The 3rd and final installment of the AI CFO bot conversation for now.

I refuse to only write about AI, but this is one of the biggest topics right now and I have something to say about it.

Week 1 was the framework — a CFO is a collection of workflows, pick one, build it well.

Week 2 was the proof — I built a marketing analytics workflow for myself. One stakeholder, small dataset, low stakes. It worked because I kept it simple.

My build worked because it was simple. One person. Structured data. Low consequences if it's wrong.

This week, same framework, applied to a real client problem harder in every dimension — demand planning.

Demand planning is traditionally outside day-to-day of finance. But finance has a huge part in it — cash flow, inventory investment, vendor payment terms, PO commitments.

And we as a firm got pulled in to help build it — demand planning isn't strictly finance, but the financial implications are enormous.

Here’s my video walkthrough for the week:

Here's the setup (anonymized, but real numbers):

The backstory: when we first started working together, they had about five different spreadsheets — each with roughly 30 tabs — all doing some version of demand planning. We took all of it and simplified it into one single spreadsheet.

We built them a solid tool. And we added and iterated over the last couple years until we arrive at today’s problem.

Every off-the-shelf “tool” they've tried has failed:

The spreadsheet we helped build 3 years ago isn’t perfect but it is still running the show. A $100M business running demand planning on a spreadsheet.

And they love it. It's not broken — it works well enough that they don't want to lose the functionality.

The thing they value most: adjusting numbers in real time and seeing downstream effects instantly. Fully customized to how they work.

"What happens if I move 500 units from one warehouse to another?" They can answer that in 30 seconds. No other tool they've tried can do that.

But we've Frankensteined this thing so many times over 3 years — new retailers, new inventory locations, new channels, new SKUs — that it's hitting real scale challenges.

Some of their forecasting assumptions haven't been updated in over a year. They know they need to evolve past it. And we do too, as a firm — we need to be smarter, better, more capable. We have better tools now.

We all know how complex these spreadsheets can be — whether it's demand planning, cash flow forecasting, modeling, scenario planning, or LTV analysis. This won't be unfamiliar to you, but just to give you a basic schema of the complexity of this particular model:

This is not a spreadsheet. This is a bespoke demand planning application that happens to run in Excel. And that distinction matters for what comes next.

They did the obvious thing first.

They tried downloading the entire Google sheet and putting it into Claude and backing their way into AI-powered analysis.

Using Claude was "really tough" — multiple distribution warehouses, orders on a monthly rotating cadence, and constant real-time decisions about where to route inventory. The spreadsheet was built for a human to scan, click around, and visually interpret.

And this is the key insight of this entire series, probably the most useful thing I'll share across all three parts:

It didn't fail because Claude isn't smart enough. It failed because the spreadsheet is built for humans, not for machines.

The Excel file communicates through layers of interconnected tabs, conditional formatting, merged cells, and formula chains that go five layers deep. A human reads it by scanning, clicking, and interpreting the layout. Claude can't do any of that.

LLMs process structured data — rows, columns, clear labels. Not visual layouts designed for a human brain.

So what actually goes wrong when you try to use AI with something like this?

Through the lens of our framework from Week 1, the problems are all the same ones — just harder.

Data Sources - You have some of the data sources, but maybe not all. The spreadsheet has most of it, but some critical context — a supplier conversation about lead times, a marketing initiative about to spike demand, a retailer launch that wasn't confirmed until the last minute — doesn't live in any system.

The operations team holds that in their heads and acts on it subconsciously within the Excel file. Teasing that out and operationalizing it is really hard.

Cadence - Claude isn't good at giving you the same structure, the same report, on a schedule. If you're just using chat or a Claude project — and you haven't built an actual app — you're starting from scratch every time.

There's no continuity. The demand planning team needs the same view, updated, every week. Chat doesn't do that.

People - You can't easily share a Claude project with your team. A Google Sheet is easy to share. People interact with it live. They can change numbers, see downstream effects, collaborate in the same file. That's not possible in a Claude project. And this brand has four stakeholders who all need different things from the same data.

Context - The tedious task of telling Claude what matters. This meeting's important, this one's not. This context is relevant, ignore that. This is the latest data, that's stale. The operations team does this subconsciously when they work in the spreadsheet — they know which tabs matter, which assumptions are current, which scenarios the founder actually cares about. Translating all of that into prompts and context for Claude is an enormous amount of work that people underestimate.

And beneath all of that, there's a structural problem. This one spreadsheet serves four functions:

One spreadsheet doing all four of those things is why it's become a Frankenstein. And it's why "just put it in AI" doesn't work — you're not replacing one tool, you're trying to replace four at once.

Here's what people aren't talking about.

Everyone's assuming that AI means AI everything. Rip out the spreadsheet, build it all in Claude, automate the whole thing.

I don't think that's right.

The spreadsheet doesn't die. And that's actually a good thing.

I've heard this complaint over and over from founders: "We just live in spreadsheets."

Well... that's okay. Spreadsheets are still amazing. They have their place, and people keep using them for good reason.

Think about what that spreadsheet does well: real-time manipulation, instant feedback, sandbox-style scenario analysis. "What if I move 500 units here?" You get the answer in 30 seconds. Try getting that experience from a database query or an AI chat.

I'm not advocating for immediately eliminating all spreadsheets. I'm advocating for using each tool for what it's great at.

If you're trying to explain to your team why the spreadsheet can coexist with AI, here's how I'd frame it:

The spreadsheet keeps doing what it's best at — dynamic, sandbox-style planning. The database handles storage and ability query. Dashboards give each person the view they need. AI can finally work with the data because it's structured.

The spreadsheet is finally just doing what it's best at instead of doing everything.

In Week 1, I said every workflow has the same anatomy: data sources, cadence, stakeholders, context, and steps. In Week 2, I filled it in for my own build. Here's what it looks like for this one — and why almost every answer is harder.

Data sources: P&L data, Shopify data, ERP/fulfillment data (once their new platform is live — about 3 months out), manual data sources. And the hardest category — context that lives in people's heads. A retail launch that wasn't confirmed until last minute. Supplier lead times that shift. Marketing initiatives that would spike demand.

Cadence: They live in this thing daily. Ordering is monthly across multiple vendors with 3-4 month lead times. Real-time allocation decisions. Structured forecast-vs-actual reporting on its own rhythm. No single cadence — multiple overlapping ones.

Stakeholders: The founder needs the high-level picture. Head of operations needs granular inventory visibility. Operations manager needs transfer costs and logistics. Controller needs forecast-actual and cash flow. My build had one stakeholder and no authentication. This one needs role-based access and probably different interfaces for different people.

Context: Mostly the data itself — but with constant ad hoc context. A one-off retail order. A supplier conversation about lead times. A marketing initiative about to spike demand. Ideally the system captures and remembers that context. This is the hardest part — systematizing context that currently lives in conversations.

Same framework. But almost every answer is harder. This is why you don't start here. You start with the simple version — like I did in Week 2 — and work your way up.

This assumption floating around that AI means AI everything.

Rip out the spreadsheet. Automate the whole workflow. Build it all in Claude.

I don't think that's right.

The spreadsheet is a great tool. The problem was never the spreadsheet — the problem was making it do four jobs at once. Storage, reporting, planning, communication. No single tool should carry all of that.

You peel those apart. You let the database handle storage. You let dashboards handle reporting. You let AI handle analysis. And the spreadsheet finally just gets to be a spreadsheet.

If you're working through this yourself or want to talk about how to apply this to your own situation, my inbox is open.

-Sam

I spoke with Joel at Uproot Clean last week and they have a posting for a Finance Manager if you know someone looking

Talk to me

My inbox is open and I love hearing from you - respond to this email with your questions, feedback, things I screwed up, or just because.

Looking for more? Some of our most popular posts:

Did someone forward this to you? If you like it, you can subscribe here