Is AI Going to Replace Accountants?
Unless you’ve been living under a rock, have fully integrated with the Amish, or have been lost at sea for the last four years, AI has been a part of life. In business, having a discussion about AI and its related development is, for better or worse, unavoidable.
You can’t open YouTube or Instagram without seeing an ad or video that has been fully generated by AI. You can no longer search for a “bookkeeper near me” on Google without first getting a Google AI Overview. There is discussion around a post-scarcity economy, the elimination of white-collar work, and, with the development of robotics, the elimination of blue-collar work as well. And, according to some of the latest headlines, Anthropic and OpenAI have even acknowledged that their products carry a non-zero chance of killing us all.
Needless to say, we’re here to focus on the “elimination of white-collar work” bit. Because as a CPA, I’ve been told more than a few times over the last 10 years that my profession is going to be automated. And believe me, there are times I wish it would be!
ChatGPT, Claude, and the other AI models you use generally run on an LLM, or large language model. These models are trained on enormous amounts of data and generate outputs based on the information contained in your prompt and the patterns they learned during training.
The neat thing about them is that they are, for lack of a better term, extremely sophisticated autocomplete engines that are similar in concept to what exists on your iPhone or Android when you’re texting.
Ask Claude a simple question like, “Guess a number between 1 and 30,” and it might output “17.” Ask it why it picked 17, and it might tell you that when someone is asked to choose a number between 1 and 30, they tend to gravitate toward numbers they perceive as random, such as 7 or 17. That, Claude explains, is the reason 17 was the output. Because 17 is a statistically likely response, far more so than 1, 12, or 22.
What these models are fundamentally doing is generating the response they calculate to be the most likely given the prompt and context. Even when the model provides an explanation for its answer, that explanation is itself generated by the model.
Now we get to the point of this whole article:
Is AI going to replace accountants?
Because of what I just explained, the answer, I’m afraid, is no.
LLMs are probabilistic models. Accounting, for the most part, is a deterministic field. We operate in a world where debits and credits have to balance, transactions have to be classified correctly, and financial statements have to tie out. There is certainly judgment involved in some instances. For example, technical accounting can be a very different story. But the underlying mechanics of accounting are largely black and white.
I’ve brought this up to a few folks. Their response is usually something along the lines of: “Well, the chat models aren’t going to replace accountants. The agents are!” But agents aren’t any less probabilistic.
An AI agent is essentially an LLM wrapped in a system that allows it to take actions, use tools, and make decisions across multiple steps. If I ask an agent to open Chrome, find me the best deal on a product, and then buy that product, what is happening behind the scenes is that the model is determining what action to take at each step and scripting it out: opening a new tab, navigating to Google, searching for the product, comparing results, clicking a link, and so on.
Here is a real life example:
A few months ago, I had what I thought was a big breakthrough with an AI agent. I automated a task. I provided detailed instructions, gave it examples, and built out skills for a GL cleanup process that had been taking me hours every month.
Important Disclosure: this involved no client data.
The agent executed the task flawlessly in month one. Month two? It took twice as long, skipped several steps, and, in its own words, told me it felt those steps were unnecessary. After enough back-and-forth, I was at an impasse and had to do the work manually.
It felt like my big AI breakthrough had been taken away from me. And by the time I finished dealing with the agent, I had spent more time cleaning up its work than I would have spent doing the task myself in the first place. That experience really spelled out for me the hurdle we’re going to face with AI taking everyone’s jobs, curing cancer, and ushering in a post-scarcity utopia (or poisoning the water and killing us all).
AI is too inconsistent.
If something can give you one result in one month, and a different one in another, it cannot be relied upon for a deterministic process. In fact, if Claude or ChatGPT were someone I employed, I would need to have some tough conversations around performance.
For the foreseeable future, we are going to need a human in the loop. And the same set of circumstances that prevented us from fully automating accounting with traditional software also creates problems when we try to automate it with AI: system incongruity, messy data, exceptions, and inconsistency.
Extrapolate that outward.
A GL cleanup exercise is relatively harmless. If the AI makes a mistake, I catch it, fix it, and move on. The inconsistency in the AI application is what you could reasonably call a bug.
But what happens when the bug is in your financial reporting?
What happens when it is in your tax return preparation?
What happens when it is in your board reporting or investor relations communications?
What happens when the AI confidently tells you that a number is correct because it looks correct? What if the analysis it provides is completely made up, but sounds compelling?
You’re not going to get away with telling an auditor or an investor that your AI messed up your financials. You are definitely not going to get away with telling the IRS it was Claude’s fault your tax return was wrong.
If you’re considering using an AI solution for your bookkeeping or tax preparation
Please be careful, as these solutions are currently half-baked and can get you in trouble without a professional reviewing the output of these programs. If you are finding success with these, passing audits, and your tax professional isn’t billing you extra for financial cleanup, more power to you! If you are finding yourself needing help, we offer book cleanup and catchup services.
What this boils down to is:
CPAs are here for the foreseeable future. AI, LLMs, and agentic tools are incredibly useful, but can land you in hot water if you rely on them entirely.
If you are interested in learning how we can help you:
Reustle Accounting and Finance is a CPA firm based in Garnet Valley, PA that offers bookkeeping and accounting advisory for small businesses and startups in the Greater Philadelphia area. We have years of experience across a variety of industries, and apply our knowledge to your business’s financial situation to do anything from monthly bookkeeping to building out a financial close process that can scale with your organization.