Build a cash-flow narrative that cannot silently change the numbers

Use AI to explain an approved cash worksheet while arithmetic, dates and assumptions stay visible.

The finance manager already has a spreadsheet showing when money might arrive and when bills fall due. The difficult part is explaining the pressure points to colleagues who do not live in that spreadsheet. A language model can help draft that explanation, provided it cannot quietly replace an uncertain receipt with a certain one.

Intuit's agent announcement includes finance reporting and scenario work. Rel-Cat research illustrates AI classification over connected financial records. BIS separately cautions about opaque data and model risk. These are foundations for experimentation, not proof that a generated cash forecast is dependable. [Product announcement](https://investors.intuit.com/news-events/press-releases/detail/1258/intuit-introduces-ground-breaking-virtual-team-of-ai-agents-to-fuel-growth-for-businesses), [research](https://arxiv.org/abs/2506.09234), [BIS analysis](https://www.bis.org/publications/fsi-summary-financial-stability-implications-artificial-intelligence-executive-summary).

In this original exercise, Harbour Furniture has 18,000 in available cash. Confirmed payments due before Friday total 12,000. A customer owes 9,000, but has only said that payment is being processed. A second customer has approved a 4,000 invoice with a documented payment date next Tuesday. The business needs a description of timing, not a prediction that either customer will certainly pay.

Start with a table whose fields are deliberately unromantic: opening balance, due date, currency, amount, evidence reference and confidence category chosen by the finance team. Separate approved obligations from possible inflows. The model receives the completed table and a short instruction to describe it without calculating new totals. Arithmetic is performed in the spreadsheet and checked independently. That separation makes an incorrect number easier to locate than asking a model to reason through several screenshots.

Harbour creates three scenarios. The base case includes the documented Tuesday receipt. The delay case moves it one week later. The cautious case also excludes the undocumented 9,000 receipt. The worksheet produces ending balances for each scenario. The assistant drafts a paragraph explaining which payment timing changes the result most. It must label the cases as assumptions and refer to row identifiers. It cannot announce that a customer is financially distressed because one invoice is late.

A manager might ask whether the company should postpone buying a saw. The assistant can organise the relevant facts: purchase date, cancellation terms, cash requirement and operational consequence. The purchase decision belongs to the people responsible for the business. It should not automatically choose borrowing, recommend an investment product or transfer money. This article concerns educational operations planning, rather than personalised financial advice.

An overlooked failure is mixing cash and accounting profit. A profitable order can require materials before the customer pays. Harbour adds an example in which a 6,000 deposit arrives this month for work delivered later. The reviewer checks whether the narrative describes cash received while avoiding unsupported statements about revenue recognition. Another test uses a supplier invoice denominated in another currency. An exchange assumption must come from the worksheet with its date; the model must not invent a live conversion.

Give the reviewer a useful comparison. Show the original worksheet, generated paragraph and every amount mentioned in the paragraph. A simple consistency check can flag numbers absent from the source table. It cannot detect every mistaken interpretation, but it helps a reader focus. Any unsupported certainty, such as payment will arrive tomorrow, returns the draft for correction. An uncertainty is valuable information, not a blemish to remove for a smoother presentation.

The operating benefit can be tested with colleagues. Ask them to identify the week of greatest pressure and the assumption most likely to change it, first from the worksheet and then from the reviewed explanation. Record misunderstanding as well as time. A faster answer that misses the delayed receipt is a failure. Keep several awkward examples, including refunds, restricted cash and invoices paid in instalments, for future evaluations.

Costs include preparing reliable inputs and maintaining the link between the narrative and the approved worksheet. A paid tool may also require a particular subscription or region; the cited announcement is historical availability evidence. No current price or account entitlement is established here. A finance team should ask whether the added explanation helps an actual meeting enough to justify those costs.

The final paragraph should carry the worksheet's version and preparation date. When a receipt changes, refresh both together. The assistant then has a clear role: turn reviewed assumptions into a readable account of choices, with the facts still close enough for everyone to inspect.