Centauri AI

The removal test leaves analysts and spreadsheets, which is precisely the state the company describes. The distinguishing claim is interpretation rather than retrieval: the platform is described as going beyond keyword search to read complex provisions across multiple sections of an agreement in order to determine, for example, whether a loan sits senior or subordinated.

That is a judgement assembled from scattered evidence rather than a field lifted from a page, and it is joined by natural language querying over past deals and agents that carry out data cutting, portfolio updates and performance reporting.

Verification is built into the output rather than offered as a policy. Every extracted term links directly to its source passage in the original document, so an analyst can confirm any value at the point of use, and the company describes supporting human reviewed, citation backed results for high value workflows. That is the right construction for work feeding investment decisions, since the reviewer checks a claim against evidence rather than accepting a summary. What is absent is any description of confidence thresholds, of which outputs are flagged for review, or of what proceeds unchecked.

A published accuracy figure with its task specified is more than most vendors here offer: over 92 percent accuracy extracting key terms across thousands of pages of real credit agreements, with one documented engagement covering 20 agreements and 18 named terms including use of proceeds, credit facilities and benchmark rate, and a footnote indicating stated methodology.

Citation linking makes every individual output checkable, which is a stronger transparency property than an aggregate score. Held at B because the benchmark is self run with no independent validation, and 92 percent on legal provisions means roughly one term in twelve needs correction, a rate the company does not discuss.

Two deployments are documented and neither customer is named: a brokerage team at a public investment bank described as using the product daily following launch, and a customer engagement processing 20 credit agreements. Backing comes from a leading accelerator and a major technology company's startup programme alongside a pre seed round of around 500 thousand dollars, with reported revenue near 600 thousand and a team of five. This is a genuine early stage profile rather than a thin one, and the evidence base remains a handful of engagements with no named institution, volume figure or retention data.

No boundary statement was located. Confidentiality is asserted as a principle without being operationalised, and the specific question goes unanswered: whether documents processed for one investment firm inform extraction models used for another, and whether deal terms extracted for a client persist in any shared layer. Firms competing for the same assets would want that stated explicitly before uploading a live deal's agreements.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located. Server location is stated as domestic, and confidentiality is named as a design principle alongside transparency and reproducibility, which is appropriate given the material: credit agreements, covenant certificates and portfolio data are among the most commercially sensitive documents an investment firm holds, and a competitor learning a deal's pricing or covenant thresholds would be a direct harm. Nothing states retention or handling.

The company reports certification under the principal service organisation control standard at the more demanding level, which assesses controls over a period of operation rather than at a point in time. For a team of five that is a deliberate and costly choice, and it is the credential an institutional investor's diligence would ask for before permitting live deal documents to leave its environment. Held at B because no trust centre, penetration testing summary, subprocessor register or further framework was located.

No regulator, statute or rule is named. The documents processed sit inside regulated activity, since credit agreements, covenant compliance certificates and securitisation data support decisions by regulated investment managers with their own record keeping and valuation obligations, and none of that framework is referenced. Audit trails are described as a product feature rather than as a response to any stated requirement.

No individual is assessed and the adapted exposure is institutional. It is still real: an extraction error on a covenant threshold, a maturity date or a benchmark rate propagates into valuation, pricing and risk reporting, and a system reading agreements at scale will fail unevenly across document types, drafting conventions and originators, meaning smaller or non standard issuers whose paperwork departs from market templates are likeliest to be misread. No analysis of error distribution by document type or issuer is published.

No guarantee, indemnity or correction process was located. The buyer is an institution with its own analysts, and the citation linking gives it a practical remedy in that any extracted term can be traced to source and corrected before use, which is genuine protection at the working level. What is absent is contractual: nothing states responsibility if an incorrect extraction reaches a valuation or an investment committee paper, which is the exposure that matters at this end of the market.

No base model, provider, hosting arrangement or subprocessor is identified. Inputs are the customer's own documents rather than licensed external data, which removes the usual provenance question and replaces it with a different one: for a company of this size the underlying language model is almost certainly a third party interface, and a buyer cannot tell whose, where it runs, or whether confidential deal documents traverse another vendor's infrastructure.

Output flexibility is good, with results delivered as database queries, structured files, live dashboards and documents in the formats investment teams already use, and querying available in natural language over past data. That addresses the reuse problem the company identifies, namely that reports exchanged as spreadsheets and slides cannot easily be queried later. What is missing is integration inward: no named portfolio, order management, loan administration or data warehouse system appears, and no developer documentation was located.

Server location is identified as domestic, which is a partial answer and more than many vendors give. No hosting provider, region selection, private deployment option or residency commitment appears, and for a platform ingesting confidential deal documents from institutional investors a private or dedicated environment is a common procurement requirement that goes unaddressed.

A subscription model covering data analytics with consulting alongside it is identified, which is more than most vendors disclose about structure, and no price, tier or unit appears. For a platform sold to investment teams the material question is whether charge follows seats, documents processed or deals analysed, and nothing indicates which, nor how the consulting component is scoped against the software.

Focus is deliberately narrow and well chosen, covering structured finance, private credit and private equity for investment firms, trading desks and banks, with mortgage securitisation named as the entry point. That specificity is a strength for the buyers it serves and it is still narrow coverage: one country, a small number of asset classes within alternative investments, and a stated ambition to reach investment teams generally that remains ahead of the current footprint.