AI Investing Workflows for Earnings Call Research

Here is one we got badly wrong, and it is the reason this guide is a list of failures rather than a list of prompts. An extraction run once produced a clean, confident, well-written analysis of Stryker's earnings call. It read beautifully. Every figure in it was internally consistent. It was also, in part, an analysis of Straker Limited — a New Zealand translation company that shares a similar ticker string and had quietly landed in the corpus under the wrong name. The same thing happened again with Altria, where a Portuguese pulp producer called Altri turned up in the file.

The model had not hallucinated anything. It did exactly what it was asked, on a document nobody had checked. That is the defining failure mode of AI earnings-call research, it has almost nothing to do with the quality of the model, and almost every article on this subject skips it in favour of prompt tips.

So the useful version of this guide is not a list of things a language model can do with a transcript. It is the list of things that break, why they break, and the workflow shape that survives them.

Why "paste the transcript and ask for a summary" fails

Not because the model is weak. Because two things upstream of the model are almost always wrong, and neither is visible in the output.

The first is that the input is dirtier than anybody expects. Ticker collisions put another issuer's call in your folder, as above. The same call gets ingested twice — a pre-recorded results segment and the live Q&A, stored as separate documents — which a naive pipeline happily counts as two quarters. Some companies publish their prepared remarks as a separate file, so the "transcript" in hand is Q&A only, and a model reading it concludes that management never mentioned something it in fact announced in a document you do not hold.

The second is that the question itself is unfalsifiable. "Summarise this call" has no wrong answer, which means you cannot distinguish a good output from a merely plausible one. Ask instead for things that can be checked — a number, a period, a speaker, a verbatim quote — and errors stop being invisible.

Rule one: verify the document is the company

Before any analysis at all, check the transcript's own header against the company you asked for. This sounds too basic to write down, right up until the week it costs you.

Automate it as a hard gate rather than a habit: if the company name inside the document does not match the one requested, the file is quarantined, not analysed. An AI workflow with no identity check is, structurally, a machine for producing confident answers about the wrong business.

Rule two: define what counts before you extract

Most bad AI research output is not false. It is uncheckable, because the model was allowed to record things that could never be graded either way. Four conditions fixed this. A statement is worth extracting only if all four hold: it was forward-looking when made, about a future period rather than a reported result; it was management-owned, meaning the company committed to it rather than an analyst asserting it in a question; it is falsifiable, so that some future observation could show it kept or broken; and it is checkable, naming both a metric or action and a period.

"We expect gross margin between 48% and 49% in the March quarter" passes all four. "We remain confident in our long-term positioning" fails three of them. Put those conditions into the prompt and the output stops being a wall of optimistic sentences and becomes a list you can actually grade.

Rule three: no outcome without a quote

The most dangerous thing an AI research pipeline produces is a resolution — a verdict that a target was met or missed. It is dangerous precisely because it looks like the end of the work, so nobody goes back to check it.

One early run marked a $100 million target as substantially achieved. The number it had been matched against appears nowhere in six transcripts of that company. The model had constructed a perfectly reasonable inference and then presented it as an observation, which is a different thing entirely.

The fix is a rule that admits no exceptions: a verdict requires a citation. If the model cannot quote the sentence in which the outcome is stated, it does not get to assert the outcome — it records the commitment as still open and moves on. That single constraint removed more error than any model upgrade did.

Rule four: never let missing data become a finding

This is the subtlest one, and it is where automated management-credibility scoring quietly goes wrong.

When a commitment is made and then never mentioned again, the silence looks like evasion. Sometimes it is. Far more often it is an artefact of the corpus: roughly a third of the company datasets assembled for the US record had at least one quarter missing, nearly all of them inside a single window between April and September 2025. A pipeline that does not know about its own holes will read a hole as a broken commitment, and will accuse a company of going quiet on something it may well have addressed on a call that was never ingested. The rule, then, is that the pipeline must be able to distinguish "we do not have that call" from "they did not say it." Anything less produces an accusation rather than an analysis.

The same discipline applies to reporting conventions. A guided effective tax rate gets reported in the 10-Q and is almost never reconciled out loud on a later call, which is a convention rather than a dodge. Treating the two identically turns a scoring system into a random number generator with a grudge.

Rule five: judge each document, not the average

Mondelez ran full prepared remarks through one quarter of 2025 — roughly 46% of the transcript came before the first analyst spoke — and then switched to a Q&A-only format from July, where prepared remarks accounted for about 4% of the document.

Average those across the window and you get 19%, a number that describes no call the company ever held and neatly hides a clean format change. Any rule a workflow applies — read the prepared remarks for guidance, the Q&A for candour — has to be evaluated per document, because companies change format mid-year without announcing it. The general form of this mistake is worth naming, because it recurs everywhere in quantitative research: an average taken across a regime change describes neither regime.

A workflow that survives

Five stages, and the order is the point, because each one refuses to pass bad work forward.

Ingest and inventory first: collect the transcripts and record what you have, which quarters are missing and which are duplicates. The gap list is an output of the process, not an inconvenience in it. Then verify identity, per file, quarantining mismatches. Then extract with citations, pulling only statements that meet the four-part test, each carrying metric, number, period, speaker and verbatim quote. Then resolve only on evidence, grading a commitment closed only where the outcome is quoted and leaving everything else open. Then — and this is the stage people skip — read the exceptions yourself.

That last stage is not a concession to human vanity. It is where the analysis actually happens. The value of the output sits in the outliers: the misses, the revisions, the topics that recur without ever resolving. The first four stages exist to get you there holding a short list instead of forty transcripts. This is the shape Promise Tracker takes, and the reason its ledger stores the quote alongside every verdict.

What AI is genuinely good at here, and what it isn't

It is good, and better than a person at scale, at finding every forward-looking statement in a long document, normalising them into a consistent structure, noticing that a phrase appeared in three consecutive calls and not the fourth, and grading whether an answer addressed the question that was asked.

It is not good, and is unlikely to become good by prompting harder, at deciding whether a commitment mattered, at knowing that a company changed its reporting basis, at distinguishing a missing document from a missing statement without being told, or at judging whether management's explanation is credible. Those all require context that is not in the transcript. The workflows that work treat the model as a very fast, very literal research assistant with no memory of last year and no knowledge of what it was not given — which is an accurate description of what it is.

How Inve implements this workflow

Everything above describes a pipeline that has been built and run in anger, which is why the failures are specific rather than theoretical. Inve is that pipeline, with all five rules enforced rather than intended.

The identity check runs per document, so another issuer sharing a ticker is quarantined instead of analysed. Extraction records only statements that pass the four-part test, each carrying the metric, the number, the period, the executive and their verbatim quote — which is why a commitment in the Promise Tracker can always be traced back to the sentence that created it. A verdict is written only where the outcome is quoted, so a target stays open rather than being closed on a plausible inference. And the corpus knows its own gaps, so a quarter that was never ingested is never scored as a company going quiet.

The result is 6,132 US calls and 37,529 tracked commitments you can read rather than rebuild — with the part that still requires you, reading the exceptions, left where it belongs.

Free for US readers while the US market has no billing rail.

Where this approach can mislead you

The honest counter-case is that every rule above trades recall for precision, and that trade is not free. Requiring a quote before resolving a commitment means genuine outcomes disclosed in a filing rather than said aloud stay marked open indefinitely. Refusing to read silence as evasion means real evasion gets scored as missing data. Those defaults were chosen deliberately — a research tool that occasionally says "unknown" is more useful than one that is confidently wrong about a company's integrity — but somebody with different stakes might reasonably choose the other way.

The failure modes here also come from one architecture over one two-year window. They generalise well as categories and poorly as frequencies: your ticker-collision rate will be different, but the fact that you need an identity check will not be.

And none of it makes the output an investment view. A perfectly extracted ledger tells you what management said and what happened to it afterwards. Deciding what that is worth remains the part with no shortcut.

Frequently asked questions

The owner's question

The temptation with a fast research pipeline is to point it at more companies. The better use is to point it at the same companies for longer. What you actually want to know about a business — whether this management tells you the truth early when things go wrong — is not a question about volume at all. It is a question about memory, held over years, on a small number of names you intend to own. Automation is worth having because it makes that kind of patience affordable, not because it makes you faster.

Related reading

• How to Read an Earnings Call Transcript
• How to Track Management Guidance Across Quarters