How to forecast your cash flow with AI and see the dips weeks ahead

TLDR: Profitable businesses also die of cash, and almost always for the same reason: the dip was seen late. The thirteen-week forecast built on real data (the up-to-date books in Holded) and played through scenarios with the assistant turns the surprise into a warning with weeks of margin. The honesty rule: the forecast is a working scenario, not a certainty, and significant financing decisions get made with professional advice.

This guide is for anyone running an SMB and living cash flow as a series of jolts: the month that squares by miracle, the big payment forgotten, the client who pays late and drags everything. Forecasting does not remove the dips: it announces them with the margin that turns a problem into management.

1. Start from real data, not wishes

The forecast is worth its data, and the data is your up-to-date books: issued invoices with their real due dates (not the theoretical ones: the ones your clients actually honor), committed payments with their dates, the monthly fixed costs (payroll, rent, subscriptions, taxes with their calendar) and the current reconciled balance.

With Holded and the previous guides’ flow (digitization and reconciliation), that data exists without gathering: the forecast builds on what is already there.

The data point that breaks the most forecasts is the theoretical due date: the client paying at 90 days though the invoice says 30. The honest forecast uses real per-client timelines, which your collection history already knows.

2. Build the thirteen-week forecast

The SMB’s useful horizon is thirteen weeks (the rolling quarter): far enough for solutions to fit, close enough for the numbers to mean something. The structure is a simple table: weeks in columns, and in rows the expected collections (per invoice, with a realistic date), the committed payments and the resulting balance week by week.

The balance line is the film: where it drops below your minimum cushion is the dip, and the dip’s date is the datum that changes everything, because it defines how much margin you have to act.

The forecast updates in the monthly ritual (fortnightly in tense stretches): the new comes in, the dates reality moved get corrected, and the line gets looked at again. January’s static forecast is worthless in March.

3. Play scenarios with the assistant

The single forecast is fragile: the value is in scenarios, and the assistant makes them conversational. YesChat (from a free plan) takes your table and plain-language questions: what if the big client pays 30 days late? What if the month’s sales drop 20%? Which is the quarter’s tensest week and what causes it?

The three scenarios worth deciding on: the probable (realistic dates), the reasonable pessimistic (usual delays plus one setback) and the one specific to your biggest risk (that client who is 40% of collections). If the reasonable pessimistic holds, you sleep. If not, step 4 starts today.

The usual verification rule: the assistant’s calculations on your numbers get checked before deciding on them, and sensitive data travels aggregated per your policy.

4. Turn the forecast into alerts and decisions

The forecast that triggers no action is a pretty chart. The system closes with thresholds: the projected balance below the cushion at X weeks triggers the protocol, and the protocol gets written before the dip, with a cool head.

The early-margin classic menu: bring collections forward (invoice the invoiceable now, the reminders from the collections guide, negotiate advances), move payments (the supplier conversation six weeks ahead is friendly, the six-day one is pleading) and, if financing is due, start it with margin, with the serious nuance: relevant financing decisions are made with your adviser, because comparing financial products and their fine print is professional territory.

The forecast shared on the monthly page (with Gamma’s formatting and the reports structure) also aligns partners and team on reality. The whole sector lives in AI for accounting and finance.

Common mistakes

Forecasting on theoretical due dates. The forecast where everyone pays on time is optimistic fiction: real per-client timelines are the honest base.

Watching only the probable scenario. The dip lives in the reasonable pessimistic, and seeing it there in time is exactly what the forecast exists for.

Updating the forecast once the problem exists. The forecast is useful while it is routine: the monthly ritual with fresh data is what keeps the early warning working.

Treating the forecast as certainty. It is a working scenario inheriting assumptions: it serves to anticipate and decide with margin, not to make promises on. The big decisions carry advice.

Frequently asked questions

Can AI predict my cash flow by itself?

It can project on your data and play scenarios with you, which is already the bulk of the value. What it does not know is what is not in the data: the order that falls through, the new client. That is why it is a working forecast, not a crystal ball.

How often do I update the forecast?

Monthly as routine, fortnightly or weekly in tense stretches. The insufficient-frequency signal is reality surprising you ahead of the table.

What minimum cushion should I set?

It depends on your volatility and fixed costs: the usual reference runs from one to three months of fixed expenses. Setting it explicitly (with your adviser if in doubt) is what lets the threshold trigger alerts.

Does it help when asking the bank for financing?

A serious, updated forecast is exactly what the bank wants to see, and arriving with it worked speeds any conversation. Choosing the specific financial product, with your adviser: the fine print there is not something guides replace.