Primer

Research, modeling, monitoring and writing all run through the agent, and the company's own engineering is the harness around a third party model: what it reads, which tools it can use and what it has been taught about finance. Remove the inference and no product is left. The vendor's own figures make the point, since the same base model scores far lower outside the harness than inside it.

The agent is designed to run on its own: recurring research tasks monitor a coverage list, update work and flag what matters without a prompt. The privacy policy states that outputs are meant to support human research and that the service is not designed to make decisions about people without human review. Beyond that, no approval step, confidence threshold, log of agent actions or limit on what scheduled tasks may do is published, and that matters most when research runs unattended.

The evaluations state their own limits. The vendor names the model underneath and reports its score on the BigFinanceBench leaderboard alone, 44.3 percent, against 79.1 percent inside the harness, which isolates what the product itself adds.

It also discloses that its own score comes from the benchmark's public questions after excluding seven it judged to have wrong reference answers, while the leaderboard covers all 928, and that its rerun of the Wall Street Prep modeling test was scored by AI judges rather than by Wall Street Prep. Those caveats make the comparisons less clean, but they are stated openly. No ongoing accuracy monitoring or validation support for customers is published.

No customer is named. Testimonials are attributed to unnamed analysts at hedge funds. The headline result is a backtest run and reported by the vendor: trading on the strongest earnings preview calls produced 34 percent alpha against the S&P 500 and a Sharpe ratio of 1.42, across 2,727 calls over 203 trading days at a 53.6 percent hit rate. It is specific and could be checked in principle, but it is a simulation, not a client outcome.

Customer files, prompts and outputs are not used to train third party foundation models, and agreements with model providers bar them from training on customer content. The homepage goes further and says customer data is never used to train models, but the privacy policy is narrower: it reserves the right to use service content, usage data and feedback to evaluate and improve Primer itself, including through human review for quality, support and safety. The boundary with outside model providers is firm, while the vendor's own improvement work can draw on customer material.

A full privacy policy sets out what is collected, the legal bases, who receives it, international transfers and data subject rights. For business customers the roles are clear: the customer is the controller for content it submits, and that processing runs under a data processing agreement. Retention is described by purpose rather than by fixed periods. Nothing addresses the Gramm Leach Bliley Act, which matters less here than for a lending or banking product, since the inputs are mostly market data and research.

The company states SOC 2 compliance and says its security controls and materials are ready for institutional review. The report type, the auditor and a trust portal are not named, so the report itself is the document to request. The privacy policy otherwise describes technical and organizational measures in general terms.

No license or authorization is held. The earnings previews produce directional calls a fund can trade on, which brings the output close to investment research. Nothing published addresses how that output sits under UK or US rules on investment research, or how a fund should treat it in its own compliance records.

No governance framework or bias evaluation is published. The risk that matters here is systematic tilt rather than fairness to individuals: an agent that learns a team's preferences can reinforce them, and directional calls can lean one way across a sector. How the agent's judgments are tested for that is not described.

No liability position, accuracy warranty or recourse is published. The marketing ties the output to trading returns, and nothing says who bears a loss when a preview or a model is wrong.

The base model is named: the product runs on GPT-5.5, disclosed alongside its standalone benchmark score. The privacy policy adds that model, retrieval, evaluation and observability providers process data, without naming them. No second provider, failover or customer choice of model is described.

Data access comes built in. The product includes filings, transcripts, consensus estimates, market data and alternative data, and also works with connected client sources and uploaded files, so analysts do not have to permission sources one by one. Models are built on the vendor's own modeling infrastructure and exported to Excel. Everything runs in the browser with no install, and it is not a system of record for portfolio, order or risk data.

The service runs in the browser as hosted software. Data may be processed and stored in the United Kingdom, the European Economic Area, the United States and other countries where the company or its providers operate, with standard contractual clauses and similar safeguards where required. No customer choice of region, dedicated deployment or on premises option is published.

No price is published. The pricing link leads to the evaluations page, and access runs through a demo request, which points to quoted pricing. The privacy policy mentions trials, paid plans and credit usage, which suggests a subscription with metered credits, but no figure or unit appears anywhere on the site.

The product is built for one kind of buyer: equity investors doing fundamental research. Testimonials come from analysts at single manager and multi manager hedge funds, unnamed. Banks, insurers, credit investors and wealth managers are not addressed, and no customer count or geography is published. The focus is deliberate, and it is narrow.