Explaining the cash forecast: AI-drafted variance commentary

A rolling cash forecast is only useful if someone explains, every week, why actual cash differed from the plan. That explanation takes hours of digging through bank activity and receivables. AI can do the tracing and write a first draft, leaving the finance lead to correct and sign it.

01the problem and who owns it

Many mid-sized companies keep a rolling cash forecast, often thirteen weeks, in a spreadsheet. Updating the numbers is manageable; explaining the misses is the slow part. Was the shortfall a big customer paying late, a timing shift in payroll, or an unplanned purchase? Answering means reconciling bank lines to forecast categories by hand.

The CFO owns the forecast and the story told to the CEO, the board, or the bank. When the commentary is thin, leadership loses confidence in the forecast, and the forecast stops being used for decisions.

02what the AI does, step by step

1. Load actuals by categoryBank transactions for the week are pulled and categorized into the same lines the forecast uses, such as customer receipts, payroll, rent, vendor payments, debt service, and taxes, using rules plus a model for ambiguous descriptions.
2. Compute variancesEach forecast line is compared with actual for the week and cumulatively. Variances above thresholds the CFO sets are selected for explanation; small ones are rolled into an other line.
3. Trace each varianceFor receipts, the system checks which expected customer payments arrived, arrived early, or did not, using AR data. For disbursements, it finds unplanned payments and timing shifts by comparing bills scheduled versus paid.
4. Classify timing versus permanentEach variance is labeled as timing, which should reverse in a later week, or permanent, which changes the forecast. Timing items carry an expected reversal week so they can be checked later.
5. Draft the commentaryA model writes a short narrative: the headline cash position, the three or four variances that matter with their causes, and the forecast changes they imply. Every sentence links to the transactions or invoices it relies on.
6. Roll the forecast forwardSuggested updates to future weeks, based on permanent variances and new information such as a revised payment date, are listed for the finance lead to accept or reject.

03systems it connects to

• Bank data. Bank portals, data feeds, or aggregators such as Plaid for transaction detail.
• Accounting system. Receivables and payables detail from QuickBooks, Xero, NetSuite, or similar.
• Forecast model. The existing spreadsheet in Excel or Google Sheets, or an FP&A tool.
• Distribution. Email or a shared document where the weekly update is circulated.

04human checkpoints

• CFO edits and sends. The commentary is a draft until the finance lead revises and approves it; nothing goes to lenders or the board automatically.
• Forecast changes. Updates to future weeks are accepted line by line by the forecast owner.
• Category rules. Finance reviews new categorization rules before they apply, since a wrong rule repeats every week.

05what to measure

• Forecast accuracy. Forecast versus actual cash by week and by horizon, tracked over time.
• Explained variance. Share of total variance traced to specific causes rather than left unexplained.
• Timing items reversed as expected. Whether variances labeled as timing actually reverse in the predicted week.
• Prep time. Hours the finance lead spends on the weekly update, compared with the baseline.

06risks and guardrails

• False confidence. A fluent explanation can be wrong. Linking each claim to transactions lets the reviewer check, and unlinked claims should not survive editing.
• Bank data access. Read-only access is all this needs. Never grant payment initiation rights to the integration.
• Confidential information. Cash position, lender covenants, and customer payment behavior are sensitive. Keep the workflow in an environment with access controls and no model training on inputs.

07build vs buy

Cash forecasting tools such as Agicap, Cashflow Frog, and treasury modules in larger platforms automate bank categorization and forecast versus actual views. If you are starting a forecast from scratch, one of these may be the faster route.

A custom layer fits finance teams that already have a trusted spreadsheet model and do not want to replace it, but want the tracing and drafting done for them, or that need explanations drawing on AR, AP, and operational data those tools do not see.

08related playbooks

Browse every finance and accounting playbook or the full library.