What a Large Language Model Actually Is, for Law and Accounting Firms
By Stacey Tallitsch | July 7, 2026
If you run a law firm or an accounting practice, you have heard the term "large language model" by now. Maybe a vendor used it in a pitch. Maybe your kid said it at dinner. Maybe it showed up in a bar association email or an article about the future of billing. You nodded along. You still are not sure what it means.
That is fine. Nobody handed you a clear definition. This post gives you one.
What the words actually mean
Let me take the phrase apart, one piece at a time.
"Model" here does not mean a fashion model or a scale model. In technology, a model is a system that takes something in and gives something back. Think of it like a very experienced paralegal. You hand over a request. They hand back a draft. You do not see every step they took in their head. You just see the result.
"Language" means the model works with words. Not numbers in a spreadsheet. Not photos, at least not in the basic version. Words. Questions, sentences, documents, emails.
"Large" means it was built by reading an enormous amount of text. Not a filing cabinet's worth. Closer to a big slice of everything ever written on the public internet, plus a lot of books.
Put it together. A large language model is a system that read a staggering amount of writing and can now produce writing of its own in response to what you ask. People shorten it to "LLM." When someone says "the model," or names a tool like ChatGPT or Claude, they are usually talking about an LLM.
How it learned, in plain terms
Here is the part that surprises most people. Nobody sat down and programmed the model with rules like "in a demand letter, put the date at the top." That is not how it works.
Instead, the model was shown mountains of text and asked to do one simple thing over and over: guess the next word. Read a sentence, cover up the last word, try to fill it in. Do that trillions of times, and the system slowly gets good at it.
The text it learned from is called "training data." That just means the writing it studied during its education. If you want the longer version of how that works, and whether your own typing gets added to it, I wrote about what training data actually is in an earlier post.
There is one more odd fact worth knowing. When you type a question, the model does not read your words the way you do. It turns them into long lists of numbers, does its math, then turns the numbers back into words. Anthropic, one of the companies that builds these systems, describes this plainly in its own research on how a language model handles what you type. You never see the numbers. You just see the answer.
An example from your world
Say you run a small accounting firm. A client emails asking why their quarterly estimated tax went up.
You paste the question into an LLM and ask it to draft a reply in plain English. A few seconds later, you have a solid first draft. It explains the increase, keeps a calm tone, and invites a call.
The model did not look up your client's actual return. It does not know your client. It wrote a general, sensible explanation of why estimated taxes rise, in the style you asked for. You still have to check it against the real numbers. But you started from a draft instead of a blank screen.
That is the honest shape of what an LLM does. It is a fast, tireless first-draft writer that has read more than any person could. It is not a lawyer. It is not a CPA. It is not your firm.
What it is good at, and what it is not
LLMs are strong at a specific set of tasks. Drafting and rewriting text. Summarizing a long document into a short one. Explaining a complicated idea in simpler words. Turning rough bullet points into a clean paragraph. Answering general questions about how something usually works.
They are weak, or flat-out unreliable, at other things. They do not know today's date unless told. They do not know your client files unless you hand them over. And they will sometimes state a wrong fact with total confidence, because their job is to produce plausible writing, not to check truth. That failure has a name, "hallucination," and it matters enough that I gave it its own post.
For a firm that bills on accuracy, this is the line to hold. Use the model for the draft. Keep a human on the facts. Never file, send, or advise off the raw output without a review.
A quick word on privacy
Because your work is confidential, one question comes up fast. If I paste a client matter into one of these tools, where does it go?
The honest answer depends on the tool. Free consumer versions may use what you type to help train future models. Paid business versions usually promise they will not. The setting exists, but you have to check it and turn it on, the same way you would check whether a new filing system is encrypted before you trust it with client data.
The safe habit for a small firm is simple. Do not paste names, account numbers, or anything that identifies a client into a tool you have not vetted. Strip the identifying details, or use a business version with a written promise on data handling. You already do this instinct check with paper. Do it with these tools too.
Why professional firms hear about this constantly
Law and accounting are word businesses. Engagement letters. Memos. Client emails. Summaries of long contracts or long returns. That is most of the day.
An LLM is a word machine. So the overlap is real, which is why every vendor selling to your profession now leads with it. The pitch is not made up. The catch is that the pitch usually oversells how much you can hand off.
The tools that actually help a small firm are narrower than the pitch. An LLM built into your email to draft replies. A tool that summarizes a 40-page document into a single page. A drafting assistant for routine letters. These are useful and boring, which is the right kind of useful.
You may also hear the word "agent," as in "AI agent," which is a step beyond a plain model. If you want that difference spelled out, I covered what an AI agent is and how it differs from a basic chatbot separately.
Does this matter to your firm right now
Here is the plain answer. You do not need to understand the math inside an LLM any more than you need to understand how your phone's chip works to make a call.
What is worth knowing is the shape of the thing. It reads and writes. It drafts fast. It does not know your specific facts unless you give them to it. And it can be confidently wrong. Hold those few ideas and you can size up almost any AI pitch that lands in your inbox.
You do not need to rush. You do need to stop nodding along to a term nobody ever defined for you. Now you have the definition. Use it to ask better questions the next time someone tries to sell you the future.
-- Stacey | The Standalone
About the Author
Stacey Tallitsch runs The Standalone, an AI Implementation Diagnostic practice for small business owners. He has 30 years of experience in technology and has written 21 books on systems thinking and decision-making. More than 30,000 students have learned from his online courses.
- Stacey Tallitsch, The Standalone