AI cash‑flow forecasting
Also called cash flow prediction, predictive cash forecasting, 13-week cash forecast, cash runway forecasting
Can AI forecast a small business's cash flow? What the research shows, what the tools cost, the data you need and why forecasts miss.
A cash forecast is worth having, but the AI part has barely been tested in small firms. In a 2016 study, the median US small business held only 27 days of cash buffer, and 51% of US small employers report uneven cash flow. In a 2025 study of a system deployed at a startup serving freelancers, a model predicting late invoices scored 0.56 balanced accuracy, barely better than a coin toss, with 297 invoices to learn from. Some accounting plans already include a basic short-range forecast; start there.
At a glance
| Status | In pilots |
|---|---|
| What it does | Projects your bank balance weeks or months ahead from open invoices, bills, payroll and past bank patterns; AI tools also predict when each customer will actually pay. |
| Typical cost | Short-range forecast bundled in some accounting plans from R450 per month incl. VAT; dedicated tools from £159 or $215 per month billed annually, excl. VAT (vendor list prices, checked Oct 2026) |
| Data it needs | A year or more of bank transactions, open invoices and bills with due dates, payroll and tax dates, and each customer's payment history. |
| Setup effort | Medium: connecting data is quick, but recurring items, tax dates and one-off events need setting up and checking every week. |
| Evidence | Thin. Academic model studies mostly on high-volume or synthetic data, one small-firm deployment with weak results; no independent study of business outcomes. |
What is AI cash-flow forecasting?
A cash-flow forecast predicts your bank balance in the weeks and months ahead, so you see a shortfall before it arrives. The simple version adds up invoices due in, bills due out, payroll and tax. Some accounting plans include one covering the next 30 to 180 days, and dedicated tools add scenarios and a rolling 13-week view.
The AI version goes further. It learns from your bank history and customers' payment behaviour to predict when money will really arrive, rather than when it is due, and fills in spending that has no invoice yet. Analysts list predicting customer payment dates among the top AI use cases in corporate finance.
The need is real. In a 2016 study of bank data from 597,000 firms, the median US small business could cover only 27 days of outflows from cash if income stopped, and a quarter had fewer than 13 days.
How does it work?
Most tools combine three layers.
- Known items: open sales invoices and supplier bills with due dates, payroll, loan repayments and tax dates, pulled from your accounting and payroll systems.
- Patterns: models trained on past bank transactions project recurring income and spending, such as card sales or fuel, that never appear as invoices. Researchers are testing language models that sort small-firm bank transactions into categories for this purpose.
- Payment timing: a model estimates when each customer will actually pay, based on their history.
The output is a daily or weekly balance line, ideally with a range, plus scenarios such as losing a large customer or making a hire.
Late payment is the biggest source of error. UK businesses received on average 17% of payments late, so a forecast that assumes invoices are paid on the due date will be too optimistic.
What it needs to work
A forecast is only as good as the ledger it starts from: reconciled bank balances, every invoice and bill captured, and enough history to show how customers really pay.
- At least 12 months of bank transactions for every account and card, ideally 24 months so the model sees each season twice
- All open sales invoices and supplier bills in the accounting system with due dates, updated at least weekly
- Customer payment history (invoice date, due date, date paid) so the forecast uses actual rather than promised receipts
- Payroll totals and pay dates from the payroll system, plus VAT, PAYE and provisional tax dates
- Loan, lease and debit-order schedules, and known one-off items such as annual insurance or planned purchases
- A weekly owner who updates assumptions and compares last week's forecast with what actually happened
Most use cases in this guide succeed or fail here, on the data underneath them, before any AI is involved.
What does the evidence show?
The models work best where data is plentiful. A study of high-volume receivables, aimed at firms such as online retailers and airlines, found a type of neural network beat traditional statistical methods. That is not a 30-person business.
The closest small-firm evidence is a 2025 study of a system deployed at a startup serving freelancers. Its late-payment model reached 0.56 balanced accuracy on the startup's 297 real invoices, against 0.72 on a larger public dataset. Its cash-flow module had an average error of about 9% to 13%, but the paper describes only simulated data for that test, and a standard forecasting tool produced 159% error with one month of history.
Across finance AI generally, 91% of respondents to a 2025 Gartner survey of finance leaders, not focused on small firms, reported low or moderate initial impact.
We found no independent study showing that small businesses using AI forecasts hold more cash, borrow less or fail less often than those using a simple ledger-based forecast.
AI forecast vs spreadsheet forecast
| Compared on | AI forecast | Spreadsheet forecast |
|---|---|---|
| Speed to update | Refreshes daily from connected data | Manual update, often weekly or monthly |
| Error profile | Can be confidently wrong with little history or after a one-off event | Formula errors and stale assumptions, but easy to see why it is wrong |
| Cost model | From £159 a month for a dedicated tool, or short-range forecasts bundled in accounting plans | Staff time and software you already own |
| Best for | Many customers, frequent transactions and two or more years of clean data | Few large customers and lumpy contracts where judgement beats pattern |
Status and evidence
Rated pilots because forecasting tools are widely sold, but the AI layer that predicts customer payments and spending has been tested mainly on high-volume or synthetic data. The one small-firm deployment we found produced weak payment predictions. Evidence is thin: a handful of academic studies and an analyst survey of finance leaders, with no independent outcome data for small businesses.
Where does it go wrong?
- Too little history. A business with a few hundred invoices gives a model little to learn from, and results can be close to guessing.
- Customers paying late. A forecast built on due dates overstates cash. Use each customer's actual payment record.
- Lumpy one-offs. VAT, provisional tax, annual insurance, bonuses and equipment purchases break pattern-based forecasts unless someone enters them by hand.
- A stale ledger. If bills are not captured and bank lines not reconciled, the starting balance is wrong and so is every week after it.
- False precision. A smooth line hides the range of outcomes. Plan around the worst case, not the middle.
- Seasonality seen once. A model cannot learn a seasonal pattern from a single year.
- Nobody acts. A forecast showing a gap in six weeks is useless if no one chases debtors, delays spending or arranges funding in time.
Numbers it should move
Ask the demo business
Northgate Supply is a fictional wholesale distributor of catering and packaging supplies: 28 staff, R46.6m in sales over the last year. Its data is generated for this demo, to 30 September 2026.
Starting from R1.65m in the bank, cash is lowest in the week ending 11 Nov, at R780 406. Supplier bills for September’s stock fall due that week, before most of October’s sales come in. Worth lining up collections from the slowest payers before then.
How this was worked out
Opening bank balance, plus open invoices expected on each customer’s recent payment pattern (overdue ones from next week, disputed ones left out) and new sales at the current run rate collected on 30-day terms plus the usual delay, less supplier bills by due date, payroll on the 25th or the working day before (with 13th cheques in December), rent and monthly VAT. Amounts include VAT.
Where the data usually lives
Where the records this use case needs are usually kept in a small business. It is not a list of recommended tools.
Frequently asked questions
How much does cash flow forecasting software cost?
Basic short-range forecasts come bundled with some accounting plans: Xero's South African plans include a 30-day forecast from R450 a month and 180 days on the R1,095 Premium plan, prices including VAT. Dedicated forecasting tools cost more; one lists plans from £159 to £199 a month excluding VAT, depending on billing period, with a 13-week view and monthly forecasts up to 36 months.
Do I need a data scientist to use AI cash flow forecasting?
No. Dedicated forecasting tools connect directly to cloud accounting systems; one lists integrations with Xero and QuickBooks Online. What you do need is a bookkeeper who keeps invoices, bills and bank reconciliation current, and an owner or finance lead who reviews the forecast weekly and adds the one-off items the software cannot know about.
How accurate are AI cash flow forecasts for small businesses?
There is little independent evidence. In the one small-firm deployment we found, a model predicting which invoices would be paid late scored 0.56 balanced accuracy, where 0.5 is chance. The same study's cash module averaged about 9% to 13% error, but on simulated data. Treat vendor accuracy claims as unproven until you have compared forecasts with actual balances for a few months.
How much history do I need for an AI cash flow forecast?
More than many small firms have. In one study, a standard forecasting method produced 159% error when given only one month of history, and published successes come from high-volume businesses. As a working rule, have at least a full year of reconciled bank data so the tool has seen each season once, and enter known one-off payments by hand.
Does cash flow forecasting software raise POPIA or UK GDPR issues?
It can, because bank lines and payroll name individuals. Where a provider processes that data on your behalf, UK GDPR requires a written contract binding the provider to you. POPIA similarly requires a written contract obliging the provider to keep personal information secure, and its definition of personal information can extend to companies as well as people. Share payroll totals rather than payslips where you can. This is not legal advice.
More in cash & finance
Sources
- 1Cash is King: Flows, Balances, and Buffer DaysJPMorgan Chase Institute, 2016Survey
- 22025 Report on Employer Firms: Findings from the 2024 Small Business Credit SurveyFederal Reserve Banks, 2025Government
- 3Financial Management System for SMEs: Real-World Deployment of Accounts Receivable and Cash Flow PredictionarXiv (Jagiellonian University), 2025Study
- 4Cash flow prediction: MLP and LSTM compared to ARIMA and ProphetElectronic Commerce Research (Springer), 2021Study
- 5Using Large Language Models to Analysing SMEs Cash FlowMacquarie University (Master of Research thesis), 2025Study
- 6Gartner Identifies 5 Top Use Cases for AI in Corporate FinanceGartner, 2022Survey
- 7Gartner Survey Shows Finance AI Adoption Remains Steady in 2025Gartner, 2025Survey
- 8Late payments research: understanding variations in payment performance and practices across business sectors and sizes (HTML executive summary)Department for Business and Trade, 2024Government
- 9Pricing Plans | Xero ZAXero, 2026Vendor
- 10Pricing for Cash Flow Management Software for Businesses - FloatFloat, 2026Vendor
- 11When is a contract needed and why is it important?Information Commissioner's Office, 2024Government
- 12Protection of Personal Information Act 4 of 2013Government of South Africa, 2013Government
This entry is educational. It is not legal, financial or professional advice, and it does not recommend any vendor. Figures attributed to vendors are their own claims.
- 1Cash is King: Flows, Balances, and Buffer DaysJPMorgan Chase Institute, 2016Survey
- 22025 Report on Employer Firms: Findings from the 2024 Small Business Credit SurveyFederal Reserve Banks, 2025Government
- 3Financial Management System for SMEs: Real-World Deployment of Accounts Receivable and Cash Flow PredictionarXiv (Jagiellonian University), 2025Study
- 4Cash flow prediction: MLP and LSTM compared to ARIMA and ProphetElectronic Commerce Research (Springer), 2021Study
- 5Using Large Language Models to Analysing SMEs Cash FlowMacquarie University (Master of Research thesis), 2025Study
- 6Gartner Identifies 5 Top Use Cases for AI in Corporate FinanceGartner, 2022Survey
- 7Gartner Survey Shows Finance AI Adoption Remains Steady in 2025Gartner, 2025Survey
- 8Late payments research: understanding variations in payment performance and practices across business sectors and sizes (HTML executive summary)Department for Business and Trade, 2024Government
- 9Pricing Plans | Xero ZAXero, 2026Vendor
- 10Pricing for Cash Flow Management Software for Businesses - FloatFloat, 2026Vendor
- 11When is a contract needed and why is it important?Information Commissioner's Office, 2024Government
- 12Protection of Personal Information Act 4 of 2013Government of South Africa, 2013Government