Cognaize

Take the models away and nothing is left. The platform is built from neuro symbolic agents, finance specific language models trained on more than 1.3 million financial documents, and a distinct class of verifier models, with deep learning described as the foundation throughout. Its founding argument is that financial decisions currently rest on the small structured fraction of available information while the unstructured majority goes unused. Given the diversity of credit agreements, filings and trustee reports, closing that gap is a modeling problem with no rules based equivalent.

Human involvement is the architecture rather than a safeguard placed around it, and the company is named for the concept. Hybrid intelligence combines the models with a purpose built interface enabling seamless engagement of human experts in the loop, so a person reviews and refines rather than receiving finished output, and the platform improves with each document a human corrects.

A separate class of verifier models sits between extraction and delivery, checking values against the institution's own definitions before a human sees them. Explainability and auditability are stated as properties the solutions enhance. Contemporaneous coverage described the proposition precisely as building a better language model for finance, one that keeps humans in the loop.

A separate class of models exists purely to check the first, described as specialised systems that validate and ensure the accuracy of extracted financial data according to institution specific definitions and rules, which is exactly the right control because the same figure means different things at different institutions and a generic accuracy check would miss that.

Around it sit human experts in the loop as a second verification layer, stated explainability and auditability, per institution ontologies governing what automation produces, and a neuro symbolic design where the symbolic component enforces rules the statistical component cannot be relied on to respect. That is verification at three levels: model, rule and person.

The customer set is strong for a company at this stage and specific enough to check. It includes two of the three biggest credit rating agencies, alongside large insurance companies, global banks and financial services businesses. Rating agencies are the most document intensive analytical institutions in finance and among the most conservative buyers of technology, so their adoption is a strong signal. An $18 million Series A closed in 2023, led by a specialist venture firm.

One investor described the company as among the first to deliver repeatable and measurable value through AI in the financial industry. The training set of 1.3 million documents and four offices across New York, Frankfurt, Los Angeles and Armenia support the picture. Individual institutions are not named.

Containment comes from deployment rather than from policy. Models run on premise or in a private cloud, so one institution's credit agreements and filings are never transmitted to the vendor and cannot inform anything served to another. The base models were trained on 1.3 million documents assembled independently of any particular customer. Each deployment is then shaped by that customer's own proprietary data, models and financial experts, so what improves through use improves inside the institution that generated it.

The deployment architecture is the privacy answer, and the company states its reasoning explicitly. Financial services firms are custodians of customer data subject to strict regulatory guidelines and compliance rules, so its models are downsized and fine tuned to run on premise or in a private cloud environment. The documents being processed include loan applications and credit agreements containing borrower information, and they need never leave the institution's own estate. For a document processing vendor, that is the strongest structural position available. The hosted alternative has no published data processing terms, retention schedule or subprocessor list.

There is no public attestation, certification, trust center or enumerated framework. Two of the three largest rating agencies and multiple global banks have completed vendor assessment on this company, which is diligence at the most demanding standard available. The on premise option also materially reduces what a customer takes on trust, since the software runs inside their own controls. Publishing the control set would still be the natural complement to a security led architecture.

Regulatory requirements are invoked as the reason for the deployment model, with financial services firms described as custodians of customer data subject to strict guidelines and compliance rules, and no regulator, statute or instrument is named anywhere.

That gap is worth noting given the customer base, because credit rating agencies are themselves directly supervised entities with prescribed methodologies and record keeping obligations, and a platform extracting the inputs to their analysis operates inside that framework.

No individual is assessed. The risk sits in extraction, and the stakes are high. Two of the three largest credit rating agencies use this technology on the documents underlying their analysis. An extraction error does not stop at a spreadsheet. It can propagate into a published rating that prices debt for entire markets and determines what institutional investors may hold. The same applies in miniature to loan applications and credit agreements feeding lending decisions.

A second exposure is format dependent quality, since sophisticated issuers producing standardized documentation will be read more reliably than smaller or less conventional ones. No error analysis by document type or issuer is published.

Error is made findable by design rather than discovered only downstream. Verifier models test extracted values against the institution's own rules before delivery, and human experts review through the interface. The company states that its solutions enhance explainability and auditability, so a wrong figure can be traced to the document and the step that produced it. For an institution defending a rating or a credit decision to an examiner, that chain is the substance of its answer. No commercial guarantee or indemnity is published, and nothing describes the vendor's obligation when the verification itself fails.

The architecture is described in technical detail. It names neuro symbolic systems, finance specific language models, AI verifiers and deep learning components, and it puts the training set at more than 1.3 million financial documents, with the document types listed. Because the models are downsized to run on premise, they are evidently the company's own rather than calls to an external service. That shortens the chain at the decisive point and is a meaningful disclosure by implication. No third party model or data provider is named, and no subprocessor list appears for the hosted deployment path.

The stated design intent is to inject extracted insight into downstream systems rather than to hold it. The company describes itself as delivering a complete solution integrated into the daily operations of business and data science teams, which is the right posture for a data layer. On premise and private cloud deployment means it installs inside an institution's existing estate rather than sitting outside it. No core banking, risk, ratings or analytics platform is named, and there is no public developer documentation.

The deployment disclosure explains the engineering behind the option rather than simply listing it. Models are downsized and fine tuned specifically for the financial industry so they can run on premise or in a private cloud environment. The company links this directly to superior accuracy and security. Its reasoning is that financial firms are custodians of customer data operating under strict regulatory guidelines. Vendors whose architecture depends on external model interfaces usually cannot offer an on premise option at all. Cognaize designed around that constraint from the start.

Pricing, packaging and the basis of charge are not published. The company is candid about why buyers hesitate. It names long demonstration and proof of concept cycles, along with costs, as reasons financial services firms have become wary of AI projects. Naming the problem is not the same as publishing terms. Nothing indicates whether charge falls per document, per user, per deployment or by license.

Five institution types are served, spanning banks, insurers, asset managers, data providers and credit rating agencies, which reaches both the users of financial analysis and the firms that produce it for everyone else. Document coverage is the second dimension and it is deliberately aimed at the hardest material: credit agreements, financial reports, ESG disclosures, loan applications, regulatory filings and trustee reports rather than standardised forms. Offices span the United States, continental Europe and a development centre. What is not evidenced is the geographic distribution of the customer base itself.