If you want to watch instead of read, here’s a companion video for this week

I got this email from a prospective client last week.

TL;DR – they decided to “build an AI CFO bot”.

And over the past several weeks, I’m seeing more signals across a full spectrum from small automations all the way to a full AI CFO bot:

What’s more valuable to you is how an actual CFO would build (or not build) an AI CFO bot with today’s capability.

And more important than the what-to-build…

…a mental model of how-to-think about building.

Whether you’re considering tools, vibe coding some automations, or embarking on your own AI CFO bot project, allow me to save you some time and tokens, because the frame of the problem determines whether you end up with the usual slop or something useful.

“AI CFO bot” is redundant. We just need “CFO” and the “bot”.

Starting with the CFO part, what is a CFO?

My therapist says “name it to tame it”. So if we can define what is a CFO is, we can have a better chance of building this bot.

It’s like every other role in the business – a collection of workflows.

Workflows like:

We can debate which workflows are included/excluded from brand to brand (except the whispering part which is always included).

But the CFO role remains a collection of workflows. And each workflow shares a common anatomy.

A handful of steps, data sources, cadences, people, and a layer of evolving context.

Cash flow forecasting for example looks something like this:

Although sometimes more messy and complicated, the anatomy is the same for all other workflows just with different data sources, cadences, stakeholders, and context.

For building a CFO bot, this is actually good news because it means each workflow, taken on its own, is theoretically possible to build – at least to some degree of effectiveness.

Making my therapist proud.

So, this all becomes a big engineering problem as long as you know what you’re building

Those are buildable. We’re already doing this with clients where we build one workflow at a time, with a human in the loop, and make continuous improvements every month.

So if we stopped here, you conclude that an AI CFO bot is basically a stack of workflow agents and you’re most of the way home, but that’s where it gets harder.

Two reasons.

1. Some of the context that matters isn’t in any database. If you don’t have 100%, you have zero percent.

The full time controller can run the close. But the controller is also holding in their head that the supplier called Tuesday and pushed inventory two weeks, that the bank wants three more months of reconciled inventory before extending the line, and that the founder is privately considering leaning heavier into retail next year.

Only parts of that context live in QuickBooks, email, and meeting transcript. Some of it doesn’t live in writing at all.

The $100M+ client who uploaded my 30-tab model to Claude figured this out fast.

The model is there. The data is there. But the model only works because there’s a human who knows which assumptions are stale this month, which scenarios the founder actually cares about, and which numbers to ignore because we already worked through them last week.

That’s why every AI CFO demo or whatever you code over the weekend looks great in Claude but falls apart fast.

2. Founders are really bad at follow-up. Bots make it worse.

I caught up yesterday with a founder friend who, by his own admission, knows exactly what to do to grow the business. He just doesn’t do it.

He wants to be the visionary, the idea person. The new product, the new ad platform, the new partnership.

The integrator work — the follow-up, the management, the repetitive tasks — he avoids. Most founders I work with are some version of this. I certainly am myself.

Now hand that founder a CMO bot, a COO bot, a media buyer bot, and a CFO bot.

What does this founder’s day now look like?

Who’s reconciling when the CMO bot’s growth plan contradicts what the controller bot is telling the bank? Who’s feeding back the context that makes next month’s run better than this month’s?

The CFO role still exists today because even if we could define all of the anatomy of each workflow, the context doesn’t all live in a database and most founders don’t care about doing finance work.

I’m very bullish on AI handling individual financial workflows at around 80% of what a good CFO does manually.

One workflow at a time, with a human in the loop is a real path. And the quality bar is going to keep moving up.

But I’m completely convinced that a monolith version of an AI CFO bot will fail to deliver the leverage founders are actually looking for.

Bad financial decisions have real consequences. And opportunity costs are still a thing.

Building and maintaining a CFO bot are hours you could spend on marketing and product. The value prop of a human-in-the-loop is still strong.

If you are going to embark on a CFO bot journey, I encourage you to begin with ONE VERY SMALL workflow. Build one workflow well (or at least good enough), then another, then another.

Save this for picking your first workflow. The right one has four traits:

Cash flow forecasting possibly fits, depending on the brand. Variance analysis is a nice option. Strategic planning isn’t. Start where the boundaries are tight, then expand.

Next week I’m getting more tactical about how we’re actually building this for a client.

-Sam

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.