Quartr

The core sellable asset is model output, and the company still sells it as data rather than as analysis. Real time transcription of live earnings calls across more than 13,000 companies in over 65 markets cannot be produced by people at that latency or scale. Strip the models and the transcripts disappear, leaving audio files, slides and filings gathered from investor relations pages. The query layer over that material, with answers linked back to source, is a further model layer on top.

The limit is that the product sold is the structured content itself, delivered through an interface for others to build inference on. The company positions itself explicitly as infrastructure for other systems, not as the system that reasons.

One control is named, and it is the right one for a research product. Answers generated in the research platform link back to the original source material, so a user can verify the underlying statement rather than trust the summary. Provenance surfaces by default at the point the answer is read, not on request. What is missing is confidence: nothing tells a user how reliable a given answer is, or when the system is working from a poor transcript. No abstention behavior is described for questions the content cannot answer. No threshold or review step is specified anywhere in the pipeline that turns live audio into a queryable transcript.

For a company positioning itself as infrastructure other systems depend on, the most important quality measure is unpublished. Transcription accuracy is what a customer is actually buying when it buys real time transcripts. No word error rate or benchmark is published for any language or market, and no sampling methodology or human review process is described. Timeliness and reliability are asserted as best in class without a latency figure or an availability measure. No evaluation of the query layer's answer quality appears either, and no retraining or monitoring cadence is described. No ISO 42001 certification is held.

The published operating metrics say more than a logo count would, and no customer is named. Net revenue retention of roughly 120 percent is a specific figure that shows whether existing customers expand rather than merely renew. Growth is stated as triple digit, and more than 700 institutions and technology companies are claimed as users. Independent parties have money at stake across successive rounds.

Altos Ventures led repeatedly and became the largest shareholder, and a Nordic bank joined the most recent round: a financial institution underwriting a vendor to financial institutions. Every customer reference in the material is anonymous, described only as a large client or a research platform, and no outcome attributable to the product is measured anywhere.

Concern about pooled client data is weaker for this product, because the training and serving material is public company disclosure, not client data. No safety practice or output screening is described, and no red teaming or data boundary. What remains unaddressed is the client side. The platform observes which companies, events and topics each institution researches and when, a signal about investment attention that competitors would value. Nothing states whether that usage data is retained, aggregated, used to shape the product or exposed in any form.

No privacy policy detail or processing agreement is published, and there is no public subprocessor list or retention schedule. The mitigating fact is structural: the material being processed is what public companies publish about themselves, so the content is disclosure rather than personal or client data, and the privacy surface is narrow. What is not addressed is the client side.

A research platform records which companies, events and topics an investment professional searches, and that query history is commercially sensitive in its own right. Nothing states whether it is retained, for how long, or whether it informs anything.

The buyers make the security gap matter. More than 700 financial institutions use the products, including hedge funds and asset managers whose own vendor risk processes routinely require an attestation before onboarding, and a European bank has taken an equity position. Yet no certification, attestation or trust portal is public. The content itself is public company disclosure, not client portfolio data, so the confidentiality exposure is lower than for vendors holding client data. That lowers the stakes and does not answer the question a procurement team will ask.

The company is a data and software provider and has no license, registration or supervisory standing of its own, which is ordinary for a business of this shape. Its obligations are contractual and reputational, not supervisory. Still, the material it redistributes sits close to regulated disclosure: the timing and accuracy of what a company said on an earnings call can matter to market abuse and fair disclosure questions at the institutions consuming it. No regulator engagement, accreditation or independent assurance over its transcription or distribution practice is published.

The exposure specific to this product is linguistic and geographic, not demographic, and it is substantial given stated coverage of more than 65 markets. Automatic transcription performs unevenly across accents, languages and audio quality. A management team speaking English as a second language on a poor line is transcribed less accurately than a native speaker on a clear one.

Nothing published examines whether transcript quality, and so how well the query layer answers questions, varies systematically by market, language or company size. No fairness evaluation, governance program or model assessment material is published.

No liability position, accuracy warranty or correction route is published. The path from error to consequence is short in this product: a mistranscribed figure or a misattributed sentence in an earnings call becomes a fact an analyst acts on, and where the data is delivered through the interface it propagates into other vendors' platforms and their customers' decisions, so an error can travel two steps from the company that made it. Nothing states what a customer may rely on, how a transcript error is reported or corrected once distributed, or whether corrections are pushed to downstream consumers of the feed.

No speech recognition provider, language model provider, family or version is named, and nothing states whether the transcription engine is built in house or licensed. The distinction matters because transcription is the product itself, not a supporting feature. A customer buying transcripts at scale cannot tell whose model produced them, where the audio was processed, or whether the answer would change if the underlying supplier did. Nothing describes the hosting arrangement for either the transcription pipeline or the query layer.

The integration story is clear on architecture and silent on named counterparties. A single interface carries consistent identifiers across audio, transcripts, filings and slide presentations, so a developer joining a transcript to a filing does not have to reconcile identifiers between sources. Entities are structured and linked across time and markets. They cover companies, events and people, plus products, topics and reported metrics.

That is a real data architecture commitment, not a connectivity claim, and it is the substance of what interface customers buy. No client platform or terminal is named as integrated, and no data warehouse or research system. No integration count is published, and the testimonials that describe the interface in use identify nobody.

Delivery is cloud based through a hosted research platform, a data interface and a mobile application, and nothing further about where any of it runs is published. No hosting region, residency commitment or tenancy model appears.

The company operates from Stockholm with offices in New York and Dublin and serves institutions across more than 65 markets, so European clients subject to data transfer rules have no published statement about where their usage and query data is processed, even though the underlying content is public.

One tier is transparent and the two that institutions buy are not. The consumer mobile application is free and carries the same event material, a verifiable commitment. Neither the professional research platform nor the data interface publishes a price, tier structure or billing basis. For an interface product the basis matters a great deal: charging by request, by company covered, by market or by seat produces very different costs for the same use. Whether the transcript archive and the live feed are licensed separately is not published.

Breadth runs across buyer types, geography and delivery model at once. More than 700 financial institutions and technology companies are stated as users. They span hedge funds, asset managers, equity researchers and investor relations teams: four different workflows on different sides of the same disclosure event. Coverage extends to more than 13,000 public companies across over 65 markets rather than one exchange, in a market where United States coverage usually dominates.

The two channel structure widens it further. The data interface serves other research platforms as customers in their own right, so the platform reaches institutions indirectly through products it does not own.