Juniper Square

A ledger built over more than a decade carries this business, and the models are a recent layer on top. GPX is a fund operating system covering fund accounting, general ledgers and payments, along with the investor portal and compliance workflows. Strip the models and all of it keeps working for 2,300 firms. The AI line is real rather than decorative: an oversight agent and an investor relations system would not exist without models. The services half of the business, fund administration over more than $300 billion, is people rather than models.

Two properties do the work. Permission inheritance is named and placed: models acting on behalf of a user can reach only the data that user is entitled to, under the same authorization model as the user facing software. That inheritance is stated to travel outward through the protocol interface as well. The current scope is also stated as a limit rather than a capability.

The vendor says plainly that its present systems do not access sensitive client data and that any future system which does will apply least privilege. By design, the oversight product works ahead of a human, catching errors in an administrator package before it leaves the chief financial officer's desk. No approval gate is described for any action the agent takes, and nothing states what happens when one of the checks fails. The headless interface is read only today, with writes on the roadmap, which is an absence of capability rather than a control over it.

The published material is easy to mistake for model risk disclosure. It is security control over models: prompts hardened against attack, third party model vendors bound by contract, and data protection controls around the models tested annually inside the attestation scope. Those are good controls, and they are not model risk management. Nothing published gives an accuracy or error rate, a validation approach or a benchmark, and there is no monitoring regime or drift policy for any model.

The closest thing to substantive transparency is the oversight agent's library of more than 150 named accuracy and consistency checks, which a customer can inspect and extend with their own. That tells a validator what the system tests for without telling them how often it is right.

A named client is quoted in a named role: the chief financial officer of an investment partners firm reports greater assurance and audit readiness on the oversight agent. Operational scale figures are very large across entities, investors, positions and capital. The quantified claim, roughly 80 percent off financial statement review time, is the vendor's own and is not attached to the named client. One provenance detail cuts in the vendor's favor.

The oversight agent was built first for its own administration team, which oversees more than $300 billion, before being sold externally. The vendor was its own first customer at real scale, a stronger form of evidence than a pilot.

Customer data does not train the external models, and a mechanism backs that. Contractual agreements with the external model vendors preclude them from using any of Juniper Square's data to train their underlying models. Vendor security reviews are performed before engaging any supplier with embedded model functionality. Access is bounded by inheritance, since a model acting for a user reaches only what that user may reach.

Data shared with models is stated to fall under the same controls as any other customer data, and those controls are inside the annually tested attestation scope, which ties the claim to an audit rather than a promise. The vendor also states the current limit: its present systems do not touch sensitive client data. Data sources are not enumerated, no retention or deletion terms specific to model use are given, and nothing states whether prompts are isolated between tenants.

The privacy position has real components rather than a policy link alone. Compliance is stated against both GDPR and the California statute, and a privacy policy is published. Encryption is specified at rest and in transit, and two factor authentication covers staff and investor users. An authorization model governs what any given user or agent can reach, and changes to sensitive investor records trigger alerts. A dedicated page answers legal and privacy questions about the AI line specifically. No subprocessor list, data processing addendum or residency statement is published.

The security posture is strong and precisely stated. Juniper Square is assessed annually for SOC 1 Type II and SOC 2 Type II, against the criteria of the national accountancy body. Independent penetration testing is routine. Data is encrypted at 256 bits at rest and with TLS in transit. Two factor authentication covers both staff and portal investors, a password policy is enforced, and an audit log alerts on changes to sensitive investor information.

Holding the SOC 1 report is the correct choice for a business whose output lands in audited fund accounts. What is missing is a standing trust portal, downloadable reports and a bridge letter covering the gap between audit periods.

Fund administration in this market is largely unlicensed as an activity, so the absence of a license reflects the category rather than a choice by the vendor. There is no supervised test of the product and no program enrollment either. The closest thing to external discipline it carries is its audit attestation.

Consumer protection exposure is low. The work is fund accounting review, investor reporting and fundraising workflow, not any decision about an individual's access to credit or a financial service. Nothing is published on fairness testing, performance variation or governance of model outputs. What remains is narrower but real. An investor relations system that surfaces which limited partners to prioritize is ranking people and institutions, and nothing describes what drives that ranking.

Selling a control creates a liability question, and nothing published answers it. The oversight agent is sold explicitly to reduce fiduciary risk and to prevent restatements, auditor rework and difficult conversations with investors. If it runs its checks and misses an error that later forces a restatement, nothing says where responsibility sits. It could rest with the general partner who bought the control, the administrator who produced the figures or the vendor whose agent passed them. There is no published allocation of responsibility and no route for any affected party.

The supply chain is disclosed in outline. Juniper Square states that its models come from third party suppliers rather than being built in house, and it names a hosting platform and a model vendor by way of example. Those suppliers are bound by contractual controls, including a bar on training over its data, and are security reviewed before engagement. The limit is the difference between exemplifying and enumerating.

Phrases like providers such as one named cloud, and vendors for example one named model company, tell a buyer the shape of the supply chain. They do not say which model serves which feature, at what version, or whether the list is complete.

Integration runs in both directions, and the outward half is the more interesting. Inward, Juniper Square does not integrate with the system of record. It is the system of record, running fund accounting, the general ledger, payments and the investor portal. The oversight agent also ingests deliverables from any rival administrator with no integration and no data onboarding, working from whatever file that administrator produced.

Outward, the headless interface exposes every platform capability to any client speaking MCP, with three major assistants named specifically. That covers fundraising, investor relations, fund accounting, ledgers, payments and compliance, and the vendor states that the same data, the same permissions and the same audit trail travel with it.

The toolset is described as purpose built for how the finance, investor relations and compliance functions actually work, not a one to one wrapper over every interface endpoint, which is a meaningful design claim.

Residency is the gap. The platform holds the books and investor records of funds raised across multiple jurisdictions, and nothing is published on region, residency, tenancy or where a given firm's data sits at rest. The cloud provider is named, and redundancy across data centers and availability zones is described for uptime purposes. Vendors of this size usually publish a residency statement for European investors, so one may exist that is not reflected here.

A buyer knows how the product is sold without knowing what it costs. The oversight agent is described as purchasable inside the platform today and available as an add on to firms that use a different administrator. No price, tier or unit is published.

Deep in one segment rather than broad across several. The buyer is the private markets general partner, addressed at 2,300 firms worldwide. Inside each firm, depth runs across the roles served: fundraising and investor relations teams, chief financial officers and controllers, and compliance. The limited partners on the other side of the portal number 750,000 accounts. Scale is large, at 45,000 entities and $1 trillion in active capital. This is fund operations for private capital and nothing else, with no banking, insurance, payments or public markets footprint.