AI Cash Flow Forecasting: Rolling Forecasts & Scenario Planning

By Christian Matumona, Kyootek AI Financial Controller

A rolling 13-week cash flow forecast is one of the most valuable tools a finance function maintains and one of the most tedious to keep current. The model itself is mechanical — receipts by category, disbursements by category, opening and closing balance, rolled forward weekly — but updating it accurately every week, and rebuilding scenarios whenever an assumption changes, consumes hours that have little to do with financial judgment and a lot to do with formula maintenance.

This is the deep dive on cash flow forecasting promised in the Complete Guide to AI for Financial Controllers.

Where the time actually goes

The build is where AI saves the most time relative to effort — generating a clean, correctly-linked 13-week structure with categorized inflows and outflows is exactly the kind of formula-heavy, pattern-based work AI tools handle well inside Excel. The ongoing weekly maintenance is where the saved time compounds: a forecast that took most of an afternoon to refresh weekly is now a quarter of that.

Building the base model

Feed AI your historical cash transactions, categorized — customer receipts, payroll, rent, supplier payments, debt service, tax — for the trailing 8-12 weeks, and ask for a 13-week rolling structure with opening balance, categorized inflows and outflows, and closing balance carried forward.

Build a 13-week rolling cash flow forecast structure in Excel formula logic.

Historical weekly cash data (last 10 weeks), by category:
[PASTE: Week | Customer Receipts | Payroll | Rent | Supplier Payments | Debt Service | Tax | Other]

Requirements:
- Opening balance carries from prior week's closing balance
- Each category should have its own row with a simple trailing-average or trend-based formula as a placeholder, clearly marked [ASSUMPTION] so I can override with my own judgment
- Closing balance = Opening + Inflows - Outflows
- Flag any week where the closing balance would go negative
- Provide the formula logic in a format I can paste directly into Excel, not just a description

Every placeholder marked [ASSUMPTION] is exactly that — a starting point built from a trailing average, not a forecast you should trust without review. Replace each one with your own knowledge of upcoming receipts, known supplier timing, and payroll dates before treating the model as live.

Scenario planning without rebuilding from scratch

Once the base model exists, generating best, base, and worst-case scenarios is a matter of varying a small number of assumptions consistently — collection days, a major customer payment date, a supplier negotiation outcome — rather than rebuilding three separate models.

Using the 13-week base forecast above, generate best-case and worst-case variations.

Vary only these assumptions:
- Customer collection days: base [X] days, best [Y] days, worst [Z] days
- [LARGEST CUSTOMER] payment: base on schedule, worst case delayed by [N] weeks
- Discretionary supplier payments: base on schedule, best case 10% deferred by mutual agreement

For each scenario, show the resulting closing balance by week and flag the first week, if any, where the balance would go negative. Keep all other line items identical to the base case — do not introduce new assumptions.

Constraining AI to vary only the assumptions you specify, rather than letting it improvise a "worst case," is what keeps the scenarios useful. A worst case that changes ten variables at once tells you nothing about which lever actually matters.

The assumption review — the part that matters most

A forecast is only as good as its assumptions, and AI cannot validate those against reality — it can only apply the logic you give it consistently. Before relying on any AI-assisted cash forecast, check three things:

Collection days against actuals. Pull the last two quarters of actual days-to-collect by customer segment and compare against what the model assumes. Optimistic collection assumptions are the single most common cause of cash forecasts that miss.

Lumpy items, not averages. Payroll, tax payments, and large supplier invoices don't smooth out into a trailing average — they land on specific weeks. Make sure these are placed on the actual week they're due, not spread evenly.

The worst case is genuinely uncomfortable. If your worst-case scenario still shows a comfortable cash position, the assumption ranges are too narrow to be useful for the conversation that actually matters — what happens if collections slow and a major customer pays late in the same quarter.

Where this fits in your broader AI workflow

Cash flow forecasting sits alongside the other Excel-heavy, assumption-driven work covered in Module 3 of the AI for Controllers course — the same calibration discipline that applies to DCF modeling and variance analysis applies here: build once, validate against history, then trust the structure while continuing to own every assumption.

For the data handling and review-control framework that should sit around any AI-assisted forecast involving real cash positions, see Building a Claude System for Your Finance Team.