Canoe Intelligence

The removal test is decisive here. Strip the models and what remains is a connector that logs into portals and drops files into a folder, which is the least valuable part of the problem. The work being sold is classifying an unlabelled document into one of dozens of alternative investment document types and extracting the specific figures inside it, across formats that differ with every general partner.

The company states its models are purpose built for alternatives and trained on millions of real documents rather than adapted from a general purpose service. The honest counterweight is that portal connectivity at this breadth is engineering effort rather than model work, and it is a genuine part of the moat.

Extraction runs without a person in the loop by default and the published control is validation rather than review. The company describes up to a hundred field specific validation checks applied to each extracted data point according to the type of field it is, which is a more concrete quality architecture than most vendors put in public.

What is not described is the escalation path: what becomes of a value that fails those checks, who adjudicates it, and whether the client sees a confidence signal on figures arriving in their book of record. The managed service tier supplies human judgement but places it on the vendor's side of the relationship rather than the institution's, which is an inversion worth noticing.

What is published describes quality control rather than model transparency. Validation checks per field, isolated processing and annual external security auditing are all stated, and none of them measures how the models themselves perform. No extraction accuracy rate appears, no error rate by document type, no drift monitoring, no revalidation cadence and no model documentation of the sort an institution's own model risk function would request during onboarding.

That is the index norm rather than a failing peculiar to this vendor, since no vendor in the index has published a model documentation package, and the gap is sharper here only because the outputs land directly in institutional books and records.

Scale is quantified, dated and consistent across the company's own material and the acquisition announcement: over five hundred institutional clients representing more than eleven trillion dollars in assets under service, over one and a half million documents processed monthly, across more than forty four thousand funds and three thousand portals.

Reported clients include Blackstone and Hamilton Lane, which are also among its investors, alongside LGT Private Banking, Hermes GPE whose chief operating officer is quoted by name, and a Geneva investment firm. The strongest evidence is not a marketing claim. Several institutions with capital at risk backed the company, and Bloomberg signed a definitive agreement to acquire it in July 2026, which is the most consequential outside examination a private vendor can attract.

Two published claims sit in tension and the tension is the finding. On one side the company commits that client material stays inside its environment, is never used to train public models, and that training uses only authorised data under privacy preserving techniques. On the other, the product is sold on collective intelligence drawn from more than forty four thousand funds, which is a network effect that exists only if something learned from one client improves the service to another.

Both statements can hold under a careful definition and nothing published says where the line falls. The question sharpens with the acquisition, since the documents crossing this platform describe private fund performance and a market data parent has an evident interest in that material.

The sensitive material here is institutional rather than consumer, which moves the axis from customer financial data to partnership confidentiality, and the company addresses that directly. It publishes an account of how it handles confidentiality provisions in limited partnership agreements, stating that documents are processed solely for the client's benefit under data isolation and enforced access controls, that sensitive documents are reachable only inside the application to prevent transfer by email, and that data is encrypted in transit and at rest. Naming the legal instrument that actually governs the data, rather than reciting a generic policy, is what lifts this above the index norm. Retention periods and deletion on exit are not addressed.

A trust centre exists on a dedicated subdomain and the attestation is named rather than implied. The company states a service organisation control type two report renewed by annual external audit, and publishes a specific control account covering encryption in transit and at rest, application only access to sensitive documents, cloud infrastructure vulnerability scanning, network and web application penetration testing, a defined incident response plan, real time monitoring and requirements imposed on its own suppliers.

That is materially more than the single footer line common across this index. It stays below an A because one attestation is named and no international information security standard certification appears beside it.

No licence is held or required, which the convention here does not penalise, and the company positions itself plainly as a technology supplier to regulated institutions. The regulatory position it leaves unaddressed is the one its buyers carry. Registered advisers and fund servicers running on the platform are subject to books and records obligations, and the extracted figures feed valuation, performance reporting and investor statements produced under those rules.

Nothing located explains how an institution evidences a machine extracted figure to an examiner, or how the trail from a source document to a delivered data point is preserved and reproduced on request.

Forcing a lending fairness frame onto a document extraction vendor would be wrong and the real exposure is different. Extraction accuracy depends on document format, and format varies with the size, age, jurisdiction and language of the general partner producing it.

A model trained on the dominant reporting conventions will do best on large established managers and worst on emerging, non United States and non English speaking ones, which means an allocator's data quality is likely to be weakest exactly where its diligence burden is heaviest, on its newest and smallest commitments. Accuracy is presented as a single global posture with no breakdown by document type, manager size or region, so a buyer cannot see where the error concentrates.

Accuracy is the entire product and no published commitment attaches to it. A mis extracted commitment, capital call or valuation does not stay inside the platform. It flows into an institution's performance reporting, exposure analysis and investor statements, so an error has a short path to a restated number and a conversation with a client. Nothing published defines a service level, a warranty, a correction obligation or a remedy.

The managed service tier makes the question harder rather than easier, because work performed by vendor staff on an institution's behalf carries a standard of care that no public document describes. Validation checks lower the chance of an error without saying who owns one.

This is the most direct answer to the dependency question found anywhere in the index. The company puts the question in its own published material, asks whether it uses public general purpose assistants, answers no, names the widely used ones directly, and states that its models are selected, trained and deployed specifically for alternative investment documents, that client documents never leave its environment and that they are never used to train public models.

Most vendors leave a buyer to infer the stack from a privacy policy. Stating the negative case by name is checkable and falsifiable, which is what this axis exists to reward. The residual gap is small: the word selected implies base architectures that are never identified.

Integration is the product rather than a feature bolted to it. The platform connects to more than three thousand general partner and administrator portals and to client inboxes to retrieve documents at source, then delivers structured data outward into the systems that consume it.

The clearest evidence is a certified integration with a major market data provider's enterprise portfolio system, launched in early 2026, which fed permissioned private fund data into cross asset portfolio analysis and preceded that provider's decision to acquire the company outright.

Depth at both ends, retrieval from thousands of counterparties and delivery into institutional systems of record, is why the platform is described as connective tissue between general partners and their investors.

Delivery is cloud hosted in the vendor's own environment and the company makes that the centrepiece of its security story, stating that data and models remain inside its walls and that processing happens in isolated environments. The same architecture is what leaves this axis thin.

No hosting regions are stated, no residency commitment is offered and no single tenant or client hosted option is described, which matters for a client base that includes Liechtenstein, Swiss and United Kingdom institutions with their own data location expectations. Nothing located separates what could run inside a client environment from what cannot.

No rates, tiers or billing basis appear anywhere in the published material, and the route to a number is a demo request. The gap is wider than usual because the company sells at least two materially different shapes, a software subscription and a managed service in which its own staff run the client's operation, and nothing indicates how either is priced or how they relate.

A buyer can reason about scale from the published document and fund volumes but cannot infer a unit of charge from them, so two institutions of similar size have no way to sanity check the quotes they receive against each other.

Coverage spans the whole allocator chain rather than one buyer type. The published client base runs across large institutional investors, fund administrators and asset servicers, wealth managers, private banks and family offices, at over five hundred clients and more than eleven trillion dollars in assets under service.

Named deployments sit in several jurisdictions, including a Liechtenstein private banking group, a London private markets manager and a Geneva investment firm, and the company staffs an institutional investor role for Europe. The breadth is structural rather than incidental, since a product that sits between general partners and their investors has to work for whoever is on either side of that relationship.