Cash flow forecasting with AI — from Excel-13-weeks to rolling forecast

Cash flow forecasting with AI means setting up a rolling 13-week forecast where AI handles routine updates — AR balances from Exact, payment calendars, recurring costs — while assumptions on new sales and exceptional items stay with the CFO or controller. For finance teams wanting a Friday-current cash forecast that's still usable on Wednesday: the difference between half-a-day-per-week maintenance and always current.

Almost every CFO sitting on a scale-up or PE-portfolio company knows the problem: a 13-week cash forecast current on Friday is already stale on Monday and not trustworthy enough to bring to the board by Wednesday. The forecast is maintained in Excel, fed with data from Exact, AR balances, payment calendars, and assumptions on sales — and exactly that combination makes it a piece of work costing someone half a day once a week, slowly decaying until the next update.

AI doesn't change that rhythm by "doing" the forecast, but by making the update work running behind it substantially cheaper. Deployed well that means: a forecast current every morning, with scenarios adjusted in minutes — not hours. Deployed poorly it means: a forecast with impressive charts that is actually nonsense because the assumptions weren't judged by a human. The difference sits in how you set it up.

Status quo — why Excel cash forecasts grind to a halt

A 13-week rolling forecast consists roughly of four blocks: opening balance, inbound cash flows (AR, advance invoicing, financing), outbound cash flows (AP, salaries, fixed costs, taxes), and ending balance per week. Building it isn't rocket science; maintaining it is.

The pain sits in four things.

The inbound side changes faster than the Excel. A debtor pays three days earlier than expected, a large customer shifts a payment, an advance invoice doesn't land. The Excel only knows that when someone updates it manually.

The outbound side sits scattered across systems. Open AP in Exact, lease contracts in a spreadsheet, VAT returns in the planning, salaries from HR. A weekly update costs an hour or more of data gathering.

Scenarios are manual copy work. "What happens if that big payment comes a month later?" — in Excel that means copying a tab, changing a number of cells, and hoping no formula reference broke. Building three scenarios takes half a day.

The forecast forgets itself. If actual cash flow deviates from the forecast, the next update isn't automatically instructive — nobody reaches back to "what did I expect last week and what was different?"

Result: a forecast that's reasonable for the first two weeks, fuzzy for weeks 3-6, and nominally present but practically not decision material for weeks 7-13. CFOs don't ignore that deliberately — they simply don't have the time to keep it current.

What AI can and cannot do in a cash forecast

AI is good at:

• Pattern extraction from historical data. Which customers typically pay on which terms? Which suppliers collect on which due date? Which months are structurally heavier or lighter? A human can do this too, but less structurally and not across hundreds of customers at once.
• Linking open items to the forecast. Combining the open AR list and the expected payment behavior per customer into a date-specific inflow.
• Scenario generation. Give an assumption ("customer X pays 4 weeks later," "we sign no new contracts in weeks 4-8") and the AI calculates the impact on the cash path without anyone copying cells.
• Variance analysis on the forecast itself. Last week we predicted inflow X, it became Y — which difference comes from which customer, which assumption was wrong, how do you adjust the pattern?

AI is bad at (and should stay bad at):

• One-off events. A refinancing that might go ahead, an M&A deal whose timing is in negotiation, a won or lost large customer. These are punctual decisions the human knows and the AI cannot predict from data.
• Strategic cash choices. Whether to pay a dividend, whether to pull an investment forward, whether to pay a supplier faster for a discount — that's CFO work, not forecast work.
• Assessing whether the forecast is internally consistent with strategy. A forecast showing cash getting tight but not accounting for planned capex or an upcoming funding round is technically correct and operationally worthless.

What an AI-assisted cash forecast looks like

Per component: what the AI does, what the human does, and where the cut point sits.

Inbound flows — AR and advance invoicing

AI: Pulls the current open AR list, links per debtor the historical payment pattern (average payment term, variability, reliability), and proposes per open invoice an expected payment date. Aggregates that per week. Human: Controller or CFO looks at the aggregation, overrides per exception (we know customer X won't pay until late next week, despite the term). Looks at the overall confidence level. HITL: On manual overrides. The base forecast can update autonomously daily.

Outbound flows — AP, salaries, fixed costs

AI: Pulls open AP from Exact, adds fixed-cost patterns (rent, subscriptions, leases, salaries, social contributions, corporate-tax provisional assessments), and places them in the right week. Human: Controller verifies that all fixed costs are included and that special expenses (bonuses, capex, one-off investments) are added manually. HITL: On manual additions.

Scenario analysis

AI: On the user's request — "what if customer X pays three weeks later," "what if revenue comes in 15% lower," "what if we push the funding round one month" — generates a new cash path within seconds with deltas and critical points. Human: CFO or controller names the scenario, judges the outcome, takes it into decision-making. HITL: N/A — these are analyses, not decisions.

Periodic variance analysis

AI: Compares the current cash position to the forecast of a week ago, decomposes the difference into specific debtors, creditors, or categories, and drafts a commentary. Human: Controller reads the commentary, confirms or adjusts, uses it as feedback for the assumptions in the next update. HITL: On the commentary for distribution.

Communication to CFO or board

AI: Generates a one-page summary of the cash position: current balance, projected balance at month-end, critical weeks in the horizon, summary commentary of the biggest forecast moves. Human: CFO reads, adds strategic context, approves. HITL: Before distribution to board or investors.

The overconfidence risk

A specific risk with AI cash forecasts you have to address actively: a forecast that looks good gets disproportionate trust — more than a manual Excel everybody knows trails reality. An AI forecast with nice charts and a personalized commentary feels like a more trustworthy document, while the underlying data is just as good or bad.

Two disciplines to counter this:

• Make the confidence interval explicit. Per week not just the expected balance but also the range (P10-P90, or a comparable measure). A forecast without a range overstates its own certainty.
• Assumptions visible in every summary. Which assumptions on payment behavior, which exceptions are added manually, which one-off events are included. A board not reading those lines isn't reading the balance properly.

Audit grade — a cash forecast also belongs in the control trail

A cash forecast isn't formal reporting in the sense of an annual report, but it is a document on which decisions are made — funding, dividend, capex timing, AR policy. That makes it a document for which you must be able to explain afterwards how it came about, especially if decisions follow from it and later come under discussion (a shareholders' dispute, a banking covenant debate, an investor asking why a forecast deviated).

Practically: keep per version of the forecast the input snapshot (which AR position, which AP position, which manual assumptions), the AI model used, and the commentary published alongside it. An AI making a new version every day without version history isn't an improvement — that's erasing evidence of how you arrived at your insight.

What this delivers

For teams that have wired AI into their cash forecasting, we see three shifts.

The forecast becomes current rather than weekly. That sounds small but changes how often the CFO wants to respond to it — a forecast holding up every morning gets used; a forecast 11 days old on Wednesday gets ignored.

Scenario work becomes incremental rather than project-like. "What if" becomes a working question in an MT meeting instead of a spreadsheet assignment for two days from now.

Variance analysis after the fact gets sharper, because the AI consistently does the same decomposition and the pattern learning accumulates over months. Not just "we're €120K short," but "€80K of that is from customer X we structurally over-estimate — adjust pattern."

Limits — when AI delivers less here

• Companies with strongly project-based cash flows (construction, M&A advisory, project engineering) whose timing per project is ad hoc. AI can support, but the pattern learning is thin because the "patterns" are one-offs.
• Companies with large, individually-negotiated customer contracts where most payment events are specifically agreed. The work stays largely manual — AI helps with the base, not the essence.
• Early-stage startups without enough historical payment data to learn patterns. Here AI doesn't make the assumptions better than the founder would.