Automating Crap Is Still Crap: What a CFO Should Actually Do About AI

Automating crap is still crap.

I have been saying that for a year and a half, and I have not softened it, because nothing I have seen since has given me a reason to.

A CFO reached out to me not long ago asking how to build an “AI strategy.” I asked him what problems he was trying to solve. He couldn’t name one. He just knew his board was asking about it.

That conversation is the entire state of the market right now, compressed into thirty seconds.

Your board asked a question that doesn’t have a product answer

The pressure is real, and I’m not dismissing it. Clients tell us their boards want an AI plan. RFPs now explicitly ask how we use AI in our daily workflows. Nobody wants to be the finance chief who shows up in the fourth quarter with nothing.

So the temptation is to bolt AI onto something. Anything. Check the box, name the vendor, put a slide in the deck.

It’s easy for a company to spend six figures on AI tools that sit unused because it skipped the fundamentals. Chasing the hype is a lot more tempting than doing the hard work of understanding your own operations.

And the purchase feels like an answer. That’s what makes it dangerous. A signed contract closes the board conversation for one quarter and reopens it, worse, in three.

Name five problems before you name a single tool

What I told that CFO was to go identify his top five pain points before designing anything.

Where are your people spending hours on manual work? Where are errors costing you money? Where is information getting lost between departments?

Make sure there’s an actual problem before you buy a solution.

The other mistake we see constantly is casting the net too wide, trying to automate everything at once. That’s boiling the ocean. Dip your toes in first. Baby steps before you leap.

When companies do this well, the candidates that surface are unglamorous and specific. Accounts payable, where manual invoice processing and approval bottlenecks eat days. Month-end close support: reconciliations, repetitive journal entries, and flux analysis — the work of explaining why a balance moved from one period to the next. Reporting, where someone is manually pulling data to rebuild the same dashboard every month.

None of that makes headlines. All of it produces outcomes a board can see. Tie each initiative to a number they already track — close time, reporting accuracy — and the AI conversation stops being a status update and starts being a business case.

What happens when you skip this step

I had a different CFO tell me he wanted to implement AI to increase efficiency. Great goal. So I asked to see his current process.

Three different departments were entering the same data in different formats. Nobody had documented who was responsible for what. And reconciliation was happening in someone’s personal spreadsheet.

Do you think AI is going to fix that? No. It’s going to amplify every inefficiency and inconsistency you already have — at machine speed, on a subscription.

Dirty data isn’t an abstraction, either. I’ve watched the same customer exist twice in a system under slightly different names, an integration run, and two invoices go out to one client. They call, furious, wondering why they’re being charged twice. Your team scrambles, and your credibility takes the hit. That’s a duplicate record. Now imagine automating on top of a hundred of them.

Why AI can’t accelerate a close that lives in Excel

This is the part most people get wrong, and it’s the most important thing in this essay.

When finance teams ask about accelerating the close with AI, they usually think the barrier is a lack of a good checklist with dependencies and approvals. Checklists are nice. They help. But a checklist is one level up from the problem you actually have to solve.

The real problem is one accounting teams create accidentally. I’m an accountant, so I’m allowed to say it.

What we see over and over is out-of-system analysis. Allocations, accrual support, revenue schedules, reconciliations — worked out in Excel, using data that may not even be in the ERP. The closing activity is happening entirely outside the purview of the system.

Here’s the mechanism, and it’s not complicated. Software reasons over data it can reach. If the work lives in a workbook on someone’s desktop, then as far as your ERP is concerned, that work does not exist. There is no transaction record to read. No field to query. No history from which to learn a pattern. No approval trail, no version, no evidence of who touched what. A spreadsheet is a room in your building that the wiring never reached — you can install the smartest switch on the market, and nothing in that room turns on.

Adding AI to that mix isn’t going to be much of an accelerant. There is nothing there for it to be intelligent about.

It gets worse when the spreadsheet starts feeding the financials. A team reports out of the system but relies on top-side entries — adjustments layered on top of what the system produces, to make the numbers work. No shame. We get it. It starts with one adjustment. Then a few more. Before long there’s a shadow version of your books living in a spreadsheet.

And then it breaks. Audits get messy. A formula error moves a reported number. And any tool you point at your ERP is now reporting on figures that are not the figures you actually publish.

There’s a subtler version of this that catches sophisticated teams. Exception detection — the feature everyone is excited about — reads what’s in the system. It does not know what never made it in. We ran a health check for one client using a connected accounts payable application, pulled the AP aging in their ERP, pulled the same report in the AP tool, and found the two didn’t tie by over a million dollars. Nobody had told them to monitor for synchronization exceptions. Their aging report inside the ERP was internally consistent. It just wasn’t complete, and no amount of intelligence applied to that report would have said so.

So before you evaluate a single tool: identify your processes, identify where your data actually lives, and decide the most purposeful way to catalogue that data in the ERP.

Then throw AI and automation at it.

But don’t expect a fast close if you’re still working in Excel.

What fixing the foundation actually looks like

Document your workflows. Eliminate redundancies. Standardize your data. Get clear on roles and responsibilities. Build a proactive process for finding and killing duplicate records before they multiply.

I say the same thing in every kickoff meeting: the devil is in the data.

And I’ll tell you that we had to do this to ourselves. A few years ago Rapid Cloud Partners was growing faster than our internal muscle could keep up with. Too much depended on tribal knowledge — the operating instructions that live in people’s heads instead of anywhere you could hand to a new hire. Decisions lived in my head. Processes lived in conversations. Which was ironic, given what we do for a living. So we slowed down and did the unglamorous work: documented how we actually operate, made decisions explicit, built frameworks so progress didn’t require me in every room.

For finance specifically, we’ve learned to start with the financial truth and design backwards. What needs to hit the general ledger, when, and at what level of detail? Then architect the upstream systems to deliver exactly that. If your operations and finance teams are constantly reconciling discrepancies, you don’t have a tooling problem. You have an architecture problem.

And then the payoff is real

I want to be clear that I am not the skeptic in this conversation. I am the one telling you to do it in an order that works.

At Rapid Cloud Partners we are actively adopting AI across how we gather requirements, document processes, analyze data, and build deliverables — and honestly, we cannot move fast enough. The goal was never to replace our team’s expertise. It’s to make that expertise more valuable, by taking the formatting and compiling and repetitive analysis off their desks so they can solve harder problems.

The same logic applies to your finance team. They would rather have their time back than have a new dashboard to learn.

Consider what that looks like when the foundation work comes first. A roughly 250-person services company ran labor cost allocation entirely in spreadsheets for five years — about 30,000 rows of time entries every month, mapped by hand to labor rates across multiple entities. It worked, sort of. It was also error-prone and couldn’t keep pace with growth. We integrated their time-entry system with their ERP so the hours ingested automatically and the system produced the journal entries. Their close went from six days to five. More valuable than the day: they stopped being blind mid-month. They now pull data weekly and can see how they’re tracking against budget before the month is even over.

Notice that none of that was AI. That’s the point. That is the work that makes AI worth buying a year later — because now the data is in the system, catalogued, and reconciled, and a matching assistant or an exception-detection model has something real to read.

The tools arriving in modern ERPs are genuinely good. On-demand bank feeds, transaction matching that recommends the likeliest candidate when the rules leave several open, close managers that surface exceptions and estimated impacts in one view. We enable these in a sandbox first, with the client’s own data and configuration, so their team learns how the features behave before anything touches production. Every release brings opportunities. Not every feature is right for every organization.

The ruling

Fix your processes. Clean your data. Get it into the system. Then automate. Then add AI. Then talk about innovation.

Process comes before platform, every single time.

And if you cannot name the five things in your business that hurt the most, you do not have an AI problem. You have a clarity problem, and there is no vendor selling a fix for that.

Automating crap is still crap.