Broadridge

Strip every model out and the core businesses keep running, because they did for decades. Investor communications, proxy and governance processing all continue. So do the post trade infrastructure underpinning more than $10 trillion of daily trading, the wealth platform and the distributed ledger repo business. The AI work is real, in production and well evidenced.

The exception is the group's corporate bond venue, which is itself described as an AI powered trading platform built around patented AI. On its own, that business would likely depend on its models far more than the group does. This profile covers it as a subsidiary line inside the parent, not as a separate vendor.

The bond trading agents act, and the company publishes the architecture that bounds them. They can generate alerts, create trade tickets and select dealers. They can also launch requests for quote and accept prices to execute automatically.

Against that the company publishes four named guardrails with their positions in the execution path stated: human in the loop approvals, policy driven limits on trade size and scope, explainability before every action, and full auditability of every action. Placing explainability before the action, not after it, makes the constraint operative rather than forensic. Limits on size and scope are a hard boundary on what the agent may do, not a workflow description.

The framing throughout is delegating selected tasks while keeping the trader in control. Two points are unresolved. No default configuration or minimum limit is published. And because parameters are trader defined, it is not fully clear whether human approval applies to every path or is one guardrail among several that a firm can configure.

The models themselves go undocumented. There is no accuracy or error rate, validation methodology or backtesting. Nor is there drift monitoring, versioning or external assessment, across a generative application answering bond questions and agents that can execute trades. The published explainability and auditability are controls on action, not evidence about model behavior. Explaining what an agent is about to do is not the same as publishing how well the underlying model performs. This vendor does the first thoroughly and the second not at all.

The deployment evidence is strong, but no result is attached to a named institution. The agentic capability is stated to be live in production, not just announced. It was refined through live deployments inside the company's own outsourcing operation across more than forty clients since 2024. Three global dealers are named as participants on the bond venue, and the repo platform's throughput is published with a period attached. Three gaps remain.

The one efficiency figure is a ceiling, not a measurement, and attaches to no institution. The named dealers are venue participants, not referenced outcomes. The throughput figures are the company's own volumes, not a customer's result.

The vendor raises the pooling question itself and then does not answer it. It states that the agentic capabilities were refined through live deployments inside its own outsourcing operation across more than forty clients since 2024. Improvement through exposure to client operations is the advertised reason the product is production ready. Nothing states whose data produced that refinement, or whether one client's exceptions and resolutions inform the agents another client receives.

Nor does anything state what the shared data ontology holds across clients, or what is excluded. The bond application is separately described as drawing on curated data and analytical models, without saying whose.

At product level nothing is published. There is no retention schedule or deletion terms, and no subprocessor list or tenant separation statement, for a group holding shareholder records, trading activity and wealth account data. One architectural fact bears on it: the choice between the vendor operating the process and the platform running inside the client's own infrastructure. That choice determines custody of the data and is presented as a commercial option, not as a data handling position.

Broadridge publishes no certification or attestation report, and no audit period or trust portal. That is an absence of disclosure, not evidence of weak controls. A listed processor underpinning trillions of dollars of daily settlement activity is examined and attested extensively. Such material sits in client due diligence channels and regulatory filings rather than on product pages. Among large processors, the larger the firm, the less findable its security evidence.

The firm holds real regulatory standing and discloses it in its own material. The corporate bond trading venue operates through a named group subsidiary acting as its broker dealer. That is a registered and supervised status, carrying examination, capital requirements and conduct obligations. It also carries the possibility of sanction against the firm itself, not only against its clients.

As a directly supervised firm that can be sanctioned, it sits inside the regulatory perimeter where software vendors do not, and the status is public and independently verifiable. Scope is stated precisely: the license sits with the subsidiary operating the venue, not with every part of a group whose other businesses are technology and outsourcing. Operating both the venue and the agents that trade on it gives the firm a regulatory position that pure software suppliers of trading agents lack.

Governance goes unaddressed: no published position, no fairness testing, no named framework and no statement on any AI regulation. Fairness exposure is lower than for vendors shipping consumer credit or collections capability, because the agents operate between institutional counterparties and decide nothing about a retail customer. That lowers the exposure without adding any disclosure.

The material does not reach two adjacent questions: whether agents acting across a venue could produce correlated behavior among participants, and what governance applies to agents a customer configures rather than ones the vendor builds.

The guardrails are published and the liability allocation is not, and what the agents can do makes that gap sharp. An agent that accepts a price and executes automatically can produce a bad fill, a mispriced trade or an unintended position. Nothing published states who bears that loss or how an erroneous execution is unwound. Nor does it say how responsibility divides between the venue, the agent platform, the firm that set the parameters and the trader who delegated the task. The published limits on size and scope are a control on exposure, not an answer to who pays when a permitted action turns out wrong.

No model, provider or version is named for the generative bond application or for the agents in production. Two dependencies at adjacent layers are named, and neither is a model. One is a strategic investment in DeepSee, whose capability is being embedded into post trade operations. The other is an infrastructure and resiliency agreement with Kyndryl. Both say more than silence would, and neither identifies what generates the output. The omission matters more because the agents transact rather than advise. A buyer assessing model risk on an agent that executes trades cannot establish what is generating the decision to execute.

In several places Broadridge is not integrated with the market infrastructure; it is the market infrastructure. Post trade platforms underpin daily trading measured in trillions, and the repo platform processed $7.2 trillion in a single month. The outsourcing arm runs client operations end to end rather than supplying software for the client to run. The AI work rests on what the company calls a completed financial services data ontology.

That is a foundation of the same kind as a published integration schema or a named industry reference architecture, and it is what lets agents act across previously separate operational domains. A named infrastructure partner, Kyndryl, is disclosed for resiliency and platform capability.

Deployment choice is answered clearly and residency is not addressed. The two named paths are materially different, not nominal variants. Either the vendor operates the process on the client's behalf, or the agentic platform is installed into the client's own infrastructure, and that choice determines who holds the data. Nothing published states region availability, residency commitments or how data is kept inside a jurisdiction. The group operates across twenty one countries, in markets with divergent rules on where trading and investor records may sit.

No pricing is published. Two distinct commercial models are named. In a full managed service the company runs operations end to end; in a standalone deployment the platform goes into the firm's own infrastructure. The structure is named and the price is not. A cost reduction ceiling is published in place of a cost, which tells a buyer what it might save and nothing about what it would pay.

The group serves sell side dealers, buy side institutions and wealth managers. It also serves corporate issuers and, through proxy and investor communications, retail shareholders at population scale. That reach spans twenty one countries, with platforms underpinning more than $10 trillion of daily securities trading and over seven billion communications a year.

The bond venue alone reports more than forty liquidity providers and over one hundred buy side institutions, with three global dealers named among participants. Buyer types span trading and post trade operations, wealth and issuer services, and outsourced operations. Those are separate procurement centers with separate budgets, not one market described broadly.