A 13-week cash flow forecast was the first AI use case I ever built back in 2024. It is still the request I get most often from my clients, and the one most likely to fail on the first attempt. Not because building a cash flow forecast is technically hard, but because every company's data is put together differently, and most attempts skip straight to building it without ever checking whether the data can support it.
This week's edition is about that gap. I talk about the specific ways I've watched cash flow forecasts fall apart in practice, and the diagnostic sequence I run before touching a single formula.
For paid subscribers, I'm sharing the Claude skill I use to run that diagnostic: what it checks, how it scores confidence, how it ranks what to go collect next, and a prompt to adjust it to whatever type of company you actually work with, whether that's one industry or a whole range of clients like mine.
EXCITING NEWS!
Some news I'm excited to share: I'm joining VAi Consulting as Practice Lead. VAi is a Practice Lead-supported network of 82 independent consultants and firm principals delivering advisory and implementation work across 14 practice areas and 13 industry verticals, including a dedicated Finance, FP&A & Accounting practice.
I'll keep operating independently through Blend2Balance, and VAi opens the door to larger-scale projects that call for a bigger network behind them.
More details on my specific focus are coming soon, but I wanted you to hear it here first.
Why Cash Forecasting Is So Hard
A 13-week cash flow forecast, or an updated version of it, is often the first thing built when a new CFO or Fractional CFO steps into the role.
Using AI to construct one seems straightforward until you attempt it in practice. Simply gathering historical bank statements, AR, and AP reports rarely works. The underlying patterns and structural dependencies run far deeper. That complexity is precisely why it remains one of the most difficult AI builds to execute successfully.
I failed more than I succeeded with this task. So after all those failures, I changed the approach and shifted where the AI work actually starts. Instead of jumping straight into building the forecast, the first step is to identify trends and run diagnostics.
Before we get into why, take a guess. The chart below shows thirteen weeks of inflow, outflow, and cash balance for four real types of companies I have worked with:
Established service business,
SaaS subscription-based company,
Venture-backed company,
Product business with seasonal swings.
See if you can match each letter to a company type. And think about the implications for the Cash Flow forecast.
Here is what actually breaks a forecast, based on the engagements where I had to go back and rebuild the assumptions:
Accounts receivable that is functionally uncollectible but never written off. The forecast counts on cash that was never coming in the first place.
Payment terms on paper that do not match reality. A client states net 30, but the actual pattern runs 45 to 60 days, so every outflow and inflow tied to that vendor or customer lands in the wrong week.
Payroll debits that bundle wages and tax remittance into a single bank line. What looks like one cash event is actually two, and three payroll months get missed entirely.
Annual costs that get miscategorized as one-off noise. An insurance renewal or a software license does not look recurring in twelve months of bank history, so it disappears from the forward view until it lands.
Non-operating cash treated as certain when the timing is not. A raise, a loan draw, or a grant with a soft close date can move a forecast by weeks.
A credit line or covenant that never came up until the numbers stopped working. The forecast is missing a lever the client already had.
What all of that means in real life is that, from week 1, you have a forecast that doesn't hold up, and over time you’ll see your team stop updating it (or looking into it).
So here’s how I handle this instead:
Start by collecting what the diagnostic needs at minimum:
Twelve months of bank statements, every operating account.
Current accounts receivable aging.
Current accounts payable aging.
Debt, lease, and loan schedules.
Credit line terms and covenants, if the company has one.
With that in hand, run the diagnostic itself:
Check data quality: all accounts, all twelve months, and enough of each transaction description to actually categorize it.
Categorize everything and roll it up to weekly totals, so patterns show up instead of noise.
Profile the company: what drives the cash, how concentrated it is in a handful of customers or vendors, and how AR and AP actually behave versus what the stated terms say.
Look at everything else that moves cash: recurring outflows outside AP, non-operating cash like a raise or a loan draw, the minimum balance, and any credit line.
Build the visibility curve: how much of the next thirteen weeks is already backed by known information versus estimates.
If that first pass does not produce a strong enough verdict, here is the additional information that tends to move it the most:
Monthly AR and AP aging snapshots going back 6 to 12 months, not just the current one.
An invoice register with issue, due, and paid dates, so actual payment behavior can be measured directly instead of inferred.
Payroll provider reports, so wages and tax remittance can be split out from a single bank debit.
GL detail and chart of accounts, to categorize with more precision than bank descriptions allow.
Major customer and vendor contracts, since a handful of relationships often drive most of the cash.
The next step is to go and collect whatever is actually missing, then run the diagnostic again. Repeat that loop until the verdict on data quality is good enough to build the thirteen-week forecast on top of it.
And yes, you can totally use your favorite LLM to run this analysis if your setup allows safe data sharing with the tool.
That sequence takes longer than opening a spreadsheet template and typing in numbers, so it gets skipped more often than it should. It is also the difference between a forecast a CFO can actually stand behind and one that looks complete for about two weeks.
Here are the answers to the chart above:
3: Company A: the venture-backed company with no revenue, a steady, moderate burn with only minimal offsetting inflow.
1: Company B: the established service business, small weekly swings around payroll dates, balance holding roughly steady overall.
2: Company C: the SaaS company, a smoother upward trend with two bumps around quarterly renewals.
4: Company D: the product business, drawing cash down to build inventory ahead of the season, then rebuilding it as sales come in.
Four different shapes, four different sets of assumptions, and the same starting question every time: can this company's data actually support a forecast worth trusting?
We've covered the shapes, the reasons a forecast breaks, and the sequence I run before touching a formula.
In the subscriber section, we get into the exact mechanics behind that sequence: the specific rule that decides each verdict, how confidence gets scored, and the formula that ranks what to go get next.
Claude skill that runs all of this is included as a download, ready to use.
Closing Thoughts
That's this week's edition. Next week, we build the actual 13-week forecast once the diagnostic says the data is good enough to work with. If you run the diagnostic on one of your own companies before then, I'd like to know what worked and what didn't, reply and tell me.
Thanks for reading, and I'll see you back here next Tuesday.
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