Yes. A cash flow forecast is mostly dated amounts gathered from systems that already hold them. It prepares the forecast at depth 2, labels every line by how reliable its source is, and compares the last forecast with what happened. A treasurer reviews and owns it, and the forecast never triggers a payment by itself.
What goes into the forecast?
A direct cash forecast lists expected receipts and payments by week or day. Receipts come from the receivables ledger, adjusted for how quickly each customer actually pays. Payments come from approved supplier invoices, payroll, tax, rent, loan repayments and committed capital spend. Then there are the estimates: sales not yet invoiced, costs not yet incurred.
The agent's strength is coverage. It reads the receivables ledger, the payables ledger, the payroll calendar and the contract register, and it does not forget the quarterly tax payment or the annual insurance premium. The weekly version is described in detail in how to prepare a thirteen-week cash flow forecast with an AI agent.
How should each line be labelled?
By the strength of its source. This is the agent's most useful habit, because a forecast that looks precise but rests on guesses misleads.
- Sealed: amounts fixed in a system of record. Approved supplier invoices with due dates, the payroll file, loan schedules, scheduled direct debits.
- Signed: amounts a named counterparty has committed to. A customer remittance advice, a signed order with payment terms.
- Declared: amounts someone told you. A customer's promise to pay next week, a budget holder's estimate of spend.
- Estimated: amounts the agent projects from patterns. Receipts from customers based on their average days to pay, sales not yet made.
The treasurer then sees how much of each week's position rests on the weaker classes. A comfortable month built on estimates deserves a different response from one built on sealed amounts.
Who prepares, who decides?
| Task | Agent | Treasurer or finance lead |
|---|---|---|
| Pull ledger and calendar data | Depth 1 | None |
| Project customer receipts | Uses each customer's payment pattern | Adjusts for known issues |
| Build the forecast | Depth 2, every line labelled | Reviews |
| Forecast-to-actual comparison | Explains each material difference | Agrees changes to assumptions |
| Decisions on borrowing, investing or delaying payments | Drafts options | Decides |
Separation of duties here is about use, not preparation. The forecast informs decisions; it does not make them. If an agent could read a forecast shortfall and delay supplier payments on its own, a forecast error would become a broken supplier relationship. The decision to delay, draw on a facility or invest a surplus belongs to the treasurer.
An illustrative example
Take a construction contractor with uneven receipts. The agent builds a 13-week forecast.
Weeks one to four look tight but positive. Week six shows a shortfall of 700,000. The agent explains why: a large stage payment from a client is expected in week five, but that client has paid each of the last four stage payments about three weeks late, so the agent places it in week eight. Two subcontractor payments of 400,000 each fall in week six.
It labels the stage payment as signed, because the contract fixes the amount, but its timing as estimated. The treasurer calls the client's finance team, who confirm the payment will be late again. The treasurer arranges a short-term draw on the revolving facility for week six, and the agent adds it to the forecast as a sealed line once the facility confirms.
What controls make it safe?
- Every line labelled by source.
- Assumptions written down. Payment patterns, growth rates and timing rules are stored and versioned.
- Forecast-to-actual every period. Differences are explained, so the forecast improves.
- No payment authority. The forecast cannot change a payment date or release a payment.
- Treasurer's review recorded. With any adjustments they made.
Where Pryme fits
In Pryme, the Pryme Context Protocol attaches a provenance class to every input, so the forecast shows sealed, signed and declared amounts separately, and the agent's own projections are marked as estimates. Forecasting is a depth 2 capability, and acting on it is a different capability held for the treasurer.
Receipts are the hardest part to forecast well, which is why clean cash application and the ability to apply cash receipts to invoices quickly make the forecast better. The daily view is covered in how to prepare daily cash positions, and the wider controls in AI agents for finance: books you can audit.
Questions people ask
How far ahead can the agent forecast reliably?
The near weeks rest mostly on sealed amounts and are reliable. Beyond that, estimates dominate. Many teams pair a weekly forecast with a monthly one for longer horizons.
Can the agent run downside scenarios?
Yes. It can model late payment by major customers, lost contracts or higher costs, and show which week the business would run short in each.
Should sales forecasts feed the cash forecast?
They can, labelled as estimates. Keep them visible as a separate line so nobody mistakes a sales hope for cash.