BRENNEN

Industries / Finance and business services / Finance and banking / Fintech

Banking for small businesses

Banking for small businesses is the accounts, cards and loans that software firms offer to small companies, often deciding on credit from a firm's own sales and bank records instead of property pledged as collateral.

6

Milestones achieved, from 2007 to 2022.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2007

    Business loans scored online

    OnDeck, a New York lender, made its first loan in 2007 and built a scoring system, the OnDeck Score, that pulls together thousands of data points about a small business from many sources. In its December 2014 share-offering prospectus it said a business could apply on its website "in minutes", get a decision at once and receive funds as soon as the same day, that it had lent more than $1.7 billion since 2007, and that about two-thirds of its loans were decided by a fully automated process.

    sec.gov

    Credit

    OnDeck (On Deck Capital)

  2. 2017

    Invoices sold on exchanges

    India's Trade Receivables Discounting System (TReDS) is a set of licensed online exchanges where micro, small and medium enterprises (MSMEs) upload invoices owed to them by large buyers, and banks and finance companies bid to pay them early. The Reserve Bank of India approved the first exchange, Receivables Exchange of India (RXIL), a joint venture of the development bank SIDBI and the National Stock Exchange, in December 2016, and RXIL says it executed India's first TReDS transaction in January 2017. In March 2025 four exchanges financed 126,424 invoices and bills worth ₹289.9 billion, with more than 135,000 MSME sellers registered, by RBI's statistics.

    rxil.inrbi.org.in

    Credit

    Reserve Bank of India; SIDBI and NSE, RXIL's owners

  3. 2020

    Data in place of collateral

    Bank loans to small firms usually lean on collateral such as property. Leonardo Gambacorta and colleagues (BIS Working Paper 881, 1 September 2020) studied a random sample of more than 2 million Chinese firms that borrowed both from Ant Group, a technology company, and from banks. Ant's credit followed each firm's transaction volumes and network scores and "does not correlate with local business conditions and house prices", while bank credit, secured or not, moved with local house prices.

    bis.org

    Credit

    Ant Group; Leonardo Gambacorta, Yiping Huang, Zhenhua Li, Han Qiu and Shu Chen, who measured it

  4. 2022

    Fintechs and pandemic loan fraud

    The US Paycheck Protection Program (PPP) lent $793 billion to small firms in 2020 and 2021, much of it through fintech firms working with partner banks. John M. Griffin, Samuel Kruger and Prateek Mahajan found that signs of misreporting, such as unregistered businesses and several businesses at one home address, "consistently concentrate in certain FinTech lenders", and that suspicious fintech lending in 2021 was four times its level at the program's start (Journal of Finance, 2023). On 1 December 2022 the House Select Subcommittee on the Coronavirus Crisis reported that Kabbage, which handled more than 310,000 PPP loans, cut its risk and account review teams by about half in May and June 2020, and that Blueacorn spent $8.6 million on fraud prevention out of more than $1 billion in fees.

    LessonThe subcommittee found that Bluevine's fraud fell steeply once its partner bank pushed it to invest in controls, so a bank partner's oversight can decide how much fraud a fintech lets through.

    api.crossref.orgcoronavirus-democrats-oversight.house.gov

    Credit

    Griffin, Kruger and Mahajan; the House Select Subcommittee on the Coronavirus Crisis

  5. 2022

    Brex drops small firms

    In June 2022 Brex, a US company whose corporate cards and accounts began with startups, told "tens of thousands" of small businesses without venture, angel or accelerator funding that their accounts would close on 15 August. Its chief executive, Henrique Dubugras, said most were brick-and-mortar firms such as bakeries, restaurants and small design agencies; some funded startups notified by mistake were reinstated.

    LessonA small business that banks with a fintech carries the risk that the firm changes which customers it serves.

    techcrunch.com

    Credit

    Brex

  6. 2022

    Business loan prices disclosed

    California's SB 1235, a 2018 law by Steve Glazer, and the regulations of the state's Department of Financial Protection and Innovation (DFPI), in force from 9 December 2022, require commercial financing providers to show a small business the amount of funding, the annual percentage rate (APR), the payment amount, the term and the prepayment policy before it signs. They cover installment loans, open-end credit, factoring and merchant cash advances (MCAs), lump sums repaid from a share of the business's future sales. The DFPI called them "first-in-the-nation protections for small business borrowers".

    dfpi.ca.gov

    Credit

    Steve Glazer; the DFPI

Watch for

  1. Who gets small-business loans

    Section 1071 of the Dodd-Frank Act requires lenders to report data on small-business credit applications, including data on the owners' demographics, so that fair-lending patterns can be studied. The CFPB's 2023 rule was held up by lawsuits and delays; its revision of 1 May 2026 covers only lenders making at least 1,000 such loans in each of two years and businesses with $1 million or less in yearly revenue, leaves out merchant cash advances, farm loans and small-dollar loans, and sets compliance from 1 January 2028.

    Achieved whenCovered lenders report under the rule and the CFPB publishes the first data.

    federalregister.gov

    Claimed by

    CFPB, May 2026

Open

  1. 01

    Cash advance costs, estimated

    Open-source code that computes the estimated APR California requires for a merchant cash advance from the offer's terms and the business's past monthly sales, following the state's historical method, with a public set of worked test cases. Use invented offers and sales figures only.

    The brief
    Credit · unclaimed
  2. 02

    Cash-flow figures, scored

    A public test of the figures a small-business lender takes from bank statements, not just the transactions: monthly sales deposits with transfers and loan proceeds left out, average daily balance, the number of overdraft and returned-payment fees, and recurring debits to other lenders or cash-advance providers. It needs invented statement PDFs labeled with those figures and a scoring script.

    The brief
    Credit · unclaimed

Open 01 · the brief

Cash advance costs, estimated

Credit · unclaimed

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

What to build

Open-source code that computes the estimated APR California requires for a merchant cash advance from the offer's terms and the business's past monthly sales, following the state's historical method, with a public set of worked test cases. Use invented offers and sales figures only.

Why it is still open

An MCA has no fixed payment schedule, so California makes the provider project payments from the business's average monthly sales over a period it fixes between 4 and 12 months, then compute the APR by the methods of Appendix J to Regulation Z. In September 2025 the California legislature passed an amendment that would shield providers whose estimates follow guidance from the DFPI or the Attorney General, a sign the estimate is still contested. This research found no open-source calculator for the method.

law.cornell.edulaw.cornell.edumanatt.com

Built with what exists

Python; the text of 10 CCR sections 930, 931 and 940; Appendix J to 12 CFR Part 1026 for the actuarial method; an AI model such as Claude Fable 5.2 (see the Note "Three working devices") or Mistral Large 4 (see the Note "A trillion open parameters") to draft test cases from the rule text, each then checked by hand.

  1. First step

    Code Appendix J's actuarial method and reproduce the APRs of 10 fixed-payment loans, then add section 930's sales projection.

  2. Done looks like

    At least 50 invented offers covering fixed daily and weekly debits, percentage-of-sales holdbacks, businesses open fewer months than the provider's period and months excluded under the rule, each with an expected APR worked out independently by two people who agree to within 0.01 percentage points, and the code matching every case.

Open 02 · the brief

Cash-flow figures, scored

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 of the figures a small-business lender takes from bank statements, not just the transactions: monthly sales deposits with transfers and loan proceeds left out, average daily balance, the number of overdraft and returned-payment fees, and recurring debits to other lenders or cash-advance providers. It needs invented statement PDFs labeled with those figures and a scoring script.

Why it is still open

FinRegLab's 2019 study found cash-flow measures predict repayment about as well as credit scores, so these figures carry weight. Bankstatemently's open benchmark scores only the extraction of transactions, has released 5 of its 15 statements and keeps its answers on its own server. AgamiAI's Apache-licensed set of synthetic Indian business statements labels each transaction but not these figures, uses common layouts only and says it is not suited to credit decisions.

github.comhuggingface.co

Built with what exists

AgamiAI's synthetic statements as a base; Python with ReportLab to draw new layouts for invented US, UK and Kenyan banks; Bankstatemently's submission format for transactions; extraction pipelines to compare, such as Mistral Large 4 run on the lender's own machines (see the Note "A trillion open parameters") and Claude Fable 5.2 (see the Note "Three working devices").

  1. First step

    Take 20 AgamiAI statements, add loan proceeds and debits to other funders, render them again as PDFs and write the script that computes the five figures.

  2. Done looks like

    At least 200 labeled statements in at least 10 layouts from at least 3 countries, each figure computed by script from known transactions; a public table for at least three pipelines giving the share of figures within 1% of the label; data and code under an open license.