Skip to content

Methodology

Check it against your own numbers.

We could put a percentage on this page. You would have no way to check it. So instead we show you how the system works and how to measure it.

The source

Every question can be traced back.

An AI that invents plausible questions is dangerous in capital markets communication. That is why no question arises in auraIR without a source.

Every generated question comes with the verbatim quote from the document it rests on, including the page reference. You look it up and see at once whether the question is justified. What is not in your documents does not appear in the question set.

No paraphrasing, no summarizing. The source is the original sentence from your document.

Question
You are cutting the revenue forecast by €40 million, three months after the original outlook. Which assumption changed?
Source
“The Management Board now expects revenue between €412 and €428 million for financial year 2026.”
Q1 2026 quarterly report, page 4 · quoted verbatim
Traceable
Model versionfixed
Temperature0
Second runidentical result

The procedure

How we test our analysis.

The most honest test is the past. A call that has taken place has a transcript. And that transcript records what analysts actually asked.

Step 1
Reconstruct the starting pointWe load only the documents that existed before the call: report, presentation, disclosure, the quarter's peer transcripts. The call itself is left out.
Step 2
Generate the question setauraIR generates the question set as in live operation, with a fixed model version and temperature zero. No fine-tuning, no cherry-picking the best runs.
Step 3
Compare against the real callThe transcript provides the questions that were actually asked. Each one is checked: was it in the question set, yes or no?
Step 4
Evaluate both directionsHow much of the call did we foresee, and how much of the question set came up in the call? Only both numbers together give an honest picture.

The procedure runs again before every release. Changes to the analysis have to prove themselves against it.

Plain talk

The questions you would put to us.

How accurate is the system?

In our tests, auraIR foresees the large majority of the questions that actually come up in the call. We deliberately do not state a percentage here: it comes from tests with one issuer, and your analysts ask differently. In the demo we run the test with your last report and your last transcript. Then you see the number that applies to you.

What about questions that don't come up?

They exist, and we do not optimize to avoid them. An extra question costs reading time and is deleted in seconds. A missing question costs seconds of hesitation in front of several hundred listeners. We weight completeness higher than a lean list.

Does the AI invent numbers?

It cannot, because every question is tied to a source. Without a citation in the document, no question arises. This is not a setting someone can change by accident, but is anchored in the analysis logic.

Do you get the same result on a second run?

Yes. auraIR works with temperature zero and a fixed model version. Same documents, same result. In capital markets communication this is not a gimmick but the precondition for an analysis to be auditable at all.

Demo

See auraIR with your own documents.

30 minutes, your last quarterly report, a finished question set. That is all the demo needs.

Book a demo

Under NDA and with synthetic sample data on request.