$50k every 20 days

Let me tell you something, the average cost of unreliable cash flow forecasts is about $465,000 every year.

And what is worse, most finance teams rely on unreliable forecasts, with cash deficits of over $50,000 every 20 days on average! (Source: Agicap Survey).

After leading finance at Thales (10k+ employees) for 7 years, the big problem was always the time it took to forecast. Plus, the fact that the forecast was already wrong by the time we needed to make decisions.

Juergen Lang, (an expert in my AI Finance Club) decided to fix this.

He built a Python-based forecasting pipeline using a SARIMA model and Monte Carlo simulations (machine learning algorithms I’ll explain later) that re-forecast in seconds.

Saving a few hours is big, but to do this in seconds? This is super impressive.

He didn’t hire a data scientist. He used:

1. AI to write the code
2. His own finance domain knowledge to guide the model.

Today I’m going to show you exactly how he did it, and how you can build the same thing in 5 steps.

The problem CFOs aren’t measuring

So, let me ask you this.

How often do you download from your accounting system, open Excel, map your accounts, filter, copy to actuals, find budgets, calculate variances, format, then present…

Only to then find that:

Your customer delays payment by 30 days.
Your credit facility terms get renegotiated.
Your early payment discounts change.

And you have to rebuild the whole thing.

I call this ‘decision lag’.

It’s the time between when reality changes and when your forecast catches up.

In a calm year, you can manage a few days of lag.

But in 2026, with everything that’s happening geopolitically plus rate uncertainty, decision lag is probably the biggest cash management problem that most CFOs aren’t even measuring.

Think about it like this. You have a team member who did great work on Monday, then went on holiday for the rest of the month. The analysis was solid when they handed it in. But nobody’s updating it as things change.

You wouldn’t accept that from a person. So why accept it from a process?

10,000 Scenarios

Juergen is not a data scientist. He didn’t take a machine learning course. He’s a finance guy who understood his numbers and figured out how to make AI do the heavy lifting.

His system is actually straightforward. Data goes in: bank balances, AR/AP, his assumptions.

Then the model does two things:

• SARIMA: Looks at your past cash data and spots the patterns that repeat, like how collections always reduce in August or increase before year-end. It learns your trends, then projects it forward.
• Monte Carlo: Takes that projection and asks “but what if things don’t go exactly to plan?” You choose how many times to run it – it could be 1,000 or 50,000, but 10,000 is the sweet spot.

Juergen runs 10,000 versions of the forecast, each time changing the assumptions slightly.

Line them all up from worst to best and you get a probability range which is displayed as P numbers.

• P10 — only 10% of scenarios came in worse than this. Your realistic downside.
• P50 — the middle. Your expected case.
• P90 — 90% of scenarios came in below this. Your realistic upside.

So instead of telling the board “we’ll have $2.1M in cash,” you say “somewhere between $1.6M and $2.5M, most likely around $2.1M.”

That’s a conversation with the board that is much easier to plan around.

Jeurgen’s forecast updates in seconds. He runs it regularly for clients. And when assumptions change? Instead of re-building it, he just re-runs it.

His first version was a flat line (it didn’t work).

But he stuck with it, iterated, told the model about seasonality, excluded anomalies like COVID, adjusted for client-specific patterns etc.

After a few rounds, the model had a much better understanding of his data (based on his finance domain expertise) and he was ready to forecast forward.

He built this going through our AI Finance Accelerator program. So if you want to know how that works, just DM me. Happy to walk you through it.

Here’s how to do it yourself…

5-Steps to Faster Forecasting

Step 1 – List your drivers

Write down the 5-8 variables that move your cash: collections DSO, payables timing, payroll cycles, debt service, seasonal patterns. Plain English.

This is finance work, not coding. You’re telling the model what matters.

Step 2 – Generate the base model

Open your preferred (secure) AI tool.

Prompt the AI:

It writes the code. You don’t need to understand every line, you need to understand what it’s doing and why.

Before proceeding to the next step, make sure to copy the code. You will need it.

Step 3 – Open Google Colab and Paste Code inside

Open Google Colab in a new tab or by clicking here. Then create “New Notebook” to get started (make sure you are signed in).

Paste the code inside the cell below:

By the way – if you want to learn more on how to use Google Colab, you can check out my newsletter on how to use Machine Learning: Machine learning in 6 steps: How to make better decisions with finance data your brain can’t see.

Step 4 – Iterate with your knowledge

See the results Google Colab produces.

The first output will probably be wrong. So, go back to your AI, and push back the same way you would with a junior: “You’re not capturing Q4.” “Exclude 2020.”

Each round takes 10-15 minutes. After 3-5 rounds, you’ll have something that’s a lot better.

Just keep improving the code and re-running in Google Colab – Super easy.

Step 5 – Add Monte Carlo

Once you’ve got a working base model, you can now ask your AI to add in Monte Carlo simulations with a prompt like this:

Now when you re-run the updated code in Google Colab, you have probabilistic forecasting!

Not “what we expect” but “what range should we plan for.”

This is the use case that makes your board trust the model.

Super important – start with one driver only.

Collections is usually the best starting point because it’s the easiest to validate.

Get that working, show the board, then expand.

The One Thing To Remember

When you focus on systems instead of spreadsheets. Cash flow stops being a 2 day project.

So when you’re asked:

“What happens to our cashflow if collections are reduced by 10%?”

You can answer, in 2 minutes, in the meeting, with scenarios.

Plus (and this is super important) Juergen didn’t become a Python developer.

He just decided to become the finance professional who builds intelligent systems instead of requesting another spreadsheet.

Because, a forecast that’s only accurate on the day you made it isn’t a forecast. It’s just a receipt.

But, now you can build one that updates itself ;)​

Formula Helper vs. Financial Modeler

If you’re using Copilot like this: type a question, get a formula, apply it, move to the next cell, repeat.

That was fine in 2024. This is not fine in 2026.

Microsoft have just rebuilt the Copilot experience in Excel and renamed Agent Mode to “Edit with Copilot” because ‘agentic’ editing is now what Microsoft are moving toward as standard.

If you’re still prompting Copilot one cell at a time, you’re running 2024 software in your head on top of a 2026 tool.

And this is costing you hours every week.

So, today I will teach you how to use ‘Edit with Copilot’ to build auditable models using only Excel, in 5 minutes or less.

Tell Copilot What to Build, Not What to Do

Stop asking for help with this cell. Start asking for help building this model.

Edit with Copilot plans a multi-step approach, executes across multiple tabs, creates formulas that reference each other, builds charts, and self-corrects when something doesn’t work.

I show this live during my webinars. And, in just 5 minutes, Excel creates a perfect three-statement financial model.

All I needed was this prompt:

And the model it builds is linked. I can test it by changing assumptions and the entire model updates. If I edit the formulas, I can check if everything stays connected.

Plus, you can now pick which AI model you want to use.

Click the model picker in the Copilot pane and choose between Auto, GPT-5.2, Claude Opus 4.5, or Claude Opus 4.6.

For financial modeling where you need reasoning chains, assumption logic, and scenario structures, Claude is the better choice.

Previously Microsoft decided for you. Now you pick the model that’s best for your finance work.

How to Start Using ‘Edit with Copilot’

Step 1: Open Excel

Click Copilot in the ribbon. Click the “+” and then select “Edit with Copilot.”

That’s where Agent Mode lives now.

Step 2: Choose model

Click the model picker at the top of the Copilot pane. Switch from Auto to your preferred model.

Step 3: Test it out

Write one prompt that describes the outcome, not the steps.

Here’s mine:

Step 4: Watch it build

Copilot shows its reasoning. You can click the “Reasoned in..” dropdown.

It created tabs, linked formulas, and generated charts.

When it finds errors, it fixes them.

Step 5: Review output

The most important step of all, you need to review it like you’d review a junior’s work. Change an assumption and check if the model updates.

But you don’t audit this by recalculating every formula yourself. You ask Copilot to build the audit for you.

Two more things to do.

1. Ask Copilot to include an assumption sheet. All your salary increases, tax rates, growth rates in one place. This makes the model auditable and easy to flex
2. Ask Copilot to add reconciliation checks:

Make Copilot document its own logic and prove the numbers work. Then you review the checks instead of the the entire model.

Treat it like delegating to a junior team member. You still sign off before anything goes anywhere.

Because, AI still makes mistakes like a human. But this way, your audit trail is built in, and much easier to review than trying to unpick everything afterwards.

The One Thing to Remember

This is a direct quote from me after building a complete financial model inside Excel with a single prompt.

I didn’t need to open ChatGPT, copy and paste. All I had is Copilot with Claude doing the work inside the spreadsheet.

Agent Mode didn’t disappear. It just evolved. And now, with Claude inside Excel, you have the best reasoning model available working directly in your financial models.

Stop asking Copilot for formulas. Start telling it to build your models.

The tool is ready. The only question is whether you’ll update how you use it.​

Before AI, I waited 6 months for a dashboard. It never came.

It took 3-4 days of work in Excel for my team to create cost centre reports for each department.

Then a BI team arrived and we were so happy we thought we’d have everything automated, and on a dashboard by the end of the month.

Six months later? Nothing.

Now AI can build the same thing in 10 minutes.

But thinking about this now, I don’t think the BI team failed because they were slow.

I think they failed because nobody ever agreed on what the dashboard was for.

And here’s what’s crazy. Procter & Gamble had solved this exact problem 14 years ago. Before AI even existed. They built dashboards around decisions (not data) and killed more than 80% of their BI reports doing it. I just didn’t know yet.

Dashboards you build too quickly become AI slop that nobody looks at. So today, I’m going to show you the method P&G proved works over a decade ago — and how to apply it to your AI dashboards this afternoon.

Avoid becoming a ‘Data Decorator’

The moment you want to show an AI-built dashboard lands in front of the CEO, the board, or your team, the rules change.

You need to have put real time into the design thinking first. Things like the decision you want to make, the KPIs that support it, and the layout.

If you skip this, you are just creating AI slop that shows your data in a different way. And your credibility goes with it.

You become a ‘data decorator’ where your CEO looks, nods, and forgets about it.

In my AI Finance Accelerator we make all participants create a dashboard using AI in week 2. But, one that they can then use to create an impact in their business. Here are some examples of what they produce:

The ‘Decision Cockpit’

Back to P&G.

In 2012, their Business Services Group President Filippo Passerini asked a different question.

Instead of “what dashboards should we build?” he asked, “what decisions do our leaders need to make?”

Then he built dashboards around those decisions. He called them Decision Cockpits. And they replaced more than 80% of P&G’s standardized reports, which were then deployed to 50,000+ employees across 50+ offices.

Source | Practical Analytics​

Passerini’s own description: his dashboards were “focused on forward-looking projections rather than historical reporting.”

That difference (building around the decision, not the data) was the focus in 2012 when P&G did it without AI.

It’s still the focus in 2026, when you can build the same thing with AI in 10 minutes.

Your edge is your brief.

You know which decisions matter. You know which KPIs drive them. You know what is important to the board.

AI can build in 10 minutes what a BI team could not build in 6 months.

But only you can decide what the dashboard is for.

5 steps to your first decision-first dashboard

Step 1. Name the one decision (on a Post-It, before AI)

Write one sentence: “This dashboard helps [role] decide [action] each [cadence].”

Example: “This dashboard helps the CEO decide whether to approve Q2 hiring requests, monthly.”

If you cannot finish the sentence, do not build the dashboard.

Step 2. Name the owner and the cadence

Who acts on it, and how often? If the answer is “everyone, whenever” – stop.

A dashboard with no named owner has no owner. Pick one name and one cadence. Weekly, monthly, per board meeting.

Step 3. Name the trigger metric

What single number, if it moved, would change the decision? That is the top of your dashboard. Everything else is evidence.

Step 4. Brief the AI with the decision, not the data

Use this prompt pattern:

Tell AI what the dashboard is for, before asking it to build.

In my last Crash Course I uploaded a file and said: “I am the finance manager, I want to understand how to increase sales. I want to build the best dashboard.”

I was not asking it to build yet. After one minute I had a proposed set of metrics and a layout, and that was my chance to push back.

One thing I push back on every time: data availability. Before you let AI do anything, walk through each proposed KPI and ask, can I pull this data from my systems, reliably, on the cadence the decision needs?

If the answer is no for even one KPI, change it out or find something different for now.

Otherwise you end up with a beautiful dashboard you can never populate and every month you waste time finding exports to create it.

Step 5. Review the layout before the data is plotted

When AI returns its proposed structure, check it against the decision.

Is the trigger metric the most prominent element? Does every other chart earn its place by supporting that one decision?

If not, keep on improving it.

The CFO of HPE did this, this year

Earlier this year, Fortune published the story of Marie Myers, CFO of Hewlett Packard Enterprise (article here).

Her weekly Monday review used to improve 100+ PowerPoint slides and hundreds of hours of manual prep across the business.

Her team built an AI dashboard called Alfred (yes, after Batman’s butler) that led to: A 40% cut in financial reporting cycle time, a 25% cut in processing costs, and 90% of the manual weekly prep removed.

The super important bit? Myers told Fortune: “The first move wasn’t to switch on AI, but to redesign the work.”

She rebuilt the process around the decisions her business needed to make – then switched on AI.

And if you’re thinking “yeah but I’m not Fortune 500” – this is good. You’ve got less complexity, less politics, and AI that costs $20 a month.

If the CFO of a Fortune 500 company had to start with a blank sheet of paper, so can you.

The One Thing To Remember

My BI team failed in 6 months. An AI dashboard can now fail in 10 minutes. The time saved is not the point. The decision behind the dashboard is.

P&G proved this in 2012, without AI. HPE proved it still works in 2026, with AI. The tools change. The method does not.

Your next dashboard should not start in Excel, in ChatGPT, or in Claude. It should start on a single sheet of paper, with one sentence:

“This dashboard helps [who] decide [what] each [when].”

If you can write that sentence, you are already the most valuable role in finance in 2026 – the person who owns the brief (and the decision).

Pareto figured this out in 1896. Most most finance pros ignore it.

So, let me tell you this.

Vilfredo Pareto worked out in 1896 that 80% of Italy’s land was owned by 20% of the population.

This same principle can be seen everywhere.

80% of your sales come from 20% of your customers.

80% of your bugs come from 20% of your code.

And, 80% of the answer in any finance analysis comes from 20% of your data.

So when you upload a 14-tab Excel file to ChatGPT (which includes every formula, every unused column and every hidden tab from the 2019 version) you are asking the model to find the 20% on its own.

Sometimes it does. But a lot of the time you have to ask again, and again, and again.

Until ChatGPT outputs it’s worse habit. An answer that looks convincing, but is wrong.

And then you’re unpicking the analysis and wasting even more time that you do not have.

So, today I’m showing you the fastest way to focus AI to improve the quality and accuracy of your results.

​

​

Why ChatGPT hates your 14-tab upload

When you upload an entire workbook, ChatGPT then has to parse merged cells, hidden tabs, and ten years of irrelevant data before it ever gets to the question you asked.

And the more data you have, the more it’s bad habit of making up convincing answers gets worse.

You wanted variance commentary on Q1 marketing spend.

But instead, you got a variance commentary that seems a lot like it was generated from the Q1 marketing spend data, but wasn’t.

And if your instincts are not strong, this is a big risk to your credibility (and your business).

Think of it like this. You don’t hand an analyst a full Google Drive folder and ask “what’s interesting?”

You hand them the cost-centre report, point at the line that changed, and say “tell me why.”

Remember, the quality of your output is a direct function of the quality of your input.

And you must always audit your responses.

For how I use AI to ensure accuracy, you can also read my previous newsletter ‘Steal my AI accuracy method (from 7 years in audit at PwC)’ here.

The screenshot-first rule

A screenshot does three things at once. It gives ChatGPT a better idea of what to ignore (important so it doesn’t hit limits).

It gives the model exactly the same focused view you’d present to a team member.

And it helps you fully formulate what questions you need answered by explaining the screenshot (which is the basis of better prompting anyway).

The end result is a specific output that you can defend. The kind that you can use in a report straight away without having to re-write anything.

Step 1: Upload your workbook

It’s still OK to upload your workbook (ChatGPT will need to use your data for the analysis) the screenshot adds more focus.

Step 2: Screenshot specifics

Open the workbook. Find the range that matters and take a screenshot of just that range.

Make sure to capture the columns as well as this will save you having to explain your column naming conventions in the prompt.
​
​Step 3: Paste the image in Thinking Mode

Thinking Mode is super important here, as thinking Mode reasons through the data and surfaces angles you wouldn’t have asked for.

Note – Whether you have extended thinking or pro depends on your ChatGPT plan. Pro thinking modes are limited to the Pro, Business, Enterprise or Edu plans.

Step 4: Ask for ten best analyses

Use this exact prompt:

You’re asking for the menu of angles to begin with, not the final output.
​
​Step 5: Pick the strongest angle

Read the ten options. Three or four will be obvious. Two or three will be useless. One or two will be the angle you wouldn’t have thought to run yourself.

Pick that one then ask ChatGPT to do the full analysis on your data.

The whole thing takes five minutes, using the same workbook you’ve been struggling to analyze for weeks.

The Pareto rule is also how I found innovative ways to analyse costs and sales. Like for a supermarket, profitability per square meter. Or software cost by headcount.

The One Thing to Remember

You don’t need a new tool. You just need to stop dumping the whole workbook in and expecting ChatGPT to find what it needs to focus on.

The smartest teams are producing 80% of their results, with 20% of the effort.

So this week, pick the next analysis you’d normally upload a workbook for. But don’t upload the workbook straight away.

Screenshot the range, paste the screenshot and run the five steps.​