FactSet

A substantial generative line sits on a business that predates it and would survive its removal entirely. Strip the agent, the pitchbook builder and the embeddable conversational interface and a financial data and analytics platform serving thousands of institutions remains untouched. The generative work is still a product line, not a feature.

Several separately named products do their work through inference and are sold to institutions rather than used internally, including an agent packaged specifically for embedding in a client's own stack.

Two real controls sit in the output path rather than around it. Generated output is explicitly labeled as generative throughout the interface. Every response carries in context source linking back to the underlying data, so a user can check any statement against its origin before acting on it. For an assistant serving analysts and bankers, verifiability at the point of use is the relevant control, and it is present.

Agentic interfaces, by contrast, are named without any description of what an agent may do unattended. No threshold, escalation route or review structure is published for the agentic path as distinct from the conversational one.

Citation is stated as universal rather than typical. Regardless of which of the five or more data sources answered a question, every response carries in context source linking for verification and identification of data lineage. Generative output is labeled as such, with linked references throughout the interface. That lets a user validate any individual claim, which is the practical model risk control for a research assistant. Nothing covers the models themselves.

There is no evaluation methodology, no accuracy or grounding figures and no benchmark, failure modes are undocumented, and there is no drift policy. Verifiability of each answer is offered in place of evidence about the system that produced it.

Scale is well documented, because the client and user counts appear in investor communications from a listed company. A figure in an investor release carries legal exposure that a marketing claim does not. None of it is evidence about the generative products specifically. No named institution is attached to an outcome from the agent or the pitchbook builder, and the productivity claims are stated in the abstract as hours reduced to minutes with no customer behind them.

Every stewardship question gets a direct, published answer. The pooled data question is answered explicitly and in the negative: user prompts and responses are isolated and protected and would never influence the responses given to another firm. The architecture is named rather than gestured at, using an open agent protocol and vector stores over the vendor's own commercial data and retrieval augmented generation over client confidential data.

Every data source feeding the system is enumerated, including the two that are optional and off by default. Model hosting policy is stated: all models are private, with no public endpoints. Supplier posture is stated too, with third party suppliers following the same approach and zero data retention and abuse monitoring opt out standard with the hosting platforms. Lifecycle terms are concrete, with twenty four month log retention and a sixty day purge.

Client confidential data is treated as a named category with rules attached: retrieval based access rather than training, isolation between firms, access restricted to a defined subset of employees for accuracy evaluation only, twenty four month log retention and a sixty day purge after termination. Those are commitments a buyer can hold the vendor to. No data processing agreement, subprocessor list or position on any specific financial privacy regime is published, and no residency commitment accompanies the retention terms.

Substantive governance and security documentation covers the generative layer, describing access restriction, isolation, private model hosting and purge terms. No certification or trust portal is public. A listed financial data provider of this size almost certainly holds SOC reports, and probably ISO 27001, so the gap says more about disclosure than about controls.

No financial services license attaches to the products. The company is a listed issuer, which is securities registration as a company rather than supervision of what it sells. One question is open: whether the group administers regulated benchmarks or indices anywhere. Benchmark administration is a supervised activity, so if the index business falls inside it, the regulatory picture changes.

A published generative governance policy exists and is detailed, but its content is data handling, transparency and labeling. Nothing addresses fairness testing, protected class analysis, or skew in what the agent surfaces or omits. The exposure is lower than for a credit or pricing model, since the user is a professional and the output is research support rather than a decision about an individual. It is not zero, because ranking and semantic search decide which companies and which themes a banker sees at all.

No allocation of responsibility for a wrong generated answer and no recourse route published. Source linking is offered as the mitigation, which places the burden of verification entirely on the user, and that is a defensible design for a professional tool but it is not recourse. Nothing states what happens when a sourced summary misrepresents the document it links to.

FactSet names its data infrastructure partner and describes its model deployment in detail, yet never names a model. Models are private instances in its own cloud or private endpoints of commercial providers, with no public endpoints and zero data retention at the host. Which foundational models those are, and which providers the private endpoints belong to, is never stated. A buyer knows exactly how the models are run and not at all whose they are. Of the two halves, the deployment terms are the one that governs data exposure.

Integration is a named strength rather than a claim. The work is wired deeply into the spreadsheet and presentation tools where it actually lands, and the agent is exposed as an interface for embedding inside a client's own stack. Packaged data feeds are built to power a client's own workflows, and the open agent protocol reaches client defined data repositories. The named integrations are productivity tools and the vendor's own protocol layer. No third party system of record, such as an order management or portfolio management platform, is named.

The deployment posture for the generative layer is described precisely. All models used are private, running either as private instances inside the vendor's own cloud or against private endpoints of commercial providers, and public endpoints are ruled out explicitly. Zero data retention and abuse monitoring opt out are stated as standard with the platforms hosting those models.

Data lifecycle terms are concrete: logs are retained twenty four months for active relationships, and confidential data is purged within sixty days of termination. Geographic residency is not addressed at all, a real gap for a client base spanning several supervisory regimes.

No price is published for any product, which is the enterprise subscription norm. One partial offset exists. As a listed company FactSet publishes annual subscription value and client counts, from which an average revenue per client can be derived. That is more than a private vendor's silence yields, though it says nothing about what a specific configuration costs.

Coverage is wide by any measure. Buy side and sell side institutions, wealth managers, private equity firms and corporates make up more than 8,200 clients and over 218,000 individual users globally. The generative products address distinct workflows across that base: banker research and pitch production, portfolio analysis, general partner monitoring, and an embeddable agent for whatever a client builds itself. Client and user counts come from investor disclosures by a listed company rather than from marketing material.