This Boring AI Automation Could Save Your Accounting Firm 40 hours (But most ignore it)

Are you running an accounting firm that's still spending hundreds of hours on manual work your team could be doing in minutes? Well, that's exactly what happened with one of our recent clients. We took a process that became 40 hours monthly and built an AI automation around it that allows them to get that same work done in just 15 minutes. Now, it's not an AI flashy skill. In fact, it's one of the most boring ones that we do, but it's extremely effective and it's specifically in the realm of bank reconciliation. In this video, I want to break down what an AI skill is, how it works, and how you can implement it for your firm in the context of that specific use case we For a client that took a process from 40 hours to 15 minutes. But first, we need to talk about why this is crucial to implement in the first place. You see, as the owner of an accounting firm, you know the biggest time drains usually aren't the difficult problems, right? This is something we see in almost every efficient audit we run. Instead of spending time on the advisory work, the reviews, the client calls, the nuances of the client, the client conversations, that effort and bandwidth usually gets used up by manual tasks no one ever stops to question. Now, one of those tasks is bank reconciliation, because the numbers never show up in the bank the way they left on your account. For example, a payment goes out as one clean number, but then it somehow gets split up and lands on your bank statement across multiple days, somehow even spilling into the next month or the next quarter, and suddenly nothing matches, which means someone has to spend their day digging through the bank statement looking up 5 numbers that potentially add up to this $10,000 batch. That's exactly what was happening to one of our clients when we first met them. Basically every transfer they sent was getting broken up before they hit the bank. So instead of seeing like a $1,000 line, they'd see $1,000 on a Tuesday, $2,000 and $3,000 on a Wednesday, and then $4,000 sometime after that. Now multiply that by dozens of transfers a month and you've got a team spending 40 hours just trying to make their bank statements agree with their books. It's a huge waste of time. So much so that the accounting firm owner of this client told the team, hey team, let's give up. This is completely out of scope. Like, we cannot do this manually anymore. By the time they came to us, they were already 18 months behind. on the client's books. Yet most firms don't realize that doing all of this manually in 2026 is completely unnecessary. Now, if your firm has a workflow like this that's been eating hours per month, or that someone already called out of scope, that's exactly what we solve here at Opsware. And we'd like to see if we can do the same for you. We've helped accounting and bookkeeping firms take their most painful, repetitive parts of their month-end and automate them so they never have to be done by hand again. Just like one CPA we worked with who took 90 hours a month of monthly bookkeeping down to 8 and kept growing her client roster 33% without actually hiring any single new bookkeeper. So if you'd like to see what we could take off your team's plate, feel free to book a call using the first link in the description below. But either way, now that you've seen how messy this gets by hand, we need to talk about how we actually automated it. Because the fix isn't just dropping a bank statement into Claude and typing reconcile this. A reconciliation like this has cutoff dates, has split payments, has some transfers that were declined and never processed, a dozen places where one bad assumption can throw off the entire quarter. So if you just sit down and prompt your way through it, you'll get an answer that looks confident, yet it could be be completely wrong. Why? Because AI, sometimes left to its own devices, will happily force 2 numbers together just to make the totals match. But this is where AI skills come in. A skill is basically a folder with instructions, the rules, and the text that Claude needs to do to do one job properly, which means we write the process down once, we hand Claude the files, and we have it run the same way every single time. Now, I'm gonna show you how it runs on some demo data in a minute. But at first, I wanna walk you through what it actually does because before it starts matching anything, it starts by auditing the files. And as I'll share later, the way it does this, it does it in a way that is deterministic. Effectively, the skill itself, you can introduce code so that it does it reliably as opposed to relying on the model coming up with a version on how to process that data. You can ensure it's processing it consistently every time, just like a software would. That means checking that every batch's individual payments add up to the total the batch says it has. Both the numbers of payments and dollar amounts. Next, it makes sure it has the ACH file. So there's like a transfer that was done for the quarter. We need all that data aggregated and then we need also the 3 monthly bank statements for that particular quarter. And it only looks at the settlement lines on the bank side. So the regular vendor payments and transfers never gets mixed in. Essentially, it first confirms it has all the inputs and necessary information, including all the data from the transfers, all the data from the bank statements, all the data from the accounting ledger, and also the detail of the vendors and customers that each transaction relates to. Once the files check out, it starts matching, working from easiest to hardest. The simplest ones are one payment and one bank line. So if a $1,250 batch goes out and a $1,250 line lands on the statement, it will make that match first. Then it moves to payments that got split across a few bank lines. So if I had a $2,200 batch, Then I have $1,200 on one day and $1,000 on the next. It will make sure to see what are the transactions that when added up match to that specific batch, which is the problem that our client was dealing with. And the trickiest are payments that go out right at the end of the quarter but don't land until the next one. For those, we actually also had a process to look for the next quarter statement to confirm the money actually arrived and it's just a timing mismatch. Now, there may be still times where we can't confirm whether payments match based on the data it has, but in the rare case that happens, it won't just try to make things work for the sake of it, right? It won't hallucinate and try to like tell you it matches when it doesn't. We specifically program it such that if a payment has no bank line it can prove, it gets flagged and tells, hey, this is only on the batch. If a deposit shows up and there's no batch behind it, it's gonna flag it as, hey, this is only in the bank. And if a batch was declined, it will also flag it instead of matching it. And if 2 matches are equally possible, it marks it as ambiguous rather than randomly picking one. In essence, we're giving the system 2 data sources and telling it compare and match everything that is one that is on the other, but then specifically telling it if it is on source 1 but not on source 2, flag that. If it's on source 2 but not on source 1, flag that too. And where there are multiple potential matches, flag that as well. So your team now can look at the items that actually need a human instead of hunting for problems hidden under a total that looks clean. Once the AI is done, it gives you finished Excel file with every batch and every bank line, the match type, the dates, the differences, all in one sheet with a summary at the bottom showing you the unexplained differences and every timing item and flag listed separately. That's like 40 hours of work condensed into 15 minutes. Now let me show you how you can actually use this for yourself. All right, so here we are looking at the example input files and also skill. So in this case, the skill, as we mentioned earlier, is really a folder, but in this case it's compressed into a zip file. If I unzip it, we're going to see a number of things, including a skill.md that has the instructions. But more importantly, as we mentioned earlier, we have a series of Python scripts that allows this process to run a lot more efficiently and a lot more consistently. So the first thing I'll do is I'll go ahead and go into this customize section. You'll notice that I am in skill and I'll go click on this add button. Here's where I'll be able to upload a skill and I'll be able to bring that zip file we just talked about. Here's a zip file. I'll just go ahead and open it. This should allow it to download. I'll see a preview. I will upload. And just like that, I have the skill up and running. Here are the main instructions where you see you have like the reconciliation being done in 2 different stages, each one with its own script. You have a number of inputs that you need, which we'll walk through and show in a second. You have a naming convention as well to make sure we understand everything that's being produced. And it even shows you and tries to explain to you how you go about reading the result so that if you as a human ask about it and the skill was used, you can understand what you're looking at. This is the main part of the skill.md file that is effectively, when necessary, pointing back to specifically which Python script it should use for certain steps when you wanted it to do consistently and reliably and not rely on prompts for those specific steps in the process. So let's actually test it. I'll just start a new session. I'll just slash and look for that ACH reconciliation engine. And I'll just then go ahead and I could either provide that access to that specific folder where I know the dummy data is right here. Or if you wanted, you can just drop those files directly as well. Either way would work. I'll just give it access to that, give it access to the folder it needs, give it access, call out the skill, and I'll do nothing else. I'll just let it run and come back. A second. All right, so approximately 4 minutes later, here's what I've got. It's suggesting the files are now in the output folder where I have 2 files, one that is reconciling the bank transfers with the bank statements and one that is reconciling the bank transfers with the general ledger. If we were to take a quick look, here's the output folder it created on its own. And as you can see here, this is the output that is telling me, hey, rather than looking for this transaction that adds up to $12,373, Guess what? It's telling me these are the 2 items that specifically add up to that. You see some of them that match one-on-one, other ones that match in group, just to give confidence to the client that, hey, all the transfers in the segments are being represented properly. The other one, a more important one, is the one that is this comparing this to the ledger. So in this case, you can see here that it's saying, hey, what are the items that match? What are the items that match but are on a clearing account? And then specifically telling you, hey, is any Anything that was not found at all is anything that was ambiguous, in this case zero. But then you can go ahead and see the exact detail. So, you know, specifically which transactions are there, which ones are not, and which ones you perhaps could just very quickly filter. So ultimately the human being here just has to look at what are the transactions that are not there, as opposed to starting from scratch with potentially thousands of transactions. And that's exactly how you take a bank rec that ate 40 hours and turn it into 15. But here's the thing, this is one of the most boring automations we built, and that is exactly why it works, because none of this needs to be flashy. It's just a clear process that Claude could follow the same way every time. So if you wanna see the other Claude skills we've set up in an accounting firm, feel free to watch this video where I break it down next. I'll see you there and stay efficient.