly qualifiers. Management says demand is strong "though we expect some moderation," or that they'll spend aggressively "only when we see real demand." Those hedges are the content — they're how guidance encodes conditionality. When an LLM summarizes a call for an agent, the qualifier is exactly what tends to get cut, and a June 2026 study shows what that costs: on real earnings-call transcripts, summaries flipped the investment decision in up to half of cases, usually without saying anything false.
TLDR:
- On S&P 100 earnings calls, naive LLM summaries flipped a bear/neutral/bull call in 23.9% of cases, and token pruning in 39.1–50.8% — against an 8.8% no-compression baseline.
- The mechanism is decontextualization: the headline claim survives, the management qualifier that conditions it doesn't.
- Returning structured, cited transcript fields instead of a re-summary sidesteps the failure — the qualifier travels with the quote.
The measurement, on transcripts specifically
The paper When Summaries Distort Decisions tested 297 S&P 100 earnings-call transcripts by feeding a model either the full transcript or a compressed version, then checking whether the bear/neutral/bull decision changed. Compression that changes the decision has lost fidelity, even if every sentence in it is true.
| Compression of the earnings call | Decision-flip rate |
|---|---|
| No compression (baseline variance) | 8.8% |
| Contextualized summary | 22.2% |
| Naive summary | 23.9% |
| Token pruning | 39.1–50.8% |
Source: When Summaries Distort Decisions, arXiv:2606.29251, Table 2 (earnings-call transcripts). "Flip" = the compressed context yields a different decision than the full transcript.
Figure 1: Even the careful, context-preserving prompt more than doubles the baseline flip rate. Source: arXiv:2606.29251.
The number that should worry anyone shipping a summarize-then-reason pipeline is the contextualized one: a prompt explicitly told to preserve context still flipped 22.2% of calls. Careful prompting helps, but it doesn't close the gap.
What decontextualization looks like
The paper's diagnosis is that compression "may preserve headline claims while stripping away decision-relevant caveats, comparisons, or expectation framing" (Lee et al.). On an earnings call, the caveat is usually a hedge or a conditional clause — and it's usually load-bearing. Here's a real one, from a neocloud CFO discussing capital spending:
Figure 2: The quoted statement is from a real earnings call (retrieved via FocusAlpha); the shortened summary is illustrative.
Management framed the bulk of its GPU spend as demand-gated: "this massive 80% spend will come only when we see real demand." A summary that keeps "plans massive GPU spend" and drops "only when we see real demand" isn't lying — it's just turned a conditional into a commitment. An agent reading the summary now over-weights a capex number that management explicitly made contingent. Multiply that across a portfolio of calls and you get the paper's flip rates.
This isn't one quirk — it's a family of dropped qualifiers, each one clause, each capable of changing the decision:
| Qualifier type | On the call | What a summary drops |
|---|---|---|
| Conditional | "spend will come only when we see real demand" | "only when we see real demand" → reads as committed |
| Comparison | "revenue up, off a low base" | "off a low base" → reads as strength |
| Expectation framing | "capex came in ahead of our own plan" | "ahead of plan" → reads as an overrun |
| Customer-specific | "that cost won't apply to every customer" | "won't apply to every customer" → reads as structural |
Source: decontextualization mechanism per arXiv:2606.29251; call phrasings illustrative of real hedging patterns.
Why structured fields avoid it
The fix isn't a better summarizer — the contextualized-prompt result shows prompting only goes so far. The fix is to not throw the source away. When an agent retrieves a structured, cited transcript field — the actual sentence, speaker-attributed, with the surrounding segment one hop away — the qualifier is still attached. The agent reasons over what management said, not a lossy paraphrase of it.
The two approaches diverge exactly where it matters:
| Free-text re-summary | Structured, cited field | |
|---|---|---|
| The qualifier | Often dropped | Travels with the passage |
| Provenance | Rebuilt or lost | Cited to speaker + segment |
| Measured decision-flip | 22–51% on earnings calls | Reasons over the source, not a paraphrase |
Source: flip rates from arXiv:2606.29251, Table 2 (earnings calls).
That's the design behind FocusAlpha's retrieve and documents endpoints: an agent gets the source-cited passage, not a re-summary, so "only when we see real demand" arrives intact. We cover the cross-company retrieval side of this in the earnings call transcripts API pillar, and the same fidelity problem on filings in when AI compresses your filings. Compression is unavoidable at some layer — but unaudited summarization of an earnings call is a decision you're making without seeing the number the paper puts on it.
FAQ
Do LLM summaries of earnings calls change investment decisions?
Yes, measurably. On 297 S&P 100 earnings-call transcripts, naive summaries flipped the bear/neutral/bull decision in 23.9% of cases and token pruning in 39.1–50.8%, versus an 8.8% baseline from the model's own run-to-run variance.
What is decontextualization in an earnings-call summary?
Compression that keeps a headline claim but drops the caveat or conditional that qualifies it — for example, keeping "plans massive spend" while cutting "only when we see real demand." The summary reads as true but supports a different decision than the full call.
Does a careful summarization prompt fix the problem?
Only partly. In the study, a prompt explicitly instructed to preserve context still flipped 22.2% of earnings-call decisions — better than a naive summary but far above the 8.8% baseline. Prompting reduces the loss; it doesn't eliminate it.
How can agents read earnings calls without this risk?
By retrieving structured, source-cited transcript passages instead of a re-summary, so the qualifiers travel with the quote. The agent reasons over what management actually said, speaker-attributed and traceable to the call.
Why are earnings calls especially vulnerable to summary distortion?
Because guidance is encoded in hedges and conditionals — "if demand materializes," "ahead of plan," "off a low base." These qualifiers are the highest-information, lowest-word-count part of the call, so they're the first thing compression drops and the most costly to lose.
How does FocusAlpha reduce earnings-call summary distortion?
FocusAlpha returns earnings-call content as structured, cited fields via its retrieve and documents endpoints — the source passage with its qualifiers, speaker, and citation intact — rather than a second-hand summary, so the agent never reasons over a decontextualized paraphrase.
What is FocusAlpha?
FocusAlpha is a SEC filings API and agent-ready financial data layer: it turns SEC filings (10-K, 10-Q, 8-K, 13F), earnings-call transcripts, and other trusted company communications into structured, normalized data where every value keeps its citation back to the source document. AI agents connect via API or MCP to research public companies from complete, trusted information.