Cash-flow forecasting a CEO actually trusts

A rolling 13-week view of the cash coming in and going out, built from your own receivables and payables, that you can stress-test with a question: "what happens if our biggest customer pays 30 days late?"

What you'll have when you're done

A forward cash picture you believe, because you built it from your real numbers and your own assumptions, not a black-box app that spits out a line you cannot interrogate. Each week you refresh three exports, the AI rebuilds the 13-week model, and you can immediately ask the what-if questions that actually keep you up at night. The output is a table you paste into a spreadsheet and a set of answers you can act on.

The past is easy to see. The next 13 weeks are where I get nervous

My accounting software is excellent at telling me what already happened. What it does not do well is answer the question I actually carry around: are we going to be fine on cash in eight weeks, or is there a squeeze coming that I should get ahead of now? The forward view is the one that matters for decisions, hiring, a big purchase, whether to chase a slow-paying customer harder, and it is the one most CEOs run on gut.

The 13-week cash flow is the standard operator tool for this, and it is exactly the kind of thing AI builds well: take your known inflows and outflows, apply payment-timing assumptions, project the rolling balance, and let you flex the assumptions. The reason to build it with AI rather than buy an app is trust. When you can see every assumption and change it, you believe the output. When a tool hands you a number you cannot trace, you do not.

What you need first

• Three exports from your accounting system: your accounts-receivable aging (who owes you, how overdue), your accounts-payable aging (who you owe), and your current bank balance.
• Your known recurring items: payroll dates and amounts, rent, loan payments, recurring revenue, anything that lands on a predictable schedule.
• Your payment-timing assumptions: roughly when customers in each aging bucket actually pay you, and when you pay your bills.
• A Claude Project on a business plan (keep financials off personal accounts).

Step-by-step

Step 1Export your AR aging, AP aging, and bank balance

Pull the three reports as clean files. The AR and AP aging reports are the engine of the forecast: AR aging tells the model what cash is likely coming in and roughly when, AP aging tells it what is going out. Grab your current bank balance as the starting point.

Step 2Give the model your assumptions explicitly

This is the step that makes the forecast trustworthy. In your Project, spell out the timing assumptions the model must not invent:

Assumptions to use (do not guess these):
- Customers in the "current" AR bucket pay in ~15 days; 1-30 late bucket in ~25;
  31-60 in ~50; over 90 treat as 30% likely to collect this quarter.
- We pay vendor bills on their due date unless I say otherwise.
- Payroll: $X on the 15th and last day of each month.
- Recurring revenue: $Y lands on the 1st.
Flag any assumption you are applying so I can see it. Never invent a payment date.

The "flag any assumption" instruction is the whole game. You want to see the model's working so you can correct it.

Step 3Build the rolling 13-week model

Prompt it to produce the forecast as a table:

Using the AR aging, AP aging, starting bank balance, and the assumptions above,
build a rolling 13-week cash-flow forecast. For each week show: starting balance,
expected inflows, expected outflows, net, ending balance. Call out any week the
ending balance drops below $[your comfort floor].

What comes back is a week-by-week cash projection you can paste straight into Excel or Google Sheets. Illustrative, the first six of the thirteen weeks, starting at a $180K balance:

Week	Start	Inflows	Outflows	Net	End
1	$180K	$62K	$88K	-$26K	$154K
2	$154K	$41K	$35K	+$6K	$160K
3	$160K	$28K	$97K (payroll)	-$69K	$91K
4	$91K	$74K	$40K	+$34K	$125K
5	$125K	$19K	$96K (payroll)	-$77K	$48K
6	$48K	$88K	$52K	+$36K	$84K

⚠️ Week 5 ends at $48K, below your $60K comfort floor. The squeeze is the payroll run landing before the $74K from your largest customer clears. Assumption applied: that customer pays in about 25 days (the 1-30-day-late bucket). Change that assumption and week 5 changes.

The value is not the table, it is that last line. Every number traces to an assumption you can see and argue with, which is exactly what a black-box forecasting app will not show you.

Step 4Stress-test the assumptions

Now ask the questions you actually worry about. This is where a forecast you built beats one you bought:

- What happens if our largest customer (the $X invoice) pays 30 days later than assumed?
- What if we lose the recurring revenue line for two of these weeks?
- If I make a $50K equipment purchase in week 4, which weeks go tight?

Each what-if re-runs in seconds. You are not forecasting once; you are pressure-testing the next quarter of cash against the scenarios that matter.

Take the week-5 squeeze from the table above and run the first what-if: "what if my largest customer's $74K invoice pays 30 days later than assumed?" The model moves that inflow from week 4 to week 8 and re-runs the rolling balance. Now week 4 ends at $51K and week 5 bottoms out at -$29K. You would overdraw. But you are seeing it in week 2, with three weeks of runway, so it is a decision and not an emergency: call the customer now and get a commit date, draw $40K on the line of credit, or slide the equipment purchase a month. Any of those is calm when you see it coming. All of them are a 6am scramble when the bank balance is the thing that tells you, in week 5, that it already happened.

Step 5Refresh weekly

A 13-week forecast is a rolling tool. Once a week, re-export the three reports, drop them in, and let it rebuild. Reality drifts from assumptions, customers pay early or late, so the weekly refresh is what keeps it honest. Put a recurring 15-minute block on your calendar for it.

Make it yours. The 13-week window suits most operating businesses, but the horizon should match how far ahead your decisions reach. A seasonal business (retail, anything that lives or dies on Q4) wants the model extended to a full year so the slow build and the big draw-down are both in view. A long-sales-cycle business where a single deal swings the quarter should model named deals individually, each with its own probability and expected close, rather than lumping them into an aging bucket. Tell the Project your shape once and it adjusts the model to it.

How you'll know it's working

You stop being surprised by cash. The tight week shows up eight weeks out, while you still have options, instead of arriving as a scramble. You also start making cleaner decisions, the "can we afford this hire / purchase / pause" question gets a grounded answer in minutes instead of a gut call.

When it breaks

• The projection feels wrong. Almost always the payment-timing assumptions. They are the model's biggest lever; tune them to how your customers actually pay, not how you wish they did.
• It invented a payment date. Restate the "never invent a payment date, flag every assumption" rule. You should be able to see every timing assumption it applied.
• The aging buckets do not match reality. Old uncollectible invoices sitting in AR will inflate expected inflows. Tell it how to treat the over-90 bucket (often: heavily discount or exclude).
• It is treating the forecast as a commitment. It is a projection, not a promise. Label it that way in any version you share, and never present a forecast as a guaranteed number.
• Two weekly refreshes give wildly different pictures. Usually a few large invoices crossing aging buckets between exports. Lumpy receivables make any honest forecast jumpy; that is real, not a bug. Smooth it by asking the model to break out your top three to five invoices individually so you can see exactly which one moved.
• It assumes you can draw on credit you do not have. It will not model your line of credit unless you tell it. Give it your available credit and a rule ("if a week would end below $40K, assume a draw to $60K and show the balance owed"), or it will happily project you straight through zero.

Where this fits in your harness

This is the forward-looking end of the finance chain. It works best when your books are closed cleanly (so the aging reports are accurate) and once you are comfortable interrogating your P&L. The same finance Project that answers backward-looking questions can project forward, then draft the board summary that explains both.