Claude Training for Accountants — AI That Works With the Systems Your Firm Runs On

Twyla Verhelst — VP of Industry Relations and Community, Karbon

00:00Setting the foundation: AI that works with your firm’s systems

I’m really looking forward to this session, where we’re going to set the foundation about how AI works with the systems that your firm already runs on, and then we’re going to spend some time actually building inside live demos with Claude. So no matter where you’re at with your AI journey — whether you’re early, intermediate, or really advanced — there should be something for everyone to take away.

My name is Twyla. I’m a CPA based in Calgary, Alberta, Canada, and a former firm owner. I now lead industry relations and community at Karbon, which in practice means I spend most of my days talking to firm leaders and their teams, learning firsthand how AI is actually showing up in the work — how often it’s there, and how often it’s not. I’ve been teaching AI training for accountants for just over two years, both inside my role at Karbon and through a side business with another accountant, Byron Patrick, called TB Academy. I’ve been teaching accountants about the intersection of technology and where it meets them in their firms for about nine years now, so I’m extremely passionate about the topic of AI and accounting.

00:05The state of AI in accounting: adoption vs. intention

Let’s get into a bit of data from Karbon’s most recent State of AI in Accounting report. Sixty-three percent of accounting professionals believe a firm’s value drops if it doesn’t use AI — that number is up seven percent from last year, which means the profession at large is recognizing the positive impact AI is having on accounting firms. For context, ninety-eight percent of accounting professionals reported that they are already using AI in some form. So we’re not having the “should we” conversation anymore. AI is happening because we’re seeing the real opportunity inside our firms.

Now here’s where it gets interesting. While almost everyone is using AI, only twenty-one percent of firms have a written AI strategy, and only twenty-one percent have an AI policy. A strategy is the shared picture of where AI fits into the firm. A policy is more like the guardrails — it’s about what goes into AI, what you don’t put into AI, what tools you use, and how AI shows up in your client work. One in five firms have either of those things.

And here’s the other interesting wrinkle: firms that have AI policies are actually more concerned about data security than those who don’t. That might be the opposite of what you’d think, but it’s because firms that have done that work aren’t more paranoid — they’re more aware. They have more understanding about what AI should be doing and where there’s potential for data gaps or security challenges, so they’re giving it more attention. So when there’s a question about data security, my answer is: build your AI policy. It makes you aware of where there might be gaps, or where there’s opportunity to tighten things and ensure your tools are set up properly to mitigate the risks of using AI in your firm.

The through line of everything that follows today is that the question isn’t about adoption — it’s about intention. This is very different from giving your teams all the tools and saying “go use AI.” It’s about being strategic and intentional with how you use AI.

08:38Everyone’s in the boat, but not rowing together

This is my best representation of what AI adoption looks like in accounting today. Picture five people in a boat, water rushing, the boat bumpy and rocky, everyone getting wet. You’ve got two people on the left rowing in some synchronous way — they actually look like they’ve got a handle on what they’re doing. You’ve got the gentleman at the front who, despite his best efforts, is rowing backwards and causing friction for his teammates. You’ve got the poor lady on the right trying her best, but if you look closely, the tool she’s using to paddle isn’t exactly the most efficient. And then there’s the guy in the back who’s either thinking “I haven’t got started with AI” or “I don’t have time” — he’s not contributing at all.

All of these folks are navigating this rapid water, getting splashed in the face, maybe not even able to see at times. Everybody’s in that boat together, but they’re not all rowing the same direction. They don’t necessarily have intention or strategy around how they’re going to navigate what feels like a lot of turbulence in AI and accounting right now. And expect that this is going to continue to feel this way for quite some time as AI evolves and we get new use cases inside our firms.

So how do we get into that synchronous rhythm of working together? Intention isn’t a soft idea — it’s a measurable one. Teams with an AI strategy save twenty-four percent more time with AI than firms without one. And firms with an AI policy save twenty-three percent more time. Same tools, same Claude. The difference is whether they wrote down what they were doing and gave the guardrails for their teams.

11:02Three zones: core, lab, and the overlap

Here’s the simplest way I’ve found to think about what AI looks like in firms today: three zones. On the left you’ll find your core — the tech stack your firm already runs on: Karbon, Xero, QuickBooks, your payroll platform, your tax software, your meeting transcriber. On the right is your lab, where you experiment, ideate, and explore — this is Claude and ChatGPT and Lovable, tools working off to the side from your tech stack.

And then there’s the overlap in the middle, where some of the magic inside accounting firms is really starting to happen. This is where you’ve got connectors — whether that’s MCP, API connections, or more and more native AI functionality inside your tooling. Over time, we’re seeing these two circles come together and the overlap getting larger. Most firms recognize that they aren’t going to replace their core tech stack with AI, because that just doesn’t work. And if you’re using AI off in a vacuum, that doesn’t compound and doesn’t give you the efficiencies. So the firms doing this well are leaning into that overlap in the middle, as we see things like connectors and MCP tooling, as well as more native functionality inside the tech you and your team are already using.

12:34Demo setup: Cedar and Vine CPAs

Let’s talk about the builds we’re going to do today. First, we’ll do some SOPs and process documentation. Then we’ll do an advisory prep — using existing information to prepare for an advisory call. And then we’ll take a messy general ledger and create something really beautiful with a dashboard. Some of these I’ve pre-baked, meaning I’ve already run them because they take time inside Claude. Some we’ll do live, so we’ll watch it cook in real time, like a baking show.

The instance of Claude we’re using is built for Cedar and Vine CPAs — specifically for one of the partners, Ryder, so we’ll work from Ryder’s point of view. Cedar and Vine is an accounting firm of about forty-four staff. They specialize in vet and dental practices, so that’s their niche. They use Karbon as their operating system, pull data from QuickBooks and Xero, have Gusto as their payroll provider, and use Vinyl as their meeting transcriber — a typical mid-sized firm with a pretty typical tech stack.

14:50Building and auditing SOPs with Claude

Let’s start with SOPs. This isn’t the sexiest place to start, but for reference, forty-three percent of firms are using AI for their SOPs — it’s the fastest-growing use case in the profession. There’s a real reason for that: SOPs are where firms have the biggest gap between what they say they do and what they actually do. If you’ve ever built or updated your SOPs, or gone to your staff to talk about their processes, you know this is true.

Most firms have a library that looks messy — a bunch of different documents, outdated processes, different formats. You’ll see a PDF, a Word doc, even a markdown file. If this is you, know that you are not alone. So in this instance, I took all the documents from that folder and dropped them into Claude. I had eleven docs, and I said: “Here’s our current onboarding SOPs. There are a few sections that are out of date. Help us bring our SOPs up to date and tell me what’s missing.” Pretty basic prompt. Claude knows we’re from Cedar and Vine already from my instance. It went through and reviewed the various docs, noting things that were missing, the various formats, and documents that exist but are out of date.

Then I said: “Let’s start going through these one by one. Ask me one question at a time.” If you’ve ever used a technique called “interview me,” it’s basically saying: ask me the questions, I’ll answer, and then use my answers to put together the deliverable. I always say “one by one” because a wall of questions coming back is extremely overwhelming. Claude walked through them in priority order, starting with the 1099 process for tax year 2026. It even tried to make my life easier — it noticed a reference to an email thread from Mike and suggested I paste that in as a good starting point.

You can also ask Claude what SOPs you should even have: “What are the standard SOPs we should have in place at Cedar and Vine?” Because it knows our services and who we support, it can reaffirm a good question and suggest the standard library you should build out. SOPs are critical to your foundation as we see more and more AI embraced inside firms.

20:15Turning a meeting transcript into an SOP

Another way to build out your SOPs — and I’ve seen this done beautifully — is to use your meeting transcriber. You begin with a call with the people who normally run that process. In this example, we built out the process for payroll runs. We recorded a call between three people — a junior, a senior, and a manager — talking through the process they go through to run payroll each time, using Vinyl to record the conversation. They use Gusto as their payroll provider.

I took that transcript, dropped it into Claude, and said: “Create an SOP from the attached meeting transcript where three of our team members discussed the process for payroll. Flag anything that was missed on the call” — perhaps an off-cycle pay run, or a journal entry. In any meeting transcript there’s always contextualizing that needs to happen, sorting through the noise, small talk, or someone going down a rabbit hole. Claude worked through it and built it out as the manager, Priya, would draft it. It flagged a few loose ends the team raised but left unresolved, and genuine gaps the call never covered: no approval or segregation of duties, no run cutoff time. It noted where Marcus said something was “gut feel” — hard to put a procedure around. If those things need to be figured out before going further, we’d add them in. The result was a nicely structured, six-page SOP for a payroll run, including the tooling, what to do before you start, and running the payroll.

23:19Projects: folders for your AI context

One thing I want to mention is projects. Honestly, projects — first inside ChatGPT and now inside Claude — were a huge unlock for me. Think of a project as a folder or bucket for things you’re doing inside AI. For example, you can set up a client folder so you keep context together for a particular client. On team and enterprise plans, you can share projects across your entire org, so you don’t need separate projects in separate workspaces — you can share one project for one client across all the team members who touch that client.

You can also set up projects for internal SOPs, marketing of your firm, candidates and HR purposes — there are endless use cases. To find projects, look for the little folder in the sidebar. You can create a new project, add a title and instructions, attach files, and then start typing inside it the same as any other Claude chat. And if you realize partway through a chat that it really should be a project, you can always click “add to project” and either choose an existing one or start a new one right from there.

25:55Why SOPs are the foundation for AI agents

Coming back to SOPs — I’m aware I’m starting with the most boring topic, but they’re the layer that makes everything else work, and nobody really talks about how critical they are. They matter not only today but as your firm gets ready for AI agents. Agents are coming, whether inside Claude or other technology, and cleaning up your SOPs is strategically the most important thing you can do to be ready for AI inside your firm — now and in the future.

These documented processes are what make your operating system so powerful. As agents arrive, the SOPs are what they’ll follow when they’re running parts of your workflow autonomously. If they don’t have an SOP to follow — the foundation of how you want that task executed — you won’t be set up for those agents to be successful. The firms with clean SOPs and documented processes are the ones where agents will be able to do real work, and your team will start to become reviewers rather than doers.

27:28Advisory prep with Cowork and the Karbon MCP

Next is a topic I’m extremely passionate about. In our firm, we were advisory-first, which means every client we worked with had advisory in their engagement — no client had work done that didn’t include advisory. For this example I’m going to use Cowork rather than the Claude web version, just to show a different use case. If you haven’t switched between Chat, Cowork, and Code before, it’s just the little toggles right there in the desktop version.

We’re going to work with a profit and loss statement and a balance sheet, both from March 2026, plus one other interesting piece: the meeting transcript from the previous advisory meeting with this client. Imagine that every quarter we meet with Acme Vet Clinic to talk about their previous and future financials. There’s always context inside those conversations that’s relevant — something small or subtle, or maybe a big theme — that you need to carry into this time. As you work with many clients and time passes, we often just forget. There’s so much gold in transcripts.

So I said: “Here’s the profit and loss, here’s the balance sheet, here’s the meeting transcript from last time.” This instance of Claude also has an MCP server connected to Karbon — we have the beta of the Karbon MCP integrated here. So I added: “The financials are attached, plus last quarter’s advisory call. These other things can all be retrieved from Karbon — any emails that went back and forth between our last calls, any internal communication that took place.” I told it I have a call tomorrow and need to prep. Claude looked at the financials and the transcript, did a little review of the financials, then went out and found Acme Vet Clinic inside Karbon and pulled the records — emails, work items, and so on.

Then something interesting happened with the brand voice. Inside this instance I’ve set up the Cedar and Vine brand voice — how we want to show up for our clients, the language we use: clear and concise, not too much jargon, speaking at the level of our dentist and vet clinic niches. Without me even asking, Claude pulled in that skill and used it. The more you build out your Claude instance, the more valuable, effective, and efficient it becomes.

33:19The advisory deliverables: prep doc and client deck

As it worked, Claude asked what I’d like for a deliverable. I asked for a prep briefing — a Word doc for me to be prepared as Ryder, the partner — and also a client-facing PowerPoint, because I know this client likes something visual. The prep doc did the cursory work I’d typically do when I get financials from a team member: it looked at growth quarter over quarter, margins, cash, AR, and the line of credit. This one’s really interesting — the line of credit’s gone up, the cash has slipped, the gross margin is slipping, and the growth has stalled. There are real insights for me to be aware of, so I feel well prepared for a conversation where they’ve got some interesting things going on with their cash and working capital situation.

The other deliverable was the presentation deck. It’s laid out nicely, with the three most important things at the top and visuals below. It also pulled in some Cedar and Vine elements from our marketing skill. I could refine it further — we have a Cedar and Vine logo it didn’t pull in, and I could request that to make it more on-brand — but it’s quite polished and professional. Depending on the client, sometimes they like the deck upfront, other times it might be overwhelming to send ahead of time.

36:17Skills: reusable prompts you can share across the team

I’ve been mentioning skills a few times. Skills are reusable prompts you can use again and again — anytime you want to do something more than once, you should think about building a skill. This is something newer inside AI tooling, but it’s so powerful because you can share skills across your team, use them inside or outside projects, and use them on desktop, on the web version, in regular Claude chat, and in Cowork. So if you’ve ever prompted something more than once, stop and ask: should this be a skill?

If the way I prep for an advisory call is the way I want my up-and-coming manager — my next great advisor — to prep, I can share that skill. And skills help with consistency, because as humans we aren’t always as consistent as we think we are. In this Claude instance, under Customize and then Skills, I’ve got a brand voice skill, a marketing skill, and an SOP library auditor.

You can turn any chat thread into a skill, or create one net new. To add a skill, click Create skill — you can upload a skill someone has given you (we’ve built skills at Karbon that we’ve given to Claude and ChatGPT users, since skills are available in ChatGPT too), create one with Claude, or write one yourself. Personally, the best way I’ve found to build a skill is inside your chat. Take that payroll example where I had Claude build an SOP from a transcript — once we’d refined and finessed it for Cedar and Vine, I could just say “turn this into an SOP-building skill,” and it builds a skill I can reference over and over.

Coming back to the framework: your client knowledge isn’t just about the financials. There’s real gold inside transcripts and inside your practice management system that goes beyond the profit and loss and balance sheet. Have Claude read across these various tools and pieces of intel to give it the context to go above and beyond for your clients — and build those skills as you go.

40:50Turning a messy GL into an interactive dashboard

One more demo, and this one always lands. I’ve taken two general ledger exports from QuickBooks Online — an April and a March GL for the same client, A1A Dental Services. Straight from QBO they’re just spreadsheet files, nothing glamorous; in fact it’s a little overwhelming, about five hundred lines each.

I said: “I have a meeting with the owner of this dental practice and I want to present their financial information in an interactive HTML page.” HTML is the language websites are built on, so I want this as a kind of website I can use to talk through the financials. I said: “Use the information in the attached files. Create something that highlights the financial outcomes and allows us to dig into the underlying details that will drive our advisory meeting. Make it interactive, give it a dental theme, and create it in light mode.”

A tip there: if you have a strong preference, give it that extra piece of information — like light mode versus dark mode — so it doesn’t build the whole thing and then you have to redo it. If you’ve been using AI a lot, you know you can burn through a lot of credits, and you have limits on certain plans. The more specific you are about what you want, the closer you get to your desired outcome and the more you save on redoing it.

For the pre-baked version with the exact same prompt, Claude went through the brand palette — pulling from the marketing skill I already have set up — caught a minor bug, fixed it, and worked through the coding, which is why it takes a bit of time. It gives you a preview, and you can download it as an HTML file to open in your browser.

45:35What the dashboard unlocks for client conversations

Here’s how it came together: the company at the top, March and April both months combined, and the key things to highlight up top. It built a cash flow waterfall, which I love — as you’re talking to a client, if something seems really off, they’ll often ask “what’s in there? Why is that twenty-nine thousand this month?” You can drill in and say, “that’s because on March seventeenth there was a large bill for twenty-three thousand dollars,” and usually that’s enough for the client to say “oh yeah, we stocked up on some things.” It also flagged uncategorized charges and a suspense item to be taken care of. And there’s a plain-language summary, using the plain language from the Cedar and Vine tone and voice, where I’ve told it to always use really plain language for our clients.

That demo is truly one of my favorites — not just because it’s a nerdy dashboard, but because of what AI unlocks for elevating the client experience. I think about times inside our advisory firm when I was trying to explain profit margin or EBITDA and I could see the client’s eyes glaze over. I thought about how I could present that visually, but I didn’t have the tools or the time to build it. Every accountant wishes they could spend more time with clients whose numbers tell a story that nobody has the time to put in a way the client will digest. This means less formatting and more time for conversation, letting the numbers showcase themselves so you can get into the details underneath and talk about what matters to that business owner.

48:11Closing: intention over tricks

We’re going to land where we started. The firms winning the most with AI aren’t the ones using it the most — they’re the ones using it with the most intention. So start inside the systems your firm already runs on. Don’t replace your core; build alongside it and even within it. Then extend with Claude for what your systems don’t do natively, or don’t do natively yet — because that’s constantly changing. That’s the overlap I spoke about in the Venn diagram. And then go back to your firm with a strategy, not just a collection of one-off tricks you try, but a strategy that compounds.

My name is Twyla. I’m VP of Industry Relations and Community at Karbon, and Co-Founder of Education at TB Academy. Thank you for your time today and for leaning into AI. If you have any questions or comments, feel free to drop me a note or connect with me on LinkedIn.