Why Finance Automation Rollouts Keep Stalling: The CFO Questions to Ask First

Full transcript

Lightly edited for readability. The audio is the original.

0:00I've watched 4 AI rollouts stall

I've watched 4 AI rollouts stall inside finance teams in the last 2 years. None of them stalled because the tool was bad. They stalled for the same 3 reasons every single time. And almost nobody's talking about it, because every vendor in the space is selling you the version of the story where the tool wins.

What I'm going to walk you through today is what's happening on the ground: 3 failure modes I keep seeing, why each one happens, and the one set of questions I'd ask before any AI ever rolls into your finance function again. If you've watched an AI pilot in your function go quiet over the last year, or if you're staring down a vendor deck right now, this episode is for you.

Welcome to Excel at AI in Finance, the show for finance pros who want to use AI to get home sooner and be the one their company can't afford to lose. I'm Mike, I run F9 Finance, and I'm a senior finance leader working in the trenches.

0:56The story: a six-figure AI tool at a regional services firm

Now, let me tell you about one of those 4 failures. Names change for obvious reasons. This one was about 6 months ago. The company was a regional services firm, finance team of about 40.

The CFO had just sat through one of those AI vendor demos, the kind where they automate the whole close in 8 minutes on stage with clean data, and you walk out thinking the function is going to look completely different by next month. She bought the tool. Six-figure annual contract with a 2-year commitment. The mandate came down from the top. Every senior analyst gets a license, and every team is supposed to use it for everything.

Now, here's what played out. Month 1, a handful of people log in. They poke around. They get 1 or 2 outputs that look impressive in a Tuesday status meeting. Screenshots fly around, Slack channels light up, and the energy is high.

Month 2, half of them have stopped opening the tool. The other half are still trying, but every time they try to use it on real work, like a real close or a real board prep, the output comes back close enough to be interesting and just wrong enough to take 20 minutes to fix. Which means it's faster to just do it the old way, and they know it.

Month 3, the CFO asks for an update on the AI project. Nobody on the team can point to a single finished piece of work the tool shipped. Not one variance commentary that made it into the board pack. Not one forecast adjustment that made it into the model. Nothing.

Month 4, the whole team is quietly 100% back in Excel and the same process as they had before. The license sits there, being paid for a year and a half longer. The vendor sends a friendly quarterly check-in email. And the CFO, who genuinely wanted this to work, is now privately telling her CFO peers that AI just isn't ready for finance yet.

Now, here's the thing that's important. That tool was not bad. I looked at the product. The model worked, and the output in a vacuum was solid. So why did this rollout stall?

Well, there was one moment in particular I keep coming back to. A senior analyst at the launch call asked a question, a good question, about what they should do with the tool first. The vendor said they should use it on everything they do. Just keep testing it, and the time savings will come.

So here's the question I want to answer today. Why does a tool that works in the demo and works in a vacuum stop working inside a real finance function? Because this is not a one-off. I've watched this exact pattern play out 4 times in the last 2 years, and the same 3 things kept showing up every time. So let me walk you through what's really going on.

3:15Why AI rollouts stall in real finance functions

Okay, so why do AI rollouts stall in finance functions? Because almost every vendor in the space is selling you the theoretical version of the story. The version where the tool works on clean data, in a finance function that has slack time to use it, owned by someone whose actual job is to make it work.

That's the demo. That's the slide deck they hand you. But that's not your finance function. That's not my finance function. That's not anybody's finance function. The gap between the demo and the real day job is enormous.

What I've watched happen across these 4 rollouts, and what I keep seeing happen in conversations with finance pros every single week, is that the same 3 failure modes show up every time. They're not really about the tool. They're about how the rollout was set up before anyone ever opened the software. Let me walk you through them.

4:01Failure mode 1: tool first, not job first

Failure mode number 1: tool first, not job first. The first failure mode is that the tool gets picked before the job gets named. Now, walk through this with me, because it's almost invisible while it's happening.

The CFO sits through a vendor demo. The vendor shows AI doing the close. Beautiful slide, clean data, the kind of close that I've personally never seen in a real company. The CFO leaves the meeting thinking, we should buy this. So they buy it. Maybe they get a small steering committee involved, maybe they pilot it with one team first.

Six months later, that team has not used the tool for the close, not in any meaningful way. Why? Because no one in the buying conversation ever stopped and asked the actual question: which specific finance job are we trying to win with this? Not the close, because the close is the demo job.

Your close is a specific shape. It has 14 entities, 3 ERP systems, 2 of which don't talk to each other, a chart of accounts that grew sideways for 10 years, and a manual journal entry analysis that exists because nobody's gotten around to standardizing it. Your close is not the demo close.

So when the team finally tries to use the tool on the real close, the gap shows up immediately. The tool can't read your trial balance because the export format doesn't match what the model was trained on. The commentary it generates is too generic for the audit committee. The variance analysis flags accounts that don't exist in your CFO's mental model, because you have legacy mapping that nobody's cleaned up.

And here's the part that's wrong. They say, this tool doesn't work, because they were sold on the tool, not on the job. So when the tool can't do the job, the tool becomes the problem.

Now, think about the rollout I told you about a minute ago. This is exactly what played out. The tool was bought to do anything, but nobody on the buying side had asked what that actually meant when you got down into the finance function, for that CFO, with those audit and board committee expectations. So when the analysts opened the tool and tried to use it for their real work, it kept giving them outputs that didn't fit, and they stopped opening it.

The fix is short and brutally simple. Pick the job first, then pick the AI that wins that job. Find the one specific finance task that you'd pay real money to make 20% faster. It could be your variance walk, your weekly cash flow update, your monthly KPI commentary, whichever one is eating your Sunday. Find a tool that gets that specific job done. The other way around almost never ships.

7:12Failure mode 2: top-down sponsor, no in-the-trenches owner

Failure mode 2: having a top-down sponsor, but no in-the-trenches owner. Now, this second failure mode is that the people who bought the tool aren't the people who use it, and the people who use it have absolutely no time built into their week to sit down and build with it.

This one is everywhere. The CFO mandate comes down. Senior analysts and managers are told to use the new tool, and their performance review now has AI adoption as a line item. Their week, however, has not changed at all. They still own the close, they still own the board pack, and they still get pinged at 6 p.m. to fix a number for the morning meeting.

So when exactly do they build the AI workflows? Well, the reality is they don't, because there's no calendar block for it. The mandate said use the tool. The mandate did not carve any time out of your current workload so you can actually build with the tool. So the analyst tries to fit AI experimentation into the cracks of an already overflowing week. It doesn't work. The cracks are too small.

And there's a specific tell I look for in any finance function with an AI pilot. I ask one question: who owns the prompts when they break? If the answer is no one, or the vendor, or "we figured it out together at first, but I'm not sure anymore," the pilot is stalled. The team just hasn't admitted it out loud yet.

The CFO mandate is necessary, but it's not sufficient. You need a sponsor at the top who will defend the investment when month-end becomes a fire drill. And you need an owner in the trenches whose job description, on paper, includes the line "build and maintain the AI workflows for this function." Not "use the AI tool when it's convenient." Build and iterate.

Now, that role doesn't have to be full-time, but the time has to be real. 5 hours a week, at minimum, blocked on the calendar, defended like any other commitment. My personal rule for finance pilots is simple. If the build time is not on someone's calendar, well, the build's not happening.

Now, in the rollout I told you about, the CFO had a mandate. The CFO had a budget. The CFO had a strategic vision. What the CFO did not have was a senior analyst with 5 hours a week carved out of their close calendar to build the workflows. The pilot stalled the way a rocket stalls when you forget to fuel it past the launch pad.

You have to have both. A sponsor at the top, who will defend the investment when the calendar has the AI build time blocked on it, with explicit budget for the hours that come out of their normal workload. Both. Not one or the other.

9:34Failure mode 3: no iteration loop (and the AI-ready data myth)

Failure mode number 3: no iteration loop. The third failure mode is the one nobody talks about. The first prompt works, the second prompt breaks, and there's no plan for who fixes it, when, or how. The whole rollout was budgeted as if the tool either works or it doesn't, instead of being budgeted for the reality, which is that the tool works after the 15th round of iteration.

Every vendor demo is 1 or 2 prompts deep. Go ahead and watch them from now on, and you'll notice this. That's the structural problem with vendor demos. They show you a prompt that they know works. They don't show you all the prompts that broke when the analyst tried it on a quarter with an outlier in it. They don't show you the prompt that hallucinated 3 line items because the analyst forgot to specify which data to use.

They don't show you the back and forth that goes from "this output is bad" to "this output is good enough to put in front of our CFO," or in front of a vendor demo. Real finance work is hundreds of prompts deep, each one slightly different, each one requiring some level of iteration. The pilot stalls at the first edge case because nobody budgeted time to iterate.

Whose job is it to fix the prompt when it breaks at the edge case? In most rollouts I've watched, the answer is "we'll figure it out," which means it's nobody's job.

There's a related myth that ties into this. You'll hear vendors and consultants talk about AI-ready data. Your data has to be cleaner before AI can work. Your master data needs to be standardized. Your chart of accounts needs to be rationalized. There's some truth to that.

But most of the time, when someone says your data isn't AI-ready, what they're really saying is your iteration loop isn't built. Because here's the secret. Your data ultimately is what it is. Your data is never going to be perfectly clean, and real finance functions have to run on imperfect data.

The iteration loop is what closes the gap between imperfect data and a usable output. The team that wins is the team that knows how to iterate on the prompt and the workflow until the imperfect data produces a workable answer.

Now, in the rollout I told you about, the team got an output that was 80% right, and 20% needed manual cleanup. Their reaction? The tool doesn't work. What they should have asked was, who on our team is going to spend an hour a day for the next 2 weeks iterating on this prompt and workflow until the output is 95% right and we trust it? Nobody had that on a calendar, so the pilot stalled.

Now, there's a third piece of this too, which is the budget piece. The way most rollouts are budgeted, you buy the license, and the implementation is treated as a project with a launch date. After launch, the team owns it. Except the team does not have launch day workflows yet. The team has a tool. The workflows still have to be built and iterated.

And nobody budgeted for that ongoing build work, because it doesn't look like a project. It looks like just using the tool. Building those workflows is what makes the tool useful, and that work has to be budgeted from day one.

And when the close hits, or when board prep hits, or whatever the next fire is, the iteration time is the first thing that gets cut. The workflows never reach the bar of usable in production, so the pilot stalls.

You have to build the iteration loop into the pilot from day one. Name the owner. Block the hours on their calendar. Give it explicit budget. Plan for 15 rounds of iteration per workflow before the workflow ships in front of anyone. Without that, the tool stops the first time it hits reality, and the rollout stops with it.

12:42The spine: 3 things that decide whether the next AI rollout ships

Again, AI rollouts in finance don't fail because the tool is bad. They fail because the tool got picked before the job got named. They got mandated without an owner, without an iteration loop. The tool isn't the rollout. The finance job is. The owner is. The iteration loop is.

When you see an AI rollout in your function, or your peers' function, or a competitor's function, or one of those LinkedIn posts where someone ironically claims they automated their whole close, your first question shouldn't be about the tool. It should be: which job did they pick? Who owned the build? And what does their iteration loop look like?

If they can answer all 3 crisply, they're shipping. If they can't, the rollout is stalled, and they haven't admitted it.

13:263 questions to ask before any AI tool gets bought

So, what should you do with this today? I want to make sure you have something concrete you can use this week. Even if you're not currently in the middle of an AI rollout, even though you should be, and maybe especially if you're not.

These are the 3 questions I want you to memorize. They take 30 seconds to ask, and they will save you and your function months of organizational pain. Next time anyone in your org talks about adopting an AI tool, at any level, from your peer in FP&A all the way up to a CFO-driven initiative, ask these 3 questions before anybody signs anything.

Question 1: which specific finance job are we trying to win with this? Not "AI for finance," not "better forecasting." Which specific deliverable does this tool make 20% faster or 20% better? If the answer is generic, the rollout's going to stall on month 3.

Question 2: who in the trenches owns the workflow when the prompts break? Not the sponsor at the top, the actual owner. Name, calendar time, budget for the hours that come out of their existing workload. If the answer is "we'll figure it out," the rollout is going to stall on month 2.

Question 3: what's the iteration loop, and whose calendar is the time coming from? Real finance AI work is 15 rounds of iteration before a workflow ships. If nobody can tell you who's spending those hours and where the hours come from, the rollout is going to stall on the very first edge case.

If the team can't answer all 3 in plain English, the pilot isn't going to work. Bring this up before the contract gets signed, or if you can, bring it up before the vendor demo even happens. Promise you, you'll be the most useful person in the room.

14:57The long game: becoming the finance pro your company can't afford to lose

And listen, this is why this matters past the next pilot. The finance pros who learn to ask these 3 questions don't just get pilots to ship. They become the people their function relies on when AI is on the table. They become the ones their company can't afford to lose, because nobody else in the building is thinking about it this way yet. And that's the long game, and that's the game you want to be playing.