Kaaj

Manual underwriting is what remains without the agents, and it is exactly what the product exists to replace. Multiple agents work together across the entire credit analysis chain rather than handling one step. They span document intelligence, business verification, bank statement analysis and cash flow, then asset valuation, fraud detection and credit memo generation. The founders describe the design as agentic workflows that mimic a lender's own team. Compressing days of work across thousands of documents into under three minutes is not achievable by any other means.

The division of labor is stated clearly and the mechanism supports it. The chief executive puts it directly: by automating the science of credit analysis, the platform frees human underwriters to focus on the art of deal making and subjective assessment, which is their true competitive advantage. The company describes itself as automating the science while keeping humans in charge of the art.

The output is a decision ready analysis and a credit memo, which are inputs to a human decision rather than a decision. The policy alignment feature shows whether an applicant meets that lender's own criteria, so the standard applied belongs to the accountable institution rather than to the vendor. Traceability is named as one of three core properties alongside speed and consistency.

Accuracy is published with a stated method. On live lender statements in September 2026, 99.7 percent of bank statements were parsed without a material dollar error, the median dollar error was $0.00 and the median time per statement was 48.8 seconds. The test reconciles parsed dollar totals against the totals printed on each statement, and statements flagged for a large discrepancy were reviewed by Kaaj's own team, most of them false positives from the check itself.

No sample size is given and no outside party has tested the figure. It also covers parsing only. The credit memo an underwriter decides from carries no published error rate, so an omission in the memo becomes an omission in the decision, and an underwriter working from generated analysis cannot see what the agents did not surface.

Customers are named and results are attributed to them. Quality Equipment Finance cut intake from 15 minutes to 2 per application, landing straight in Salesforce. Harbour Capital moves a package from inbox to its system in two to three minutes, against four to eight hours before. Capital Solutions Providers finishes underwriting on a complete package in 15 minutes rather than 45, and at Hub Funding Solutions the share of applications stuck in process fell from 70 percent to 10.

Executives are quoted by name and title, among them Amur's SVP of Technology Jeremy Gillam and Quality Equipment Finance chief executive Paul Fogle. Volume is stated at more than $20 billion of loan originations processed, a figure the company reports itself. A $3.8 million seed closed in November 2025, led by Kindred Ventures with Better Tomorrow Ventures participating, for a company founded in 2024.

Agents improve with exposure to more borrower packages and more lender policies. The platform serves equipment financiers, brokers, community banks and private credit teams who compete for the same small business borrowers, so what is learned from one lender's applications has direct value to another. Brokers add a further dimension, since the same application may pass through the platform on behalf of several prospective lenders. No data boundary statement is published, and nothing states what is contained.

The payload is substantial. It covers complete borrower packages with bank statements, business and often personal financial records, and thousands of documents per application, processed for lenders across more than $20 billion of originations by the company's own count. A data retention policy is listed in the trust center behind an access request.

No data protection agreement, retention schedule or deletion commitment is public, and nothing published states what happens to the file of a declined applicant or how borrower documents are held between institutions.

Kaaj states SOC 2 Type II compliance on its security page and runs a trust center on Sprinto at kaaj.trust.site. The trust center shows SOC 2, a set of named controls and 34 policies, among them data retention and business continuity, and lists eight subprocessors including Microsoft Azure, Cloudflare and GitHub. The report itself, the policies and penetration test summaries sit behind an access request. No ISO certification or audit period is published.

No supervisor, statute or rule is named. Equal credit opportunity obligations apply to business credit as well as consumer credit, including the requirement to give an applicant reasons for an adverse decision, and United States small business lending is additionally moving toward mandatory data collection designed to expose disparities in who receives credit. A platform generating the analysis on which those decisions rest sits inside both, and neither is addressed.

The account of credit exclusion is structural, identifying the mechanism rather than asserting an outcome. Underwriting a $100,000 loan consumes the same time and resource as a $5 million one, so loans under a million are unprofitable for most lenders and simply do not get made. The central bank's own 2024 small business credit survey records roughly half of applicants failing to receive the full amount they need. That frames exclusion as an economic problem rather than an analytical one.

Fixing unit economics widens access without loosening the credit standard, which is a different and more candid claim than finding more creditworthy borrowers in the same pool. What is absent is evidence. There is no outcome data on who actually got funded, no fairness testing, and no analysis of whether errors scale alongside the fortyfold volume increase the platform enables.

The lender is reasonably placed. The policy alignment feature ties the analysis to its own stated criteria, so it can see which of its rules an applicant met or missed and defend the resulting decision. The small business applicant has nothing described. There is no statement of what a declined borrower is told, no route to correct a misread bank statement or misclassified transaction that shaped the memo, and no notification that agent generated analysis informed the outcome. No guarantee, indemnity or falsifiable commitment is published.

No model provider is named for any of the agents. The public subprocessor list runs to eight entries (Microsoft Azure, Cloudflare, GitHub, Office365, Microsoft Teams, Entra ID, Porter and Hatchet), all infrastructure and tooling, and none of them is a model provider. The primary input is the borrower's own submitted package rather than purchased data, which shortens the chain, though that is implied rather than committed. Business verification draws on Secretary of State filings, web presence and address signals, and the data sources behind those checks are not named.

Results land in the systems lenders already run. Extracted fields, risk summaries and credit memos sync into CRMs (Salesforce, Microsoft Dynamics, HubSpot) and into equipment finance platforms (Leasepath, LTi Aspire, Solifi). Packages arrive by forwarded email, an embeddable Kaaj credit application or a REST API with webhooks, in whatever form the broker sends them. That matters for this buyer set, since equipment financiers and brokers run lean technology teams and cannot absorb a parallel system. No bureau or bank data aggregator connection is named.

Infrastructure runs on Microsoft Azure, named as a subprocessor in the trust center alongside Cloudflare. Exposure is domestic, which keeps residency simpler than for international vendors. Lenders outsourcing analysis of complete borrower packages, bank statements included, still document for their own examiners where that processing occurs. No Azure region, residency commitment or private deployment option is published.

The pricing model is published and the prices are not. Most deployments combine a recurring platform fee with a monthly allowance of applications or usage credits, and overages scale with applications processed, documents analyzed or products used. Implementation fees apply where a workflow needs configuration or CRM and LOS integration. Pricing is stated not to be per seat, so credit, sales and operations staff can work the same deal without adding cost.

Four structures are described: a workflow pilot, platform plus volume, usage credits and an enterprise workflow. No rate or price per application is published, and for a product whose argument is loan level unit economics, that price is the number that decides adoption.

Six buyer types are served, spanning equipment finance companies, small business lenders, brokers, private credit teams, community focused institutions and merchant cash advance or revenue based finance providers, which is a coherent set defined by a shared characteristic the company names directly: high volume, document heavy credit workflows where speed, consistency and traceability matter.

Functional coverage runs the whole analysis chain from intake to credit memo rather than one step, which the founders position as the differentiator against competitors handling parts of it. The limits are geographic and categorical, being United States small business credit specifically.