What problem does it solve?
Many small businesses struggle to get bank credit. Traditional underwriting asks for audited financial statements, tax returns and collateral, which many micro and small firms do not have, and relies on manual spreading and credit memos for loans that are small relative to the effort. The result is a high cost to serve, slow decisions and owners who turn to more expensive finance or go without. At MYbank, which lends on its own data and models, over 72 percent of the 3 million borrowers it added in 2023 had never had a business loan from a bank before.
Most of these businesses do have a detailed financial record: their bank account, card acquiring, marketplace or accounting software. Cash flow underwriting reads that record directly. It can decide simple, small facilities in minutes and give credit officers a much better picture for larger ones, but only if the data connections, the model and the credit policy are designed together.
How does it work?
- Connect the data. With the owner's consent the engine pulls bank transactions, acquiring or marketplace sales, and accounting data through APIs, or reads uploaded statements with document AI.
- Build the cash flow picture. Transactions are categorised into revenue, payroll, suppliers, taxes and existing debt service; seasonality, volatility and concentration are measured.
- Score and size. A model estimates default risk, and policy rules translate free cash flow into an affordable limit and tenor.
- Decide or refer. Small, clean applications within policy are approved automatically; the rest go to a credit officer with a prepared summary, the key ratios and the reasons.
- Keep the owner informed. Status updates and requests for missing documents go out on the owner's channel of choice.
- Keep watching. The same data feeds monitor the borrower after the loan is drawn.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- API and system to system, Mobile app, Web chat, Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Users served | Not pooled | at least 53 million | 1 | 1 organization |
Value drivers: Speed and cycle time, Inclusion and access, Lower cost to serve, Revenue growth, Risk and loss reduction.
Indicative value
A bank receiving 20,000 small business loan applications a year
USD 600,000 to USD 3.2 million
Underwriting effort released per year
How this is calculated
Formula: applications * automatedShare * hoursSaved * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Small business loan applications per year applications, applications per year | 20,000 | 20,000 | The reference bank. |
| Share of applications decided through the automated cash flow path automatedShare, fraction of applications | 0.2 | 0.33 | iTnews reported that NAB's QuickBiz platform had been held up for deciding one in every three small business loans, and a NAB executive later said 45 percent of small business lending accounts were opened through it. Not every application on such a platform is decided without an underwriter, so the high bound stays at one in three. The low bound of 0.2 is an editorial floor, not a reported figure: replace it with your own measured automation rate. Source |
| Underwriter hours saved per application on that path hoursSaved, hours per application | 3 | 6 | Editorial assumption for spreading, analysis and memo writing on a small facility. Replace with your own time study. |
| Fully loaded cost of an underwriter hour hourlyCost, USD per hour | 50 | 80 | Editorial assumption. Replace with your own cost. |
What it leaves out: Counts underwriting effort only. It leaves out additional lending volume from faster decisions, changes in credit losses, data access fees, and the cost of building and validating the model.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
MYbank
China · Banking · 2024
MYbank, the Chinese digital bank associated with Ant Group, lends to small and micro businesses with its "310 model": a collateral free business loan that takes under three minutes to apply for on a phone, under one second to approve and no human interaction. The bank says AI, including Ant Group's Bailing foundation model and a supply chain knowledge graph, informs its lending decisions. By the end of 2023 it had served over 53 million small and micro businesses, and over 72% of the 3 million new borrowers it added in 2023 had obtained a business loan from a bank for the first time. It also uses satellite imagery to estimate farm output and an AI conversational system to manage credit lines.
- Users served: at least 53 million, small and micro businesses served, cumulative to the end of 2023
"the bank has cumulatively served over 53 million small and micro-sized enterprises (SMEs) as of the end of 2023"
Claimed by: organization
OakNorth Bank
United Kingdom · Banking · 2020
OakNorth Bank, a UK lender to small and mid sized businesses, underwrites with human credit officers supported by systems that pull in and analyse public and alternative data, and monitors each borrower continuously against a peer group in the same sector and geography rather than waiting for audited financials every six months. By late 2020 it had lent GBP 4.6 billion to 750 businesses since 2016 and sold the same software to other banks. The published figures describe the lending book, not a measured effect of the AI.
No outcome disclosed.
Sumitomo Mitsui Banking Corporation
Japan · Banking · 2020
In November 2020 Sumitomo Mitsui Banking Corporation licensed the credit underwriting and monitoring software built by OakNorth, a UK SME lender, and invested USD 30 million in OakNorth equity. The software pulls in public and alternative data and compares each borrower with sector and local peers, so lenders can underwrite businesses and watch them continuously rather than waiting for periodic audited financials. SMBC's group CFO said the alliance would bring more sophistication to its corporate lending platforms and that the group was harnessing AI through big data and machine learning across its strategic markets in Southeast Asia, such as Indonesia. No outcome figures were published.
No outcome disclosed.
National Australia Bank
Australia · Banking · 2019
NAB's QuickBiz platform decides unsecured small business loans and overdrafts online. In 2021 its product page said NAB reviews the applicant's cash flow, credit score and time in business, and that businesses using Xero, MYOB or QuickBooks can link their accounting data. iTnews reported that the platform uses machine learning to make decisions faster. In 2019 the general manager of digital and sales transformation in NAB's business and private bank said that 45 percent of NAB's small business lending accounts were being opened this way, and that credit decisions often came the same day.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Historical small business applications with repayment outcomes
- Consent based access to bank transaction, acquiring or accounting data
- A written credit policy for the automated path (eligible products, maximum amounts, exclusions)
- Reason codes that credit and compliance have approved
Systems to integrate
- Open banking or account aggregation provider
- Accounting software connectors (for example Xero, MYOB, QuickBooks)
- Acquiring, marketplace or point of sale data where the bank has it
- Loan origination system and decision engine
- Credit bureau and business registry
Complexity: High
Data connections, categorisation quality and credit policy design carry most of the effort. Decisions are regulated in many markets, the model needs validation, and the automated path must be tightly bounded by product, size and risk grade.
- 1
Define the automated lane
Write down which products, amounts, industries and risk grades may be decided without a person. Start narrow, such as unsecured facilities up to a set limit for existing customers.
- 2
Get the categorisation right
Test transaction categorisation on a few hundred real businesses per sector. Revenue, owner drawings and transfers between own accounts are where errors hide.
- 3
Backtest and validate
Score past applicants and compare with outcomes, then take the model through independent validation and add it to the model inventory before any live decision.
- 4
Prepare the referral pack
For cases outside the lane, generate a summary with cash flow charts, key ratios and the reasons for referral, so credit officers start from analysis instead of raw statements.
- 5
Launch with existing customers
The bank already holds their transaction history, which removes the consent step and gives a cleaner first measurement of approval, speed and loss rates.
- 6
Extend to connected new customers
Add accounting and acquiring connections for new to bank businesses once the lane performs as expected.
Guardrails
- Automatic approvals only inside the documented lane; everything else goes to a credit officer
- Specific, recorded reasons for every decline, reduced limit or referral
- Consent recorded for every external data source used in a decision
- Deterministic affordability and exposure limits outside the model
- Monitoring that compares automated approvals with manually underwritten ones
KPIs to instrument
- Share of applications decided in the automated lane
- Median time from application to decision, by lane
- Default and arrears rates of automated versus manual approvals
- Consent and data connection completion rate
- Decline reasons distribution and appeal overturn rate
Human in the loop
Credit officers own every decision outside the automated lane, every appeal and every exception to policy. Credit risk reviews the performance of automated approvals each month and can close the lane for a segment at any time.
Common failure modes
- Misread cash flow
- Transfers between the owner's own accounts or a one off asset sale look like revenue. Categorisation must be tested per sector and suspicious patterns flagged.
- A lane that creeps wider
- Pressure to grow volume pushes larger or riskier loans into automatic decisions. Change the lane only through credit committee with fresh backtests.
- Declines without a real reason
- A small business told only that it "did not meet criteria" cannot act and may complain. Give specific reasons tied to the data.
- Blind spots after drawdown
- Underwriting uses live data but monitoring still waits for annual accounts. Connect the same feeds to early warning.
What are the risks and rules?
EU AI Act
Depends on design
Annex III point 5(b) makes AI systems that evaluate the creditworthiness of natural persons or establish their credit score high risk. Scoring a company is outside that point, but a sole trader is a natural person, and a model that also assesses the personal credit of owners, partners or guarantors evaluates natural persons. The tier therefore depends on who the borrower is and whose creditworthiness the model assesses.
Rules that apply
Guidance
- Guidelines on loan origination and monitoring (European Banking Authority, Europe). Covers creditworthiness assessment for micro and small enterprises and the governance of automated models used in credit decisions.
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 5(b) on creditworthiness of natural persons decides whether a given small business model is high risk.
- MAS Guidelines for Artificial Intelligence Risk Management (Monetary Authority of Singapore, Asia Pacific). Consultation paper of November 2025 proposing supervisory expectations for financial institutions in Singapore on AI inventories, risk materiality, fairness, explainability and human oversight, which apply to credit models.
Controls to put in place
- Model inventory entry, independent validation and an approved scope for automatic decisions
- Written automated lane policy approved by credit committee
- Consent and data lineage records for every decision
- Reason code library reviewed by compliance
- Monthly performance review of automated approvals against manual ones
Frequently asked questions
- How fast can a small business loan be decided with cash flow data?
- For small, simple facilities, very fast. MYbank describes a loan that takes under three minutes to apply for and under one second to approve with no human involved, and a NAB executive said QuickBiz credit decisions often came the same day as the conversation with the banker. Larger facilities still need a credit officer, but with a prepared analysis instead of raw statements.
- Does cash flow underwriting replace financial statements?
- For micro and small loans it often can, because transaction data shows revenue, costs and debt service more currently than annual accounts. For larger facilities it complements statements and supports continuous monitoring after drawdown, as OakNorth Bank does by comparing each borrower with peers in the same sector and location.
- Is SME credit scoring high risk under the EU AI Act?
- Annex III point 5(b) covers creditworthiness of natural persons, so scoring a company is not listed. Sole traders are natural persons, and models that assess owners or guarantors personally also fall inside it, so classify each product by whose creditworthiness is assessed.
How to cite this page
Blits.ai AI Use Case Library, "AI cash flow underwriting for small business loans", last verified 26 September 2026, https://www.blits.ai/ai-use-cases/sme-cash-flow-underwriting. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published