For every AI need, a small accounting or bookkeeping firm has three options: build it, buy it, or leave it alone. Firm leaders from LiveCA, VP CPAs, Juna Financial Solutions, Accounting Therapy, First Steps Financial, BKC and Grow Group Global explain how they decide, and why maintenance should be the deciding factor. Part of Two Cents Magazine Issue 02: Artificial Intelligence.
Chad Davis’s most expensive lesson in building tools came before AI hit the mainstream. In 2019, Davis, co-founder of the remote Canadian firm LiveCA, acted on a longtime plan to make the firm more efficient.
The idea was to consolidate all its spreadsheet-based processes into a single app. He carved out budget, hired a developer, and put four or five employees on the project. Then they scrapped the whole thing.
Chad Davis
But looking back, Davis saw a more fundamental issue. The firm’s data wasn’t centralized enough to support what it wanted to build.
Chad Davis
So when Davis tried again after AI arrived, he started with the data, and he started small. Today Davis builds tools for his firm and leads AutomationTown, a community that helps accountants learn to build their own automations and tools.
Yet he’s the first to ask whether he should build anything at all. “This tech exists now that can pretty much do the majority of anything you wanted to do,” Davis says. “It’s just like, should you?”
Davis’s 2019 project raised a question every firm has to answer: what will this actually return? Most firms using AI still can’t say. In Financial Cents’ 2026 survey of 486 bookkeeping and accounting professionals, only 1 in 5 firms using AI could point to a measurable return on investment. About half called the payoff promising but hard to quantify.
When nearly anything can be built, the pull to build everything gets stronger. Davis knows how that plays out: automate everything, he says, and you “cause more work than it’s actually saving.” The key to an efficient AI tech stack is deciding, for every need, what to build, what to buy, and what to leave alone.
Build, Buy, or Brush Aside: Three Choices for Every AI Need
Ask most firm leaders what they want AI to do, and the answer will sound like a shopping list: draft emails, summarize meetings, reconcile accounts, answer client questions.
What the list leaves out is the decision behind these items. For each capability, there are really only three options:
- Build: Create something custom—a script, a custom assistant, or a full app.
- Buy: Adopt an off-the-shelf tool, or switch on the AI inside software you already use.
- Brush aside: Skip it, or leave a working tool exactly as it is.
Instead of weighing all three, many firms default to one. Some go straight to build. Others buy whatever’s newest. And a few stop to consider the third option.
Build is way easier today, and that ease is what makes the pull to build everything so strong.
Buy is where most firms using AI already are. In the AI survey, 96% use a general-purpose assistant, and 58% use AI features built into their general ledger software. Only 20% use a tool built specifically for bookkeeping.
Brushing something aside gets the least consideration, but it can be just as smart a call as building or buying.
“Not everything’s going to be solved by AI,” says Deborah Harris, CEO of Grow Group Global. “Sometimes we should just actually look at the process and abort some of the crazy things we do in the first place.”
When new AI tools appear every week, doing nothing can feel like falling behind, and few firms are willing to do that. The same Financial Cents AI survey found that 95% of firms are somewhere on the AI adoption curve, and only 5% are sitting it out.
The best stacks come from asking, need by need, which of the three moves each one actually warrants. Sometimes that’s a custom build. Sometimes it’s an off-the-shelf tool. And sometimes it’s leaving things alone.
Buy: Why Off-the-Shelf AI Should Be Your Default
Before you build anything, take a look at what you’re currently paying for. There’s a good chance AI is already working its way into your firm through software you use every day.
Davis suggests starting with a simple question for your existing providers: “Are you using AI?” As he points out, firms may already be sending work through AI models without thinking of those products as AI tools.
Chad Davis
Run that check before any new purchase, and you’ll likely find AI spread across several tools—not one.
Gena Graziano, integrator and chief of staff at First Steps Financial, runs a paid stack that includes Gemini, Claude, ChatGPT, and Fireflies, plus the AI built into their existing software.
Gena Graziano
Her firm pays for its tools very deliberately. “We are a Google company, so Gemini is our paid version,” she says, and they have a firmwide rule behind it. “No one’s allowed to use free versions. No one’s allowed to use personal versions.”
Patti Storms, who supervises the client accounting team at BKC, agrees that no single tool will do.
Patti Storms
She also knows she can’t find every option on her own, so she networks. “It’s almost impossible for me to believe that any one person or one firm can research every product out there,” Storms says. She advises smaller firms to lean on professional networks, such as state CPA associations, and to look to people who’ve made it their business to find the best, most secure tools. “I usually find out something that I’ve never heard of,” she says.
Buying what the market already does well means someone else maintains and improves it, so your team can focus on getting the most out of it. And since your competitive edge is unlikely to come from rebuilding a tool you can already buy, you can save the building for what’s peculiar to your firm.
Build: When a Custom AI Tool Is Worth It
The strongest case for building your own AI tools isn’t just that you can, but that your firm has a problem so particular to the way it works that nothing on the market solves it well.
Bri Como, operations and systems manager at VP CPAs, has built two custom AI agents inside her firm’s workspace software. The first lets staff log billable time by messaging a bot. The second sends her a daily digest of everything still outstanding.
The time-tracking agent came from a simple problem.
Bri Como
A staffer now messages the agent something like “track forty-five minutes on this task, and here are the notes for the time entry.” The agent finds the task, or offers two likely matches and asks which one is right, then logs the time.
The digest agent needed tuning at first. Como had to tell it to ignore anything she’d tagged “NTS,” her shorthand for notes to herself. Neither agent is a product you could buy. They are shaped to how her firm works, and each sits on software the firm already pays for.
Jean Zick, who runs Juna Financial Solutions, an 18-person remote firm serving life science, biotech, and SaaS clients, builds for a different reason. Her biotech clients budget by department, which gets very granular, and she couldn’t find many products that handle it.
Jean Zick
The firm is building a budgeting and forecasting tool. Earlier this year, they also built an automation for one client and turned it into a service. Zick’s team debated selling the IP, but she pushed back. “I don’t want to sell IP,” she remembers thinking. “What are we really selling here?”
Pricing it was tricky, so she asked colleagues in the industry who had done something similar. She told the client the firm had built it, so that the client would now take over the work, and that the price was a fraction of what it would cost to hire someone to do it by hand. A custom build can become more than an internal efficiency tool when it solves a valuable problem for a client.
Taylor Ammons shows what building looks like at the earliest stage, along with its risks. Ammons is the integrator at Accounting Therapy, a remote firm of eight. She has been using vibe-coding tools to turn a clunky Excel capacity planner into something more automated, built on a database.
Her next idea is a plug-and-play app. A staffer would drag in a fixed asset schedule or a raw data export and get back a file ready to upload into QuickBooks. She hasn’t built it yet, and her team isn’t using this kind of automation on client work.
Before she does, she’s counting the cost.
Taylor Ammons
Deborah Harris’s reason for building goes beyond efficiency: value. Harris is also an angel investor, and she has sat through pitch after pitch built around AI. Early on, she says, they all sounded the same. “AI, we do this, we do that, and then you went, ‘But you don’t actually own the IP.’”
Grow Group’s answer is a context engine Harris has spent four years refining. It’s a body of knowledge that captures how the firm works and thinks about clients, and her team uses it to give AI systems a deeper understanding of the business.
Deborah Harris
She expects the old rules of thumb for valuing service firms to change.
Each of these builds is tied to something particular about the firm: how it bills, who it serves, what it wants to be worth. Build where a custom tool becomes a lasting, firm-specific advantage, and only after you’ve priced the upkeep.
Brush Aside: When Not to Use AI
Jan Haugo builds for a living. An AI implementation expert, she has built tools that tie a year of payroll to the general ledger in under 10 minutes. So when she asks whether something should be built, the question carries weight.
Say her AI can handle accounts payable, but she already has Bill.com doing the job; does she replace it?
Jan Haugo
That last question is the heart of brushing something aside. Whether AI can do the job isn’t the test, because it often can. The test is who maintains the tool after it exists. Every custom build can become a standing obligation: fixes, updates, and a person who still has to understand how it works.
Davis has also built plenty with AI, and he’s candid about what happens when firms ignore the math. “The first version is always the easiest,” says Davis. He warns that the cost comes later, if a build doesn’t work well for the firm.
Chad Davis
Davis’s own 2019 app taught him the scale version of the lesson. “You just can’t try to replace 30 things,” he says. It didn’t turn him against building. His advice is to build small, off to the side, and leave what works alone.
Davis sees room for small automations in the gaps between the systems. “Your project management system’s not going to deal with something your billing system’s not going to deal with,” he says. “Not everything has to be a big, large sweeping change in your firm.”
The fact that you can build something doesn’t mean it’s reasonable to take on. Sometimes the smarter choice is to let a specialist product keep doing its job.
Every tool and system you leave alone gives your team one less change to absorb. The systems that already work can keep doing their jobs, while the custom work fills the gaps between them.
The Build-or-Buy Decision Tree
Run each need through these questions in order. The first one that stops you is your answer.
- Is the need real? If you did nothing, would anything actually break? Or is it a process you could simplify or drop?
- No: Brush aside. Skip it, or fix the process instead.
- Yes: Go to question 2.
- Is a working tool already doing the job?
- Yes: Brush aside. Leave it exactly as it is.
- No: Go to question 3.
- Does something you can buy handle it well?
- Yes: Buy. Ask your providers what’s already switched on before you add a new subscription.
- No: Go to question 4.
- Is the need particular to how your firm works?
- No: Brush aside for now. If no product fits and the need isn’t specific to you, it may not be worth building. Revisit when the market catches up.
- Yes: Go to question 5.
- Can you name who will maintain it?
- No: Brush aside, or buy the closest option.
- Yes: Build, with three rules: automate the first piece, run it alongside the people doing the work, and measure performance.
The Maintenance and Authority Test
The decision to build, buy, or brush aside doesn’t end when a tool starts working. Three questions need answering: Is the gain obvious? Who will maintain it? And how much authority should it have?
Davis is most interested in the authority. He warns that small firms are prone to one mistake.
Chad Davis
To define what the 20% mark looks like, Davis advises firms to pull the data out of their practice management software and run the tool in parallel with the people doing the work to “see where the edge cases are, see how it performs.” Only then “slowly give it authority to make changes in the tools.”
He believes that firms should keep access and authority separate. Giving AI access to a system “doesn’t mean that we should give it the authority to change the stuff in [it],” Davis says.
Then there’s upkeep, and Ammons has already weighed it. Even when an app like her QuickBooks import tool works, she says, “you still have to validate that information and make sure it’s right.” And when the software it depends on changes, someone has to rebuild it. “So there’s still somebody doing work somewhere,” she says.
The payoff has to justify both, and Haugo’s ledger-review tool shows what a clear gain looks like: a year of payroll tied to the general ledger in under 10 minutes, and a full ledger reviewed in about 15.
When the gain is that visible, it’s easier to weigh against the upkeep. When it isn’t, she tells firms to look for solutions “that are going to move the needle for you.”
So before anything joins your stack, ask three things: Is the gain obvious? Who maintains it? How much authority should it have?
Together, these questions keep the build, buy, or brush aside decision more practical. The long-term value depends on whether your firm can maintain a tool and trust it.
Access ≠ Authority
Access means the AI can see your data. Authority means it can change it.
“Just because we can give the AI access to [a system], it doesn’t mean that we should give it the authority to change the stuff in [it].”
— Chad Davis, Co-Founder of LiveCA
How Davis says to earn authority:
- Don’t automate the whole workflow. Stop at the 20% mark and pull the data.
- Run the tool in parallel with the people doing the work for a while.
- See where the edge cases are and how it performs.
- Then slowly give it authority to make changes.
Assemble a Stack You Can Maintain
A well-designed AI stack doesn’t rest on one or a pile of custom apps. It’s layered: a general-purpose assistant for flexible work, the AI capabilities already built into your core software, and a few custom solutions for problems that genuinely call for them. Each layer has a purpose, and each addition should earn its place.
Davis offers one more habit worth imbibing. At the end of planning any build, he says, add one sentence: “How else could we do this using different technology? You might be like super surprised at the path that it shows you,” he says. It’s a cheap check against falling for the solution you’ve already started to build.
Maybe an existing feature can do the job. Maybe a specialist product handles it better. Or maybe the process doesn’t need another tool at all.
The goal is to assemble the smallest stack that solves your problems, one your firm has the time, resources, and capacity to maintain.
Enter your email address to receive new issues of the magazine straight to your inbox.