BRENNEN

Industries / Finance and business services / Finance and banking / Equipment leasing and factoring

Credit checks

Credit checks are the work of judging, before money goes out, whether a business that wants to lease equipment or sell its invoices is real and will pay and, in factoring, whether its customers will pay the invoices.

5

Milestones achieved, from 1968 to 2026.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1968

    Ratios predict bankruptcy

    Edward Altman combined five ratios from a company's accounts, among them working capital, retained earnings and sales, each divided by total assets, into one number, the Z-score, first published in the Journal of Finance in September 1968. Tested on 66 US manufacturers, 33 that had filed for bankruptcy from 1946 to 1965 and 33 matched firms that had not, it sorted 95% of them correctly from the accounts one year before failure and 72% from the accounts two years before, by his own later summary. In 2000 he wrote that the model was "still being used by practitioners throughout the world", and he had made versions for private firms and for companies that are not manufacturers.

    Lessona score fitted to one sample looks better on that sample than it is, so it has to be tested on firms outside it, which Altman wrote "cannot be overemphasized".

    moodle2.units.it

    Credit

    Edward I. Altman

  2. 2001

    Liens found by name

    Revised Article 9 of the Uniform Commercial Code (UCC), the US model law on loans backed by movable property such as machines, stock and invoices, set 1 July 2001 as its start date. It also covers sales of invoices, so a factor that buys them files the same short public notice as a lender, a financing statement, which needs only the debtor's name, the creditor's name and a description of the collateral. Anyone checking a business searches these notices by its name; a notice with the name wrong is "seriously misleading" and loses its effect, unless a search under the correct name with the filing office's standard search logic still finds it.

    law.cornell.edulaw.cornell.edulaw.cornell.edulaw.cornell.edu

    Credit

    the US states that enacted the revised text

  3. 2016

    UK banks share small-firm data

    Regulations made under the UK's Small Business, Enterprise and Employment Act 2015, in force from 1 January 2016, require banks designated by the Treasury to pass credit data on their small and medium-sized business customers, with the customer's agreement, to designated credit reference agencies. Those agencies must give it to other lenders that agree to share their own data in return, so a lender that is not the firm's bank can see its borrowing record.

    legislation.gov.uk

    Credit

    HM Treasury

  4. 2019

    Lease payments join a bureau

    On 24 April 2019 Equifax, a credit bureau, bought PayNet, which sold credit ratings on small businesses to leasing companies, commercial lenders and online lenders in the US and Canada. PayNet's database of small-business loans, leases and lines of credit held over 24 million contracts worth over $1.6 trillion, by its own count; Equifax said PayNet's leasing payment data would strengthen its own commercial data.

    investor.equifax.combiia.com

    Credit

    Bill Phelan, PayNet's co-founder and president; Equifax

  5. 2026

    A score gate removed

    For small loans under its main 7(a) program, which can pay for equipment, the US Small Business Administration (SBA) screened applicants with the FICO Small Business Scoring Service (SBSS) score and raised the minimum from 155 to 165 from 1 June 2025. From 1 March 2026 it dropped the score. Lenders must now analyze each applicant's credit history, two recent months of bank activity and its cash flow, which must cover debt payments at least 1.1 times, as they do for their loans without an SBA guarantee, and may use their own scoring models if those do not rely only on consumer credit scores.

    Lessonwhen a shared score is dropped, each lender's own method decides, so the same file can get different answers from different lenders.

    everycrsreport.comnaggl.orgexperian.com

    Credit

    SBA

Watch for

  1. Invoices approved in seconds

    Triumph Financial says its "instant purchase decision model", launched at scale for its smallest trucking clients in 2024, uses machine learning to screen invoices and approve them for purchase "in seconds without human intervention". For each invoice it judges the broker or shipper that owes the money as well as the carrier selling it, since the customer is the real credit risk.

    Achieved whenan outside party, such as its auditor or bank examiners, reports loss and fraud rates on invoices bought this way against those bought after human review.

    sec.gov

    Claimed by

    Triumph Financial, annual report, February 2026

  2. Credit files in five minutes

    Kaaj, a US start-up selling AI agents to equipment finance companies, private credit funds and other lenders, says it compresses the processing of a small equipment-loan file, which "traditionally took days", into under 5 minutes once all documents are in. It describes that work as verifying the business, reading bank statements, checking UCC filings and valuing the equipment.

    Achieved whena lender using it publishes, or lets an outside reviewer measure, decision times and error rates on real files against its earlier process.

    kaaj.ai

    Claimed by

    Kaaj, May 2026

Open

  1. 01

    Test lien searches by name

    A public test suite for the name searches that decide whether a filed lien counts. Each case pairs an invented business's correct legal name with a name as filed, such as "Acme Tools, Inc." against "ACME TOOLS INCORPORATED", a dropped "The" or a typo, and gives the expected result under a named filing office's written search rules. Add a reference program that applies those rules, then score name-matching programs and AI models against it.

    The brief
    Credit · unclaimed
  2. 02

    Same file, different answers

    A noise audit of small equipment-lease decisions: give the same invented application files to many working credit analysts and to AI models, and measure how far their answers differ. Each file holds a made-up business's profile, two years of figures, a bank-statement summary, a bureau-style report, and the equipment, its price and its vendor. Each analyst decides every file alone (approve or decline and, if approved, the term and any down payment), and 15 of the files come back six weeks later to measure how far each analyst agrees with their own earlier answers. Each AI model decides every file ten times.

    The brief
    Credit · unclaimed

Open 01 · the brief

Test lien searches by name

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test suite for the name searches that decide whether a filed lien counts. Each case pairs an invented business's correct legal name with a name as filed, such as "Acme Tools, Inc." against "ACME TOOLS INCORPORATED", a dropped "The" or a typo, and gives the expected result under a named filing office's written search rules. Add a reference program that applies those rules, then score name-matching programs and AI models against it.

Why it is still open

Under UCC section 9-506 a misnamed financing statement survives only if the filing office's "standard search logic" finds it under the correct name, so the answer depends on each office's rules. Kaaj notes that the legal name on an application often does not match the state registration. The nearest public test, published by CPT Engineering in October 2026, checked its own lien tool on 20 inputs against live records in three states, not against search rules, and says it is not an independent audit.

law.cornell.edukaaj.aigithub.com

Built with what exists

Filing offices' written search rules, from their websites or by request; Python; invented business names only; an AI model such as Claude Fable 5.2 (see the Note "Three working devices") as one of the matchers tested.

  1. First step

    Write one filing office's search rules as code and run 100 invented name pairs through it.

  2. Done looks like

    At least 1,000 name pairs covering the rules of ten or more filing offices, each expected result agreed by two people reading the rule text, and a public table of each tool's agreement with the expected results.

Open 02 · the brief

Same file, different answers

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A noise audit of small equipment-lease decisions: give the same invented application files to many working credit analysts and to AI models, and measure how far their answers differ. Each file holds a made-up business's profile, two years of figures, a bank-statement summary, a bureau-style report, and the equipment, its price and its vendor. Each analyst decides every file alone (approve or decline and, if approved, the term and any down payment), and 15 of the files come back six weeks later to measure how far each analyst agrees with their own earlier answers. Each AI model decides every file ten times.

Why it is still open

A noise audit gives the same case to many people and measures the spread of their answers; Olivier Sibony and Daniel Kahneman describe one at an investment firm in which any two analysts valuing the same company differed by 44% on average. Shawn Cole, Martin Kanz and Leora Klapper paid Indian bank loan officers to assess real past small-business loan files, 14,675 decisions in all, to measure the effect of pay incentives, not the spread between officers on the same file. Raffi Khatchadourian and Rolando Franco (November 2025) measured how often AI models give the same output twice on financial tasks, with no people for comparison. Since March 2026 the SBA has let lenders use their own methods in place of a shared score (see 2026 above). This research found no published measure of how much equipment-finance analysts, or AI models, disagree on the same file.

karriere.mckinsey.denber.orgarxiv.orgnaggl.org

Built with what exists

Invented files drafted in a spreadsheet and rendered to PDF; LimeSurvey, an open-source survey tool, to collect decisions; AI models such as Claude Fable 5.2 (see the Note "Three working devices") and Mistral Large 4 (see the Note "A trillion open parameters") run with fixed prompts; Python for agreement measures such as Fleiss' kappa, a standard measure of agreement among several raters.

  1. First step

    Write ten files, ask five analysts at one lessor to decide them independently, and run one model ten times on each.

  2. Done looks like

    60 files decided by at least 30 analysts from at least six lessors, with a public table of how often two analysts chosen at random disagree on approve or decline, the median gap in the terms they set, each analyst's agreement with their own earlier answers, and each model's agreement across its ten runs; the files and the anonymized answers released under an open license.