A 13-week cash flow forecast was the first AI use case I ever built, and it is still the one most likely to fail on the first attempt with a new client, because every company needs its own version. Last week, I shared the diagnostic I run before any build. This week is Part 2: the build itself, and how to keep the forecast improving.
In the free section, I look at what the forecast has to do in four types of business, lay out the sequence I follow, and open the cash flow diagnostic skill to everyone, with a short guide to running it. In the subscriber section, you get the prompt I use to build the first version of the forecast, plus a new skill that reviews it against actual cash every week and recommends changes.
What The Lunch Version Leaves Out
Last week's edition on cash flow diagnostics got more replies than most. The same reaction kept coming back: building a 13-week forecast is hard, and checking the data before the build helps. Many of you also said you are tired of posts that make it look easy. One reader put it this way:
"I also get frustrated with those posts. One client actually asked me why it took me so long to build their 13-week forecast when he said ChatGPT could do it for him over lunch."
I have had versions of that conversation too. A forecast over lunch is possible for a simple business with clean data, as long as nobody compares week six to what actually happened. The trouble starts when you put different businesses side by side and look at what each forecast has to do.
Four businesses, four different forecasts
Last week I showed four cash balance charts and asked you to guess which company was which. Here is what the forecast looks like for each of them.
These four are a small part of the variety I see in client work. Each has its own blind spots and risks, and each starts with a different set of information; inside every type, each company has its own version of both. So how would one prompt or one skill cover all of that, and how would anyone build it during a lunch break? The build is yours to do, for your own business, and here is how I would approach it.
How to build yours
The build is a sequence, and each step makes the next one easier.
Run the diagnostic from last week’s newsletter (see below). It tells you what your data can support, so you know which parts of the forecast will be solid and which will be estimates.
Understand the gaps. Look at what is missing and how much each gap matters to the forecast.
Collect the information that is available. Much of it sits outside finance: the invoicing outlook with sales or delivery leads, the purchase plan with operations or production, the hiring plan with leadership. Ask for a date and an amount for each item, since a general request for updates rarely gets you numbers.
State your assumptions clearly. Where information is missing, write down what you assumed and why, and keep that list next to the forecast.
Track what replaces each assumption. As answers come in, swap the assumption for the real number and note where it came from.
Create an improvement loop. Compare the forecast to actuals every week and feed what you learn back into the assumptions.
This is a step-by-step process, so do not expect to have everything at once; the first version will have gaps, and each week closes some of them.
Last week, paid subscribers got the cash flow diagnostic skill I use with clients. This week I am opening it to everyone, since every other step depends on it.
Download it here:
A short guide to running it:
Add the skill to Claude. Download the file and upload it: Customize→Skills. It runs Python to build its outputs, so it needs a plan with skills and code execution turned on.
Collect the inputs: twelve months of bank statements for every operating account, current AR and AP aging, debt, lease and loan schedules, and credit line terms if you have a credit line. If something is missing, the skill asks for it before it starts.
Start a new chat or project, upload the files, and ask Claude to run the cash flow diagnostic for your company as of your forecast start date. Mention any financing you expect in the next thirteen weeks.
Read the dashboard: a verdict, a confidence level, the visibility curve showing how much of each week is backed by known items, and the three to five missing pieces that would improve the forecast most.
Go get the top items on that list and run it again. The Excel data file it produces becomes the starting point for your build.
The diagnostic tells you where you stand and what is missing. The rest of the process is the build and the weekly loop, and in the subscriber section I share both: the prompt I use to build the first version of the forecast and a skill that reviews it against actual cash every week and recommends changes.
Closing Thoughts
Thank you to everyone who replied to Part 1; this edition started with your messages. If you run the diagnostic or the review on your own numbers, I would like to hear what it found and where it got things wrong, since that is how the next version of both skills gets better.
We Want Your Feedback!
This newsletter is for you, and we want to make it as valuable as possible. Please reply to this email with your questions, comments, or topics you'd like to see covered in future issues. Your input shapes our content!
Want to dive deeper into balanced AI adoption for your finance team? Or do you want to hire an AI-powered CFO? Book a consultation!
Did you find this newsletter helpful? Forward it to a colleague who might benefit!
Until next Tuesday, keep balancing!
Anna Tiomina
AI-Powered CFO