Google's bank agent can read FactSet. The licence, not the model, decides how much.

Six of the seven data sources named at launch are commercial licences: FactSet, Moody's, S&P Global, MSCI, PitchBook and Dun & Bradstreet ^[5]. The seventh, SEC EDGAR, is a free public filings repository ^[6]. That ratio is the dependency graph an operator should draw before drawing the architecture diagram, because entitlement-aware access means the connector checks what the requesting user is permitted to see on each call ^[4]. The agent's reach is therefore a function of whose seat it is borrowing.

Crowdfund Insider's account of the announcement says permissioned data stays restricted and licensed information remains controlled ^[4]. It does not say how any of those six vendor agreements treat an agent as the reader, and it carries no price and no general availability date ^[15]. Those are the two documents a capital markets COO actually needs, and neither is Google's to publish.

The skills are the part with a half-life. More than 50 of them encode report formatting, targeted retrieval and house research method, and they can be lifted out of the managed agent into agents a customer builds ^[2][3]. Deutsche Bank supplied the regulatory-requirement and data-protection view during design ^[7], so a firm that runs the managed agent unmodified inherits one large European lender's reading of what a defensible research step looks like. CME Group also appears in the preview cohort ^[9], which puts an exchange operator alongside a corporate bank in the same early-user list.

The governance description is specific enough to test in a pilot: a central control plane for policy enforcement, isolated customer data, source citations and data snapshots for audit, explainability of agent reasoning, and a statement that customer content is not used to train Google's foundation models ^[12]. The training exclusion is the clause legal reads first; the snapshots are the clause that outlives the pilot.

Set against that, the workflow claims are vendor-shaped and unmeasured. Bond risk assessments completed in minutes with hedging suggestions, mispricing detection for credit analysis, and faster pitch preparation are all asserted ^[13], with no error rate, no benchmark and no baseline in the material ^[15]. Seven financial-sector institutions are named in total, and five of them are users of the wider Gemini Enterprise platform rather than this product ^[10][11]. Reference density is doing some work here that measurement is not.

Preview status also means the interesting piece of paper is not the launch note ^[14]. It is the next licence amendment a data vendor sends a client whose desk wants an agent reading on its behalf ^[5][15]. Until that arrives, the cheapest agent in a bank is the one pointed at EDGAR.