The cash forecast is due. Your team is chasing inputs, leadership wants an AI success story, and you have approximately eleven minutes to become the finance department of the future.
“Copilot, forecast our cash flow.” Tempting, right? It can certainly produce something that looks like a forecast. The awkward part comes when someone asks where the numbers came from.
One of the questions I get most often is how to forecast cash flow. People used to ask about Python; now they ask about Copilot. I understand the appeal, especially for an FP&A manager who is already overwhelmed and wondering what AI means for their own job.
My concern is the leap from “AI generated these numbers” to “we have a forecast we can defend.” There is quite a bit of work hiding between those sentences.
A convincing answer still needs a basis
A table of plausible numbers, a currency format, and a tasteful chart can make almost anything look ready for the board meeting. Unfortunately, none of those tells you whether your customers will pay.
An unexplained AI forecast risks becoming RANDBETWEEN() with a better vocabulary.
Even when the numbers look plausible and the explanation sounds professional, you still need to establish why either deserves your confidence.
That is the point behind my cash-flow post on X:
For roughly the same reason, you probably wouldn't use RANDBETWEEN() to come up with next quarter's cash forecast.
— George Mount (@gjmount) September 30, 2026
The “generative” part is doing some work here
That comparison is about accepting an unexplained number, rather than how Copilot calculates. It can use your workbook, generate formulas, and help construct a real model. Microsoft documents those workbook-building capabilities; availability depends on your subscription and Excel setup.
But generative AI has room to produce different responses and proposed approaches. Microsoft describes these systems as probabilistic. Leave the method, missing inputs, and assumptions unspecified, and you leave consequential choices for the system to make.
It might extend a historical trend, suggest a collection assumption, or generate a convincing explanation for a number. A confident paragraph does not tell you whether the underlying calculation is appropriate.
I sometimes wish we called it “creative AI.” That phrase lands differently in a finance department. Nobody hears “creative accounting” and thinks, “Wonderful, we should put that on the agenda.”
Do you want undocumented creative space and variability in your numbers, reports, and workflows?
— George Mount (@gjmount) September 29, 2026
The useful distinction is between generating a proposed model and calculating results from an agreed model. Copilot can help write a formula. Once checked, an ordinary deterministic formula calculates from its inputs without making a fresh modeling decision each time.
A better prompt leaves less room for improvisation
A customer expected to pay in six weeks will not fund payroll due in two. Your forecasting process needs that context. A chatbot does not acquire it just because you asked confidently.
Too much creative freedom
“Forecast our cash flow for the next three months using this spreadsheet.”
Start with a smaller assignment. Here is a prompt to adapt to your team’s existing process:
A more useful first request
“Review this workbook to help us prepare our cash forecast. Do not generate forecast amounts or change cells yet.
Identify the source tables, date fields, existing formulas, and documented assumptions. Cite the relevant cells. Separate what you found from what you are inferring.
List missing inputs and decisions the forecast owner needs to make. Do not invent amounts, payment dates, or explanations. Propose the next steps and wait for my review.”
You have given Copilot a job and a stopping point. You have also made it easier to catch misunderstandings before they become twelve beautifully formatted columns.
After the finance team agrees on the method and assumptions, ask for a specific formula, scenario, or documentation task. Require visible inputs and calculation logic. “Act as an expert financial modeler” is a role description, not a quality-control procedure.
Inspect the work, including the explanation
Ask where a result came from, then follow the reference yourself. Check that the cited cells exist and support the claim. Inspect generated formulas or code. Change a known input and confirm that the result responds as intended.
Keep the finance team’s existing model checks and review process. Copilot saying “I verified everything” is another generated statement to check. It has not hired an independent reviewer.
Also check whether you received a working calculation or a static output. Microsoft notes that some analysis responses contain static tables or images. A nice chart does not establish that the forecast recalculates.
Give Copilot some useful work around the forecast
You can make progress with AI without delegating the entire forecast. Pick one task with an output your team can inspect:
Prepare: Flag missing dates, inconsistent categories, and possible duplicates. Review proposed corrections before applying them.
Explain: Describe an existing formula and identify its inputs. Check the explanation against the workbook.
Present: Draft a management summary from approved results. Require cell references and keep observed changes separate from possible explanations.
Those tasks can give you more time to think about the business. Count review and correction time when judging whether Copilot helped. Your AI success story can be pleasantly boring.
Where Python in Excel earns its place
Python in Excel gives you code you can inspect, parameters you can adjust, and results you can compare. Its included pandas, NumPy, and statsmodels libraries support several useful approaches I have covered before.
Model recurring patterns. In my exponential smoothing example, Holt-Winters accounts for both trend and seasonality. That makes it a candidate for forecasting recurring receipts or drivers such as sales when the historical pattern supports it. A sales forecast still needs collection-timing assumptions before it becomes a cash forecast, and known one-off payments belong in the picture too.
Check whether a method earns its complexity. My forecast accuracy comparison holds back 12 months of retail sales, then compares Holt-Winters and ARIMA/SARIMA predictions with those actuals. Python can calculate measures such as MAE, the average absolute miss, and RMSE, which penalizes larger misses more. For a cash-flow application, use a time-ordered test and compare against a simple baseline. The model with the most impressive acronym does not automatically win.
Explore the risk of a cash shortfall. My IRR simulation example varies cash flows and makes assumptions adjustable through worksheet cells. You could adapt that approach to simulate collection delays and payment variations, then count how often projected cash falls below your chosen buffer. The Monte Carlo introduction also explains random seeds, which help reproduce a run with the same code and inputs. Your chosen distributions and relationships still determine what the simulation means.
Copilot can help draft and explain that code while your team approves the method and checks its behavior. My forthcoming book, Python in Excel for Data Analytics, develops these preparation, forecasting, and simulation skills further.
Put Copilot to work around your forecast
Download three focused prompts and four checks for reviewing Copilot’s contribution. Your finance team keeps ownership of the method and assumptions.
Give your team a useful place to start
I help teams choose practical Copilot and Python in Excel tasks, write clearer prompts, and review what comes back. Bring the business knowledge, and we can work on the AI skills together.