Windows Broke Claude Overnight. The AI Tools Backup Plan Every FP&A Pro Needs

Full transcript

Lightly edited for readability. The audio is the original.

0:00A Windows update stopped Claude Cowork overnight

Wednesday morning, I sat down at my desk, opened Claude, and nothing I had built would run. A Windows update had installed overnight, and every process I'd rebuilt around Claude stopped with it. My daily briefing, my weekly finance update, all of it just sitting there.

I've spent months telling finance pros to go all in on AI, and I'd go all in again tomorrow. But I had never once asked, what happens on the morning it just isn't there? So in this episode: how you go all in on AI and still keep a way back.

Welcome to Excel at AI in Finance, the show for finance pros who want to get home sooner and be the ones 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. Each week, I share the AI tools, workflows, and mindsets that get you where you want to be. All right, let's get into it.

0:53Rebuilding finance work to be AI native

Let me walk you through that morning and the sheer terror I experienced watching Claude crash on task after task. Now, I'm obviously doing AI for finance, and I like to practice what I preach, so I've rebuilt almost everything I do to be AI native.

Now, what's interesting about that is I'm not going in and saying, here's this Excel file I built, run this. I'm physically redesigning all of my processes to be AI first. And what that means is some of the files don't look great and aren't fully usable by humans.

They're designed for clean data that can be processed by AI. They're designed to support deterministic calculations and formulas, so AI is not doing the math. And they're designed to be easily auditable, so I can tell where all the numbers are coming from. But without the AI component, they don't necessarily produce themselves.

It's for building AI files where you're not replacing the human, you're building it alongside the human, and it's doing what it's best at. But when everything's shut down, the files aren't ready to go and move right to a manual process.

1:56The Claude Cowork shell error and the markdown workaround

So I'm sitting there. I started my day like normal. I was absolutely slammed. Had way more to do on the list than I could have ever gotten to.

And I started kicking off my Claude processes. Some of them are automatic. They start right when I turn on my computer. Some of them I was kicking off. And one by one, they started coming back saying they were failing.

And it was a shell error. Now, I've seen occasional shell errors before. Typically, you just shut Claude down, reopen Claude, and it comes back up. I did that. Still wasn't working.

Then I went to my next tactic, which is I pop over to Claude Code. I ask it to research what's happening. It can actually go. It runs on a different environment. It doesn't run on the shell in Linux the same way that Claude Cowork does.

I had to hop in and check it out. It saw the failure and had no idea what was going on, because it was traditional functionality and it had never seen the error before. I sent it to look on the internet: is anyone else having issues?

I was finding this out along with everybody else, because the update had just come out. Nobody had a solution, and only one or two people were seeing a similar error. Now, keep in mind, as I'm doing all this research, I'm losing 30, 45 minutes, an hour, because I'm trying to get this working, because everything is built around AI. So not only is my work not getting done, I'm not doing the things that I had on my list, the things that AI can't do, because I'm trying to troubleshoot AI.

What I'd like to tell you is that I jumped right into action. I had a fantastic backup plan and I just kept going on about my day. But in reality, I actually kind of crashed out a little bit. I spent another hour trying to look for any possible way, because Claude Code had always been able to get the shell and the Claude Cowork virtual machine back up.

Finally, before accepting that (hey, this isn't coming back up, Claude has no idea what's going on, people are just starting to talk about this online, we might not have a good way to figure this out), I finally figured out that markdown files don't require the shell. The shell is really used to work with files like Word and Excel. So while I couldn't produce those, I could do markdowns.

And I had a quick hack way. I would have Claude Cowork do the work that didn't need the shell and create a markdown file. And then I had Claude Code jump in, which uses Python to build documents, and I would have it do the conversion. So it was a really hacky way.

And then the next day, as there was a lot more internet traffic on this, we started figuring out it was the Windows security update blocking it. So right from Windows, something to make our computers more secure was actually blocking the way Claude Cowork works. And as of right now, as I'm filming this, it still hasn't been patched by Anthropic. So if you didn't roll back that security update, Claude Cowork would still not be functional.

Hopefully by the time you're watching this, that's been fixed. But again, I'm filming this a couple of days later, and this hasn't been resolved.

So I rebuilt all of my work around AI the way that I tell everybody to, but I skipped something that finance is usually really good at. And I want to make sure you don't skip the same thing. So let me walk you through what that looks like and how you avoid my mistake.

5:23Key person risk, the finance rule we dropped for AI

Finance has a rule called key person risk. And what that means is that you don't have any process relying on one key person. Nothing business critical should fail if one key person is, to say it bluntly, hit by a bus.

But for some reason, when we moved to AI, we just dropped that rule, because AI is a tool. And we would never have a backup for Excel, because Excel just works. But AI is a lot more complicated.

AI we need to treat, as I always say, like a finance intern or a junior analyst. And just like you wouldn't have that finance intern or junior analyst being the only person who can get a critical process done, you can't have AI doing that either.

Traditionally, we solve for this with things like cross-training, desk procedures or standard operating procedures, and making sure we have vacation coverage. We even have to answer to auditors on Sarbanes-Oxley control processes in the U.S.: who else can run this?

Like I said, I would still go all in on this tomorrow. I'm still staying all in. But I need to check myself, just like I'm offering you the advice, and make sure there's not this key person risk where everything is centered around one single AI tool.

The tough thing is it only matters when there's a deadline or someone waiting. So we really have to intentionally carve the time out before we get there.

There's three things that can help us solve this, and I'm going to take you through each one. The first is going to be instructions. The second one is going to be a backup tool. And the third one is going to be a manual version.

6:49Backup 1: AI instructions in files you own

The first thing to ask yourself is where your instructions for the task live: prompts, the context, everything you've taught about the business. If all of that lives inside one tool, the day the app stops working, you lose the ability to run that process.

So to back this up, you need to make sure that the instructions don't just live in the single tool. Now, the great thing is you've got the AI tool trained. It knows all of this, and it can help you document it. You don't need to physically write out the procedures yourself.

So ask the AI tool to do two things. Ask it to give you a plain language standard operating procedure that you could hand to anyone on the team to do the work it's doing.

That's going to give you all of the steps that you need to do. That's going to give you all of the guidance, all of the "this is what the CFO or the vice president likes." It's going to give you all that information.

And then ask it to give you a markdown file of how it's running it. If you haven't gotten as advanced as using AI skills, it's essentially creating an AI skill. It's the same thing, just a markdown file of how it's running the process. By putting it in markdown and not in a specific tool's format (for instance, .claude if you're using Claude), you can move it to any other tool.

And keep those two documents: the one for a human to read and the one for another AI tool to read. With these instructions in hand, any person and any tool can pick up from where your AI tool left off.

8:12Backup 2: a second AI tool like Copilot that breaks differently

So the next one I touched on before with instructions, but now we need to pick a second tool. We've never needed a backup to Excel because Excel just works. I can't think of a single time, aside from our physical computers and shared drives not working, that it ever stopped working. It just never lets us down. And so we've been trained on that, that there's always this backup.

Now, your second tool can be a completely different AI tool with different operating rules. It can be a manual process you run in Excel. It just needs to be completely disconnected.

Now, a good example of this: let's say your primary tool is Claude Cowork, and that works. It's largely designed to work on your desktop with the shell, like I mentioned. It does have some options to work in the cloud, but the main way is through the shell on the desktop.

Now, contrast that with Copilot. Microsoft Copilot works entirely in the cloud through OneDrive. Now, that on its own has some different risks, but they're just that: they're different risks.

You wouldn't want to entirely rely on that, because you could have OneDrive issues, or Microsoft could change something. But because it's working in the Microsoft cloud, it's highly unlikely that both would break at the same time. So that makes a good backup.

The critical piece is just always making sure it breaks for different reasons, and that you've got those instructions ready to go where you can very quickly move the work over. Even better if you just have it set up as a backup. Now, that's not feasible for every single process, but for your really critical processes that absolutely need the AI to run, you might want to have those staged and backup ready to go. And then you have your instructions as a third backup if you need to do something else.

9:47Backup 3: the manual version and how long it takes

Now, the last one can be the toughest, and that's making sure that you've got the manual version documented and ready to go with the time it takes. As we roll out more and more AI tools, staffing and individual workloads are adjusting for the work that the AI tools can do. And that means as you look at a close calendar, the close calendar is built around how you've been doing the work, not how you might have to do the work if something goes wrong.

So if something does go drastically wrong that the instructions can't solve and the backup tool can't solve, it's critical that you have the time documented, so you know what the business critical processes will take to get done if they have to be done entirely manually. Now, your instructions from the first step are a key piece of this. That's how you do it. But then you need to know how long it's going to take, because after 6 months of having AI do it, you're frankly not going to remember.

Now, again, this isn't feasible for every single process you're running, but the really critical ones: the reconciliation that is needed to close the books, the month end report that has to go to the audit committee on a certain day. There's those critical processes that will stop every other process in its tracks or miss regulatory or governance deadlines.

You need to make sure that at least every 6 months, you're testing to see if the manual version is sustainable and how long it will take. Other work might have to fall off, but at least you'll know what you're getting into.

I'd go all in on AI again tomorrow. I'd just keep a way back: the instructions in files I own, a second tool I've tested that breaks for a different reason, and the manual version, written down with the hours next to it. You would never let one person be the only one who can run the close. So don't let one tool be that person either.

Everything on that list is a finance habit you already have. Just point it at a tool. The hard part is doing it on a week when everything works, because that's the only week you'll have the time.

11:43Three actions to run before your next close

All right, so what do you do with this before your next close? Three things, about an hour in total, depending on how complex it is. With this, you'll know how exposed you are.

Action one: write down every process that runs through your AI tool and circle the ones with a deadline or a person waiting on the other end. For the top one, move its instructions into a file you own.

Action two: run that same process once through your second tool this week, start to finish, and write down what broke. And action three: do it once by hand with a timer running, and put the steps and the time on one page somewhere your cover can find it.

The first two feel like progress and you'll probably do them. The third one feels like a wasted afternoon because the AI version works fine today. But that page is the one you'll be looking for the next time an update installs overnight.

The finance pros who keep a way back don't just keep their seat. They become the ones their companies can't afford to lose.