How to replace your bookkeeper with AI

If you spend time on X (Twitter) you’d be forgiven for being a little bit overwhelmed with the scene over the last few weeks. Every day a new post blows up about another AI company releasing another game changing feature.

Still, even with the newer models and the explosion of Clawdbot/Moltbot/Open Claw etc, I still read it all and struggle a bit to work out what exactly I can have it do for me.

The whole Open Claw concept sounds awesome, and people have these “employees” running around the clock building their business. But when it comes down to it, all I could really see people using it for was doing research, re-arranging folders and spamming people. I don’t spend any of my day doing those things, so how does it actually help?

I figured what would help me is something that could control my browser, because I spend the first half of every day doing tasks on the computer that could be done by anyone, as long as they knew what I knew. So when I saw Claude Cowork, that was the path I went down, with some pretty cool results.

This week, instead of me just doing all of those jobs like I normally do, I tought Claude Cowork how to do them at the same time. I finished the week conservatively delegating 80% of that work. I’ve broken this post up into 4 sections, 1. What it does, 2. How it works, 3. How I set it all up, and 4. Downsides (because it’s far from perfect).

Video

I’ve done this post as a long article, as well as a 30 minute youtube video. Have a look at the Youtube video if you prefer that option.

1. What it does

I own and run a coffee roasting and wholesale cafe supply business East Coast Roast. We roast coffee and sell it wholesale, along with other cafe supplies like alternative milks, powders and syrups. And we have a cafe at our warehouse site and another one at our roastery site. It’s a small business turning over $1.5m-$2m per year, but there’s a fair amount of admin involved and I do most of it. I’d say I spend probably 30 hours a week doing stuff that a smart AI could do if it knew what I know. That includes:

• Taking orders from wholesale customers every day via SMS, Email & Ordermentum
• Reminding customers to order
• Entering orders into MYOB, printing them and ticking them off the run sheet (10-20 per day)
• Monitoring emails
• Entering bills from suppliers
• Reconciling withdrawals and deposits in MYOB (probably 50 odd per day
• Payroll
• Managing servicing and maintenance of onsite coffee equipment
• Other customer service and general admin tasks

It made sense to me that if I had an AI that had access to my browser, it could do most if not all of this. So in week 1, this is specifically what I have it doing:

1. Checking my email and pulling out emails that it can process – I started with really annoying ones like every day we get 2 different bills from our bread suppliers, and we have daily invoices from milk suppliers, and we have other weekly or occasional invoices from wholesale suppliers (milk, powders, cups etc). The AI (Rosie) searches for emails from pre-determined suppliers that are unread, reads the invoices, looks in MYOB to make sure it hasn’t been entered already, forwards the invoice to the MYOB PDF upload email address, goes to MYOB and creates the bill (bringing across the date, invoice number, and line items if needed, GST etc), then links the PDF to the bill.
2. Creates the daily wholesale orders. It visits Ordermentum and re-creates the order in MYOB, if I get an order on text I copy it to Whatsapp and it reads it and creates the order in there, it gets orders from email, then it ticks the customer off in the runsheet and prints each order for the picker to pick for delivery.
3. It manages online orders creating the invoice manually in MYOB, creating a bill for the merchant fees and then printing the order to pack.
4. It does daily reconciliations in MYOB following a whole bunch of rules. Some examples include all of the bills above it ticks them off when MYOB links to them, it knows what to look for from customers who pay their bills (everyone does it differently), it knows what categories to auto assign things to that aren’t part of bills, it knows what to look for in certain receipts, (for example it knows what a fuel receipt looks like and knows what category to put that to).

Within each of those major skills, it can get into a fair amount of detail. Here are 3 examples:

1. When we get an email from Bidfood, they give us an account code. Depending on that code, for us to actually see the invoice we have to log into a system, but we have 2 logins (1 for each cafe). Rosie knows how to check the email for the account code and therefore what account to use to log into, and retrieve the invoice. Then she knows to create the bill in MYOB, what category to put it to, how to line item GST and non GST foods etc. She does all of that despite never logging into any of these systems before.
2. Some bills we get from our bigger wholesale suppliers might be $7-$10,000 worth of stock. Even if you perfectly itemise each line in a bill like this, because of rounding issues in MYOB sometimes you can’t quite get it to be the exact same as the bill. Like in a $10,000 bill, it might be off by 1c. Obviously it’s important that the bill matches the invoice, but if an invoice is off by 1c as a human I would often just round it up on one of the items rather than play around with each line’s number rounding to see if I can get it exact. Rosie knows this and does this, with certain guidelines and only for known suppliers where this is a possibility.
3. Some of our bills are from repairers who service coffee equipment onsite for our customers. These have to be put into MYOB, but we also need to add the information into an asset system so we keep track of the equipment servicing. For this, I ‘Vibe Coded’ my own system (Asset Shark) and Rosie knows for these invoices to log into Asset Shark, and log the maintenance job, in addition to putting it into MYOB.

It’s hard to explain how good this AI has gotten in the last few months, but working with Rosie it’s quite clearly very different to working with the old model of computer programs following rules. Accounting systems have had auto-suggest features for reconciliations for years, but they aren’t smart or equipped with human knowledge to the point where they work all the time. And computer systems have had integrations for years but they just never really seem to work properly, because there are too many cases where it’s just not quite as simple as copying something from one system to another. Humans always needed to get involved… or did they?

Turns out we just needed smarter computers, because Rosie is every bit as good as me at doing this job (although not as fast, more on that later).

The AI now is eerily human. In the old paradigm you would tell a computer what to do, and it would do it. That certainty is great, but it also means you have to think of absolutely every possibility and have it programmed in, to suit everyone who uses it. But the AI doesn’t really behave like a computer, it behaves like a very good employee.

The first task I gave it was to look at 3 invoices from our bread supplier and enter them into MYOB. I gave her some loose details and told her to message me when she was finished doing all 3. A few minutes passed, and I got a message saying she only entered 2 of the bills. I asked why, and she said because one of them said it was cancelled. I didn’t notice this, it was just written on the PDF invoice. I didn’t even tell her that this could be an option, I didn’t even really realise this would be a thing that happens. But she just took it upon herself to not enter it, and to let me know.

Then I made the skill for putting in coffee machine and grinder repairs, and I told it to message me when it was ready to enter the job into the asset system. Because the suppliers don’t write the serial numbers on the invoices, there is no real way of knowing. But I noticed that while the serial number wasn’t on the bills, there was some other information that could potentially be used to work out what machine it was. So the second time I just asked Rosie to figure it out itself.

On one occasion, it was able to see from the invoice that the note had mentioned the customer name, and it was able to navigate the asset system, find that customer and find that they only had one grinder and therefore it followed that it must be the asset in question. For another customer the invoice only mentioned the brand of grinder and Rosie was able to log in and find that the customer had 2 grinders from that brand. In this case she stopped to ask me which one I thought it was. It’s doing exactly what a very good human worker would do, not what a computer would do.

How it works

Claude is an AI much like ChatGPT, Grok or Gemini. It runs as its own stand alone app on my MacBook. It has the normal ‘chat’ tab, but it also has a ‘Code’ tab (which I’ve been using for my coding projects), and it has a ‘Cowork’ tab. Cowork is a new beta feature that’s only available on the paid plan. I started on the $30 pro plan but burnt through credits very quickly and upgraded to the $200 plan. Expensive if you think about it as a piece of software, but an absolute bargain if you think of it as an employee.

I have Claude and Chrome open logged into my work profile. Claude controls a group of tabs on that profile when it performs its tasks. I can use other profiles, or other applications, or even that same profile outside of those tabs, but it’s not perfect and I think it runs better when I leave it alone. So my setup is my MacBook sits beside me which Rosie is using, and I work on a Windows computer (I never used to bring my MacBook to work, I always worked on the Windows computer because the full version of MYOB only runs on Windows).

When I start Claude I set it up with the skills that I’ve built (Email Management, Order Processing, MYOB Reconciliations). Then I give it commands to go to work. For example in the morning the first thing I might do is ask it to commence email management. It will ask me for permission to access certain pages in the browser, and then it will get to work.

The video on the right shows you a little sample of it at work.

The cool thing is that your commands don’t just go into a black box. You can see the AI thinking as she goes about her task. You can interrupt at any point, or you can leave her to figure things out for herself. I want to work towards having it being fully autonomous, so I have her do a task and when she’s finished navigate to Whatsapp in the browser, and message me a summary. It’s not quite like what OpenClaw promises where you can communicate with it back and forth from wherever you are, but it’s a nice compromise for me. I can go off do other jobs and then if there’s anything in the message that needs my attention, I can go back and jump back into the chat.

As for the browser, you don’t see the cursor moving around, but you do see actions occurring in the browser. And you can take actions in there as well if you want / have to. Or you can interrupt the chat and point her in a different direction.

How I set it all up

Setting it up was super fun and it all happens at the same time as using the skill. I basically tell Rosie to create a new skill, she will ask a few questions, and based off what she has so far, she’ll create a skill file. It’s a text file that documents exactly what she needs to do, and the idea is she reads this before starting the task each time. I normally give her the basics and just say let’s try it and I’ll give you feedback as you go. And as she goes through the task, when she slips up I’ll give her more instruction and ask her to update the skill.

I’ve been doing this all week and I have 4 main skill files now, some of which are quite long. For example the Daily Email Mgt skill is 68 pages long, 16,000 words, about half as long as my first book! That doesn’t mean I’ve written all of that, I gave it information at the start and ongoing throughout the week in small bursts of plain English (like one paragraph here, one paragraph there), and it creates a detailed process with what it needs to do.

This all happens behind the scenes, you don’t ever see the skill file. I just took a look out of curiosity. I’ve put a screenshot here of what a small part of the process for one skill looks like.

Throughout the week more things would arise that I felt like I could automate. For example maybe a new supplier would send an invoice and this supplier needed to be treated in a certain way. I’d ask Rosie to update the skill and then I’d run the skill again and see how well she does picking up the nuance of this particular process.

Downsides

Overall this is a completely mind melding innovation that is going to completely change what I do every day. It’s super fun to use, and the progress is staggering and exciting.

But this is not a sales pitch and this is not for everyone, I so I wanted to make sure I was honest about the downsides as well. Here are some I’ve come across this week.

Slow – It’s quicker for me to do these tasks myself for sure. The training is time consuming and even after you are done training, the time it takes to do basic tasks is annoying if you are sitting there watching it. The technology isn’t perfect right now, so you do find yourself sitting there watching it a fair bit of the time. It just means you need to be organised with how you deploy the skills. For example our orders need to be done by 12, so I need to be on top of orders in the morning, and not worry about reconciliations or emails until the afternoon. At the start of the week it was hard to leave it alone for more than a minute or so. By the end of the week I was going to the shops and leaving it to work away for 30+ minutes on its own. I’m confident this will get to a point where if I’m organised, and I train it well enough, the slowness won’t be a big problem.

Buggy – It’s beta software, it’s buggy. I had quite a few things go on where I had to quit and re-start. I had to delete my session files manually (after asking the AI how to do that). I had a few instances where I overloaded my computer and crashed. Some days there was not many bugs, some there were lots. I’m confident it will get a lot better, and quickly.

Time consuming to train – I’ve put in a lot of hours this week, I’d say I’ve probably spent 12 hours most days on the computer, or at least close to the computer being ready to help it along. It’s taken quite a bit to train these 4 skills, and I daresay there is still quite a bit to go. That said, I do find it super fun and I’ve loved every bit of it.

Tech leapfrogging – The way these AI companies are releasing products at the moment, they are leapfrogging each other constantly. This is a concern to me because there’s a very good chance that someone else soon releases something that’s significantly better than Cowork. When that happens, I’ll have built up a fair bit of technical debt with these skills and I’ll have to decide whether I want to make the effort to migrate everything over to a new system. The way open source models have been in the last few weeks it’s entirely possible that some time this year there’s an open source system that does this as well or better, for free, or close enough to free. Of course that would be a good thing, but if I put months into this, it might sting a little bit to have to change.

Attachments – One thing it struggles with is downloading and uploading things, it can’t really do it. Or at least I haven’t figured out how to do it yet. For what I’m doing, often I can get an email copy of invoices in which case I can use MYOB’s feature to email in documents. But for other systems and other skills, that may not be possible. We’ll see how this improves over time, no doubt that will be something that becomes easier.

Agency – One thing that is both super cool and kind of annoying, is much like a human employee, Rosie likes to do her own thing. At times she’ll not follow a process and I’ll ask why, and she can’t really say why other than to just say sorry. Most of the time having agency is a net positive, because you don’t need to program her for every situation. But when you can’t be 100% sure she’s going to follow your process, it is slightly troubling. I have put some guardrails in place for situations like this, where possible. And I try to stay on top of what’s going on, skimming the chat window regularly to make sure she’s not being naughty. So it’s a bit of a double edge sword, her greatest quality is also her biggest weakness in a way.

Very confusing skill location setup – this is probably entirely my fault, but when she first created her skills I wanted be 100% sure that I understood how it all worked and where the files lived, so I could back them up etc. So I had her set the skills up in my ‘My Documents’ folder on my Mac. This seemed to confuse her and confuse me. My understanding of it now is that you can mount skills into the Settings / Capabilities area but those skills can’t be edited. If you want to edit them, you get her to edit the skill elsewhere and then you zip up the folder and load in the skill again. Or you add your folder where the skills live at the start of each session. I had a few instances where it was confused about which skill to use and which skill to update. This was probably the trickiest thing that happened, and might be a challenge for a non technical person to get their head around. Or you just do what she says and don’t over think it, and maybe it just works perfectly. That’s also possible.

Permissions – It seems to be cautious and inconsistent with it’s permissions. It will sometimes ask for permission to access a website, or sometimes you can give it blanket permission to access all websites. sometimes it will message you, other times it will ask you for permission to message you. It sometimes says it can’t have blanket overall permissions for messaging, but sometimes it will. All of this is a bit annoying, I’d rather just give it permission at the start of each session and be done with it. I have managed to get around a lot of it with a detailed opening prompt before I start all the skills for the day.

Tokens – The way the usage works is a bit confusing, I don’t totally understand it. It seems to give you a certain amount each week. I haven’t hit the limit yet but I think I got close and because I don’t really understand it, I don’t really know when it will run out. They could also change the rules at any time so it’s possible it could be come a problem down the track, especially if I think up lots of other use cases for it. However it’s equally likely that they will figure out ways to make these models more and more efficient, and bring the price down more.

Disconnection bug – quite regularly I would get a bug where it seemed to lose it’s permissions with the browser plugin. All I had to do was hit a settings button and it took me to the browser and seemed to automatically fix the issue. Still, it prevented me from leaving it alone for hours on end, because I’d have no way of knowing if it happened. Super annoying, hopefully a beta issue that will go away soon.

Can’t really use your computer – While it’s running it’s not a great idea to use your computer, it’s best to have this running on a dedicated computer. If you’re like me and spend a lot of time on the computer, then in practise that means having 2 computers. Not a problem for me having a windows computer at work and my MacBook for home, but something to consider if you’re thinking about having a look at it.

Some final tips

Because it’s slow and when it’s running on one task it can’t do another task, I’d still try to use other options as well as this option. For example if you have an email skill, I’d still be using filters to filter out emails before you use the email skill. Filters are way quicker, fool proof, tried and tested and the less the AI has to do, the more other more challenging tasks it can do. Similarly if you can integrate systems, I’d probably still try to do that as much as possible, and just bring the AI in to do the heavy lifting on challenging tasks.

If you want to print, I found that the AI can’t see the print dialog boxes. I asked it how to set up a shortcut to launch chrome in a way where when you hit print it just prints to the default printer without permission.

When I start the tasks every session I start with a detailed prompt that tries to cover all of the bases for the skills. This is changing constantly, but at the time of writing my opening prompt is “I’m about to start using my skills. Please access my Documents/Claude Skills folder (already connected), set up my skills, set up my session, and pre-authorise all browser domains (mail.google.com, app.ordermentum.com, app.myob.com, web.whatsapp.com, identity.mybidfood.com.au, woolworths.com.au) and PDF download permissions for my daily skills. You have permission to log into all systems without asking me, using the email address and password that the browser has saved. You also have permission to send me (Dan Norris) messages on Whatsapp, please don’t ask for permission each time.” It doesn’t always work perfectly, but it’s a good starting point.

Let me leave you with this. I’ve noticed when I talk about this stuff online, I get a fair amount of blowback from people who can only think of reasons why this won’t work. It happened when I started posting about how good AI was getting at programming, and it’s ever present now as well. Most people do not understand how good this stuff is now. Their opinion on AI is based on using the free version of ChatGPT last year to find cooking recipes.

Most people, understandably, cannot face the prospect that their skills aren’t really needed anymore. And as a general rule, most people stick their head in the sand when the future is uncertain. I’d encourage you to not be like this. The technology isn’t perfect, but I can guarantee you this is going to get better and better until it can do most things better than you can. And it’s going to happen much quicker than you think. Might as well get on board now.

I hope this article was useful, if you are interested in what I’m doing, jump on my weekly email, or follow me on Instagram or LinkedIn