AI Cash Flow Forecasting: the 13-Week Model

What AI cash flow forecasting means

AI cash flow forecasting is software that pulls the dates and amounts already sitting in the ledger, open sales invoices, open supplier bills, the payroll calendar and the VAT or sales-tax due date, and turns them into a week-by-week cash position, then drafts a narrative explaining the weeks where cash runs tight. It matters because building this by hand every week in a spreadsheet is exactly the kind of repetitive, data-heavy task a person should not have to redo from scratch.

The same stance applies here as everywhere else in this knowledge base: AI assembles and computes, a named person sets the assumptions and owns the number that goes to the bank or the board. This guide covers why the forecast is thirteen weeks and direct rather than indirect, the building blocks it needs, a worked week, how to run a scenario, and where AI adds value against where it must be checked, including the case for using a tool rather than the model's own arithmetic.

Why thirteen weeks, and why direct

A thirteen-week forecast covers roughly one calendar quarter, long enough to see a VAT payment, a payroll run and a seasonal dip in the same view, short enough that every line is still a real, dated item rather than a guess. It answers a different question from the income statement: not whether the business is profitable, but whether it will have enough cash on the days it needs it.

A direct forecast lists actual cash in and cash out by source, collections from customers, payments to suppliers, payroll, tax, rather than starting from profit and adjusting for non-cash items the way the indirect cash flow statement does. Our guide on cash flow is not profit covers why a profitable business can still run out of cash; a 13-week direct forecast is the tool that catches it early enough to act.

The building blocks: AR, AP, payroll and tax dates

A forecast is only as reliable as the dates behind it, and not every line carries the same confidence. Sort every input into confirmed, meaning a date and amount already fixed, or modelled, meaning AI or a person has estimated it from a pattern, and never blend the two without saying which is which.

• Accounts receivable: open invoices due this quarter, with the collection date modelled from each customer's actual days-to-pay history rather than the invoice's stated term.
• Accounts payable: open bills, confirmed to the date the business intends to pay them, which may be later than the due date within agreed terms.
• Payroll: confirmed dates and amounts from the payroll calendar, the most reliable line in the whole forecast.
• VAT or sales tax: confirmed once the return is prepared, modelled from the prior period's pattern before that.
• Fixed costs: rent, loan repayments and financing costs, confirmed from the contract or the loan schedule.
• One-off items: a planned capital purchase, an asset sale or a financing draw, each dated and owned by the person who approved it.

Drivers and assumptions behind the model

Behind the dated lines sit a handful of drivers that decide how the forecast moves if the business does not: expected sales growth, how many days later customers pay than their stated term, any planned change in supplier terms, and known one-off items such as a hire or a piece of equipment. AI can build the mechanics of a driver-based model quickly, reading twelve months of the ledger to estimate a customer's typical days-to-pay or a category's seasonal pattern.

What AI should not do is choose the driver values. A sales-growth assumption, a decision to stretch supplier payments, or a view on which customer might pay late this quarter are judgement calls that belong to the finance team, informed by what AI found in the data but not delegated to it. Write the driver values and who approved them into the model itself, so anyone reading the forecast in week nine can see what week one assumed.

Worked example: building one week of the forecast

Week six of a distributor's forecast opens with EUR 82,400.00 in the bank. Collections expected that week, drawn from open invoices and each customer's typical days-to-pay, total EUR 46,500.00. Confirmed outflows for the week are EUR 31,200.00 of supplier payments, EUR 28,750.00 of payroll on its fixed date, a VAT payment of EUR 9,600.00 and rent of EUR 4,200.00, an outflow total of EUR 73,750.00.

The net movement for the week is a negative EUR 27,250.00, collections of 46,500.00 less outflows of 73,750.00, which brings the closing balance to EUR 55,150.00: 82,400.00 less 27,250.00. Because the business holds a minimum cash buffer of EUR 40,000.00, a closing balance of 55,150.00 needs no action this week, and the model carries that balance forward as next week's opening figure.

Scenarios: what happens if a large customer pays late

A single base-case line is a forecast of one future; a scenario tests what happens if a specific assumption breaks. The most common scenario in a 13-week forecast is a large customer paying later than modelled, because one delayed payment can swing a week's result more than a dozen small ones.

In the same forecast, week eight's base case shows a closing balance of EUR 72,000.00, comfortably above the EUR 40,000.00 buffer. The downside scenario removes a EUR 60,000.00 collection from a customer who has paid later than terms on several invoices over the past year, pushing it out three weeks to week eleven. Week eight's closing balance falls to EUR 12,000.00, a shortfall of EUR 28,000.00 against the buffer.

The assistant's draft memo lists three ways to close the gap: draw EUR 30,000.00 on the revolving credit facility ahead of week eight, ask the customer for a partial early payment, or delay that week's VAT payment by requesting a filing extension. The treasurer accepts the credit-line draw and the early-payment request, and rejects the third option outright: delaying a VAT payment risks interest and penalties and is not a lever to use for a liquidity gap, whatever a model proposes.

Why a person owns the assumptions

A forecast that nobody can defend in front of the bank is not useful, however quickly it was produced. Every driver value, every modelled collection date and every scenario choice needs a name attached, both to answer the obvious question when the forecast is wrong and so the assumption gets corrected the next time rather than repeated.

AI's contribution is speed and consistency in assembling the dated inputs and running the arithmetic each week without drift; the business's contribution is judgement about what could plausibly happen next and how much risk it is willing to carry. A weekly ten-minute review of what changed since last week, which invoices moved buckets, which assumption proved wrong, keeps the model honest without becoming a project of its own.

Where AI helps, and where to use a tool instead of a guess

Reading twelve months of payment history to estimate a customer's typical days-to-pay, drafting the narrative behind a variance, and proposing mitigation options are all places AI adds real value. Adding up seventy line items across thirteen weeks by asking a model to do the arithmetic in plain text is not; large language models can state a wrong total with the same confidence as a right one.

As of September 2026, Anthropic's documentation describes a code execution tool that runs Python in a sandboxed container with no internet access, and states that the environment can read CSV, Excel, JSON and text files uploaded through the Files API, which suits building and re-running a cash flow model far better than asking a model to compute totals from a prompt. The same documentation lists current models as supporting adaptive or extended thinking, letting a model spend more internal reasoning on a harder question, such as reconciling several conflicting drivers in one scenario, rather than answering as though it were a single lookup. Structured outputs can force each week's line items into a fixed shape for a spreadsheet or a database to consume, which is a guarantee about the shape of the answer, not about whether the underlying assumption was right.

Security and data protection for AI in cash flow forecasting

A cash flow forecast pulls together some of a business's most sensitive figures in one file: individual customer balances, supplier terms, and a payroll total that a person could use to estimate average pay. As of September 2026, Anthropic's documentation lists the code execution tool and the Files API as not eligible for zero data retention, with a code execution container kept for up to 30 days, so treat any file uploaded for a forecast the same way as any other sensitive attachment, not as a disposable scratch file.

Payroll figures are personal data under the GDPR, and Canada's PIPEDA keeps a business accountable for personal information it hands to a service provider for processing, so a lawful basis, a processor agreement and a view on where the data is stored all matter before payroll totals go into a forecasting tool. Anthropic states that, by default, it does not use inputs or outputs from its commercial products, including the API, to train its models, which is a helpful default and not a substitute for sending aggregated totals where a total is all the forecast needs. This is general information, not legal advice.

Choosing AI cash flow forecasting software

Cash flow forecast software ranges from a spreadsheet template with an AI assistant bolted on, to a dedicated treasury platform that pulls bank balances automatically. The right one depends on where the underlying data already lives and how often the forecast needs refreshing, so test any candidate against last quarter's actual cash movements before trusting its next quarter.

• Does it pull open AR and AP directly from the ledger, rather than requiring the totals to be typed in each week?
• Can every line be tagged confirmed or modelled, with the modelling method stated?
• Does it support at least a base, upside and downside scenario, with the changed assumption visible?
• Are driver values and who set them recorded, so the forecast can be audited later?
• Does it produce a short narrative a non-finance reader, such as a lender, can follow without the spreadsheet?
• Where is payroll and customer data processed and stored, and is a processor agreement available?

Cash flow forecasting in Skyline Nexus ERP

Skyline Nexus ERP does not yet build a rolling 13-week forecast automatically, but it already holds the inputs the model above needs. The AR Aging Report and AP Aging Report give the collection and payment buckets a forecast starts from, Treasury's Daily Cash Position, Cash Flow Report and Bank Balance Summary give today's treasury position, and Treasury's Scheduled Payments list gives the outflows already queued for approval. Fiscal Authority's Cash Flow statement then shows, once a period closes, how cash actually moved between operating, investing and financing activities, from the Beginning Cash Balance to the Ending Cash Balance, which is what a forecast should be checked against after the fact.

The in-app assistant answers how-do-I questions from the product's own help library and, within a user's permissions, gives live figures such as customer dues, supplier dues, expenses for a period and the sales trend on request. Claude Desktop and other MCP clients reach the same figures through the Skyline Nexus ERP MCP server once an administrator issues a personal access token. Building the rolling forecast itself, and having Claude draft or refresh it automatically inside the ERP, are being rolled out on the Skyline Nexus ERP roadmap; export today's aging and treasury reports in the meantime, and tell us your priorities so we can confirm a go-live date.