Seeing cash flow 90 days out

“Will we be short of cash in the next three months?” In a mid-sized business this question is usually answered with a spreadsheet and an instinct. The file is updated in the first week of the month, and by mid-month reality has moved away from it.

Here is the interesting part: nearly all the data needed to answer it is already inside the ERP. What is missing is not a model — it is that the data has never been brought together.

The three inputs of a projection

A cash projection is made of three things. First, the certain ones: receivables with known due dates, payables with known due dates, known fixed costs. These are not a forecast, they are a calendar.

Second, the probable ones: when open orders will be invoiced, and whether invoices will be collected on time. Past behaviour governs here — and this is where a model genuinely contributes.

Third, the unknowns: new orders, unexpected costs. Rather than pretending to forecast these, it is more honest to leave them as a range.

Collection risk is a customer attribute

Assuming every receivable is collected on its due date makes a projection systematically optimistic. In practice customer payment behaviour is stable: some always pay on time, some are always ten days late, some slip further as the order gets bigger.

A simple delay profile built from collection history changes projection accuracy noticeably. No elaborate model is required; what is required is holding the data per customer and not collapsing it into an average.

One caution: this score is not a credit decision. Rather than labelling a customer “risky,” it is enough to place that receivable slightly later in the projection.

Abnormal costs, the hidden enemy of a forecast

The other source of drift is one-off, large costs. A forecast that looks at historical averages flattens these and misses the real troughs inside a month.

Anomaly detection helps here: when a cost line that does not fit its own pattern is flagged, the finance team can place it into the projection by hand. The model is not deciding; it is only saying “look at this.”

Why 90 days

Thirty days is already visible in most businesses — no projection needed. A year cannot be predicted to any useful accuracy; that is a budgeting exercise, not a projection.

Ninety days is the longest horizon you can still act on: enough time to have a collection conversation, reschedule a payment, or agree a credit line in advance. The finance module builds the projection on that horizon.

In closing

Cash flow projection is one of AI’s least flashy and most concrete contributions. Unflashy, because there is no impressive chart. Concrete, because when it is set up properly the question changes from “will we be short?” to “which week, by how much, and what do we do now?”