Aiera

Graded on a feasibility argument rather than a features list. The product is live transcription at under half a second of delay across tens of thousands of events a year, plus generative summaries and sentiment on top, and no arrangement of people produces that. The company describes a finance trained speech model and an ensemble of financially trained models for summarisation and sentiment.

The honest counterweight is stated rather than smoothed: event sourcing, licensing and human review are real assets that are not model work, and about a quarter of the content is exclusive because of the sourcing operation, not the models.

Human review is built into the pipeline rather than offered as an option. The company describes owning the workflow from event sourcing through automated capture to human review, and sells both a real time machine transcript and a human edited transcript tier, which is a published two speed accuracy model. Outputs are research inputs consumed by analysts rather than decisions executed on a client's behalf, so the oversight burden is lighter than for vendors acting in the market. Nothing describes review of the generative summaries themselves, only of transcripts.

Rare on this axis: the company has published its evaluation methodology in engineering literature, describing how it built a labelled dataset from its own transcript archive, which model it used to generate reference insights, and a comparison of overlap based and embedding based scoring metrics before selecting a production model. Publishing how a generative output was evaluated, including the limits of the metrics, is more than any model documentation claim in this index. It stops short of an A because no production error rates, drift monitoring or ongoing revalidation cadence is published.

Volume is quantified and internally consistent across sources: over forty five thousand events a year, thirteen thousand companies, a hundred plus macroeconomic entities, split into roughly thirty one thousand corporate events, twelve thousand conference presentations and three thousand macro and regulatory briefings, with transcription accuracy stated at ninety nine percent.

It is also the subject of a business school teaching case, which is an unusually independent examination of a company this size. What is missing for an A is a named institution with a measured outcome, since customers are described in categories rather than by name.

The stewardship story here is content rights rather than model training, and it is more developed than most: entitlement aware access so users see only what their firm has licensed, transparent attribution and source traceability, standardised legal access frameworks for proprietary content, consumption metrics intended to preserve the value of that content, and validation work conducted with the content providers themselves. That addresses the question most research assistants leave open, which is whose material the model is answering from and whether the rights holder agreed.

The data class here is market and research content rather than consumer financial data, so the privacy exposure that shapes this axis elsewhere in the index is largely absent, and the axis is graded on what is published rather than penalised for the category. Nothing was located on handling of customer research activity, watchlist contents or query history, which is the sensitive material in this product, since what an institution is researching is itself signal.

Searched the product, platform and company pages and partner announcements for an enumerated certification, an attestation report or a trust centre and found none. Governance language on the site addresses content entitlements and auditability rather than information security controls. For a supplier whose clients are institutions with vendor security review processes, the absence is conspicuous rather than neutral.

No licence is held or claimed and none is required for a research data supplier, which the convention here does not penalise. What is graded is the clarity of the regulatory position, and the word compliant is used throughout in the sense of content entitlements rather than securities regulation.

Research distribution to institutional clients sits near research unbundling and inducement rules in Europe and near supervised records obligations in the United States, and no source located addresses either.

The bias question for a transcription and sentiment vendor is not lending fairness, and forcing that frame would be wrong. It is accent and dialect. Coverage spans thirteen thousand companies worldwide, where a large share of executives speak English as a second language, and speech recognition error rates are well known to vary by accent, while sentiment scoring reads tone through the same channel.

A transcription error or a sentiment score on a mis heard sentence becomes an input to somebody's investment model. No per accent or per region accuracy is published, and the ninety nine percent figure is a single global number.

One phrase in the marketing does work the terms do not: the human edited transcripts are described as carrying guaranteed accuracy, and no published warranty, service level or remedy defines what that guarantee is worth if a transcript is wrong. Errors here have a direct path to loss, since a mis transcribed number or a summary that inverts guidance can be traded on. Attribution and auditability make an error traceable after the fact, which is genuinely useful, but tracing is not recourse.

More candid than most about running on other people's models. The company states an ensemble approach that uses the best of what is available in the market alongside what it builds internally, its engineering write up names the specific third party models used to build its evaluation set and to serve summarisation, and it publishes connector libraries for two major model providers and one cloud platform. That is the dependency stated openly rather than behind proprietary. It falls short of an A because there is no maintained disclosure of which model serves which production function today, so a buyer cannot tell what changes when a provider deprecates a version.

Integration is the distribution strategy and the named surfaces are specific: research authoring platforms used by sell side publishers, a regional business intelligence provider licensing results and transcripts for ten thousand listed companies, enterprise interfaces for downstream data providers, embeddable interface components, a mobile application, and a model context protocol server with connector libraries for the major model and cloud providers so the content can be consumed inside a client's own artificial intelligence stack. Meeting institutions inside the tools they already run, rather than asking them to visit another destination, is the pattern this axis exists to reward.

Delivery is cloud hosted with interface and protocol access, and one older account describes offering trained speech and language models for secure use against a client's own private datasets, which would be a meaningful private deployment option if it is still offered. Nothing current states hosting regions, residency commitments or a single tenant option, and no source located distinguishes what runs in the vendor's environment from what can run in the client's.

No rates are published. The billing basis is visible only through a third party case study, which describes seat subscriptions for the front end and interface licensing for data delivery, and the company site mentions usage and consumption metrics designed to preserve content value. A buyer can infer the shape of a deal but cannot price one, and the disclosure sits outside the vendor's own material, which is why this stays below the peers that publish a tier ladder.

It sells across both sides of the market, asset managers and hedge funds on the buy side and banks and brokerages on the sell side, plus corporate investor relations teams and downstream data providers who license the transcripts. Its own advisory group is described as senior leaders from asset managers, long only firms and hedge funds. Coverage is global across thirteen thousand listed companies. Held at B for the same reason as evidence, since no institution is named.