The AI Terms Vendors Hope You Don't Understand

So you keep hearing the terms LLM, API, MCP, AI agent, and maybe you're not alone, but secretly you don't know exactly what those things mean. My hope is that that ends today. Hey, my name is Isaac from Opsa.co, where we help accounting firms support more revenue without increasing their headcount or adding expensive software. And in this video, I'm going to break down 20 AI and automation terms that every accountant and bookkeeper needs to know in 2026. I'll attempt to explain them in as simple terms as possible and hoping to use some accounting knowledge or analogies so everything clicks. Let's start with the first one, token. What is a token? A token is like a billable unit. It's technically the smallest chunk of text an AI can process. It's estimated to be a third, three-quarters of a word. And just like you as a firm, an accounting firm, might bill a payroll service per employee or an AP service per bill, or perhaps there are services that you bill per hour, right? That is the billable unit you're using. In the same way, AI uses and measures the number of tokens as its billing unit. Just to give you some examples, the word tax might be a token, accounting might be 2 tokens. If you have a 200-page book, you might be talking about 100,000 tokens. If you're talking about the tax code, you might be talking about 5 million tokens. Both the inputs, what you tell the AI, and the output, what it tells you, consume tokens. So the longer your response and the longer its response, the faster you hit your limits on your AI models, and you might need to pay for extra usage. Second word, LLM. This is really the AI engine that's reading and generating the text. Think of it as a smart accountant or a smart team member who's read and memorized a lot of textbooks, the tax code, the client file, and now you can effectively ask it for the knowledge that you want. The LLM, or the large language model, AI companies launch over time, right? You have GPT-5, Claude, Opus 4.6, Gemini 3 Pro. Like, they typically have different LLMs within their tools. And again, think about it as different accountants, different people that may have different strengths, different weaknesses. For example, most of the time they have models that are trained to provide a quick and fast response. Other times you might have models that are meant to think deeper and give you a more thorough response. Those typically consume more tokens. So we can start to discuss the different terms. Term 3, hallucination. This is where AI confidently tells you something that it completely made up. And it's one of the biggest risks of AI because it's sounding confident when it's completely wrong. Some common examples might be people creating a document and all of a sudden it's assigning citing quotes for research that doesn't exist or is referencing a number that is actually not there. Hallucination is becoming less of an issue, and it's gone from like 20% back in GPT-3 days to like 3% in GPT-4. And at least based on the last features back in later 2025, hallucination rates are somewhere around the 1% mark at this point, and we expect them to continue to go down. Turn number 4, temperature. Temperature is a setting that controls how creative or predictable the AI output is. The lower the temperature, the more safe, the more it'll go by the book. The higher the temperature, the more creative, the more it will take risks. You can think of it as a tax strategy, how conservative or aggressive it might be. It's a little bit of an spectrum. Similarly, the temperature's typically like 0 to 1, and you might say, hey, 0, 0.1, 0.2, that might be very predictable and consistent if you wanna extract data or categorize things. If you wanted to create poetry and, and brainstorm marketing copy, then you might actually, you might want a much higher temperature closer to that 1 that we talked about. Term number 5, context window. Context window is the maximum amount of text or context that an AI can see at once in a conversation. It's effectively just like human memory. You cannot really hold a lot of information at once. You have a limit. Similarly, AI models have a context window limit. For example, Claude offers like 200,000 tokens. That might be like 500 pages of text. GPT, depending on the model, might be between 8,000 and 128,000 tokens. The idea is that once you exceed that window, the AI is going to forget about that initial thing of the conversation. You may have to repeat it. Now, and the reason it's not an infinite context window, because the more it has to read, the more it's going to cost and the less quickly it's going to give you back a response. So it has to limit itself to some extent. Term number 6, RAG, or Retrieval Augmented Generation. I like to see this as an AI's ability to generate a table of contents on the fly. So imagine if you have like a 10,000-page library and you're asked a question, you're not going to read every book and every word. You're going to look for the most appropriate or relevant book, and you're going to read a table of contents to find the right chapter and go straight into the most relevant pages. That's exactly what RAG is. When the context is too big and you ask AI a question, it might try to scan the documents and kind of build its own table of contents, see where the most relevant sections are to pull that into the context window so that now it can answer Your question correctly. Term number 7, prompt engineering. This is really the skill of writing instructions that will get you the best output from AI. I do believe that this is essentially good old communication. I'm convinced that the term prompt engineering won't be a thing in a few years, but it's the difference between you just telling it, hey, just help me with my taxes, versus you asking it, hey, what should be the depreciation schedule for these, you know, Tesla Model 3 2023 that's used half of the time for business in the state of Florida, right? The more specific you Or just like you would have to be for a, for an intern or for a team member, the more you provide examples, the more you break down the task into steps, then typically the better the prompt and the better you're communicating with the AI to avoid a lot of back and forth. Turn number 8, system prompt. These are hidden instructions that tell the AI how to behave before you even tell it anything. So think about it as your firm's SOPs, right? You will perhaps tell a system prompt, you are an AI assistant for an accounting firm. Always ask for the client's entity type before giving tax advice and never give a specific dollar amount without a disclaimer. If you have that, for example, as we'll talk later in a project or somewhere that is like beneath the kind of the prompt that of the thing you're working with, the more consistently AI behave regardless of who uses it. So it's really just a set of instructions that's loaded behind the scenes before the conversation starts. And just like you could have a prompt for your firm, sometimes companies themselves like Claude or OpenAI may have systems prompts. Behind their models that we don't see. Term number 9, projects. This is in the context of Claude and ChatGPT. It's really more like a workspace where you have a custom set of instructions and some files that are going to persist and be referenced across conversations. Really, the best analogy is literally a folder in the physical world where you might have specific files there, and perhaps you have a specific checklist or process on how to handle the things that are within that particular folder. So the AI is gonna remember the instructions, is going to— if you set it up that way only, you can set it up so it only remembers its isolated context. It doesn't reference anything outside of it. And it's really the difference between you asking a random staff member on the regular chat that perhaps doesn't have that unique context and someone's dealing with a client for 3 years, aka the project. It tends to perform better. Term number 10, embedding. Embedding is turning text into numerical codes so that AI can measure how similar 2 things are. And the analogy I like to propose is imagine assigning codes to a general ledger. You might say, hey, office supplies, it's 6,000, computer equipment, it's 6,050, right? They're close together, they're related because they're both operating expenses, and you might kind of put them closely together. This is actually what behind the scenes allows RAG, the Retrieved Augmented Generation search that we talked about possible. Because now if the app's converting text into numbers, it can mathematically know how similar they are, even if they don't actually have similar words. So if you say accounts receivable and outstanding invoices, those might have very similar embeddings under the hood. So AI knows they are related, but the words are totally different. You will rarely interact with embeddings directly, but it's worth knowing that they power a lot of the AI search features that you see. Term number 11, fine-tuning. This is really retraining an AI model on specific data so it changes its default behavior. This tends to be expensive and complex. Most small firms won't need it, but this is the difference between effectively just giving AI a better instruction and more like rewiring its brain and giving it more examples and data so that its internal knowings are fundamentally different. Term number 12, API or Application Programming Interface. An API is really a structured way for 2 systems to exchange data. I like to see it as a login page for a robot, just like you have a login page for a human being where you first give credentials and now you have access to do stuff like you can create an invoice, create a journal entry, create a customer, create a vendor, edit a vendor, create a rule, delete a rule. It's really how you allow a software to talk to another software. So when QuickBooks is pulling bank transactions from a bank feed, it's making an API call. Every app that's using kind of has an API working behind the scenes, whether it's Stripe, Gmail, Slack, QuickBooks, Xero. You don't need to know how to create those APIs yourself, but it's worth knowing that every time an app is interacting or pulling data from another, the existence of an API between them is what allows them to connect. Term number 13, webhook. Webhook, it's an automatic notification sent from one app to another when something happens. So think about it as just like perhaps you signing up for notifications or your bank when a transaction is successful, right? You don't have to ask. Every time that happens, it tells you, it sends you a message. In a very similar way, a webhook is like a message you're able to send to another app or somewhere else in the web. It's similar to API, but a little different, right? API, it's I'm able to call and see the bank at some point in time when I want. A webhook is more like when something occurs, For a specific event, I'm gonna notify you at that point. Some common examples might be, hey, when this work item in our practice management tool is completed or it's changed the status, I want you to tell our automation platform to kick off these series of things that perhaps I wanna notify a team member or want to assign it to somebody else. Term number 14, connector. Connector is a prebuilt integration between 2 specific apps. So connectors can be between an app and an AI model, right? You have connectors within ChatGPT, connectors within Claude, but you also have connectors connections between apps and automation platforms like Zapier, Make, or N8N. It means someone already wrote the code for you so you can go in and just click connect. And in a way that's native and prebuilt, you'll be able to connect, say, Gmail to Claude or Outlook to Make. You'll be limited to what the connector supports and wouldn't be able to customize beyond that, but it allows you to piggyback off of what someone's already built. Turn number 15, NCP, or Model Context Protocol. This is like GAAP but for AI. It's essentially a standard that lets AI plug into any external tool or data source. Just like GAAP, someone decided to say, hey, let's sit down and let's have generally accepted accounting principles so we have a shared set of rules that anyone reading a company anywhere will be able to, you know, follow through. Now NCP does the same and gives every AI a shared standard on how to plug into any tool in the same Way. It's a relatively new concept that was sort of launched or popularized in late 2024, and it's gained a lot of traction in recent months. Turn number 16: SAPs, scenarios, flows, workflows. The term might vary, but what do you call a specific automation or automated procedure in a playbook or in a tool? This is essentially what represents that single unit of an automated workflow, and depending on the tool you're using. It will be a different thing, but it's essentially a procedure saying, hey, when the engagement letter is signed, I want you to create a folder. I want you to send the organizer. I want you to update our tracker. I want you to notify the client. In Zapier, they're called Zaps. In Make, they're called scenarios. In Power Automate, they're called flows. In N8N, they're called workflows, but they all represent the same thing. And the good thing is you can have as many of them as you want. You might be running dozens of these workflows at once, each handling a different procedure, just like you would have a human being. Knowing that, hey, when this happens, you have to perform these series of steps. Term number 17, no-code or low-code. This is really a term that really means visual development. What that means is that you're building automations or building digital solutions by dragging and dropping, by filling forms and simpler visual elements instead of writing code. When they say no-code, it's because it's totally drag and drop, perhaps like Zapier or Make. For example, although you technically can add some code blocks, low-code is like mostly visual where you're dragging and dropping, but you can add code for more advanced logic. This is really what has allowed a lot of automation to become a lot more accessible for accountants and bookkeepers, cuz now they don't need to be a developer to give that a shot. Although a developer could still assist them if they wished. Term number 18, trigger. This is the event that kicks off an automation. A common example is a client accepting your proposal. That's the signal that now you have to start the onboarding process. And just like that, an automation needs to have a trigger to assess at what point do you want it to do what it's supposed to do, right? And that trigger will kick off that entire chain of actions. And that trigger can be either event-based, right? When something occurred, we talked about webhooks, that could be its trigger. And we can have it scheduled as well, running every Monday at 8:00 AM or on a regular cadence. Turn number 19, action. Okay, this is essentially after the trigger occurs, The task that automation will perform on its own automatically and without human involvement. A single trigger can have 1 action or 100 actions in sequence. And some examples might be, we might be creating a folder, sending an email, updating a spreadsheet, posting a Slack message somewhere. Each action usually is gonna be connected to a different app or to the same app via the connectors we talked about earlier. So ultimately with a combination of triggers and actions, you can create turn number 20, a workflow automation, right? A predefined sequence of steps That are gonna run the same way every time given a specific trigger. I like to think of workflow automation specifically without any AI components like formulas in an Excel, right? You're saying if this cell says space, then I want you to perform these calculations. It's gonna do exactly what you told it to do and nothing more than that. If something unexpected happens that you didn't account for, it may error out like some formulas sometimes do. So it is fast, it is reliable, it is observable, and you know what will happen, but it can be a little dumb because it doesn't have AI in it. So it'll, it will struggle to handle exceptions that you have not anticipated or programmed. That said, it's still where most firms should start to automate the repetitive things that don't require a lot of judgment because this tends to be more reliable and more cost effective over the long term than its AI-related equivalents, which I want to talk about next. Term 21, agentic workflow. This is like an automation, a workflow automation with some AI steps, right? It's a predefined workflow. That has AI embedded in a specific step or series of steps to handle some judgment call. Think about the spreadsheet formula example I gave you where now you might have a human being looking at when this thing errored out or something that perhaps we couldn't really put in with a formula, like flagging and something doesn't look right. You have now that human being, that intelligence making that determination, making that decision. So that's really an agentic workflow where at key steps you're adding intelligence. And an example might be, hey, a receipt came in automatically, but now I wanna categorize it. That perhaps requires a little bit more intelligence and I want to like flag it for anything unusual that might need a little bit of intelligence, but perhaps I know that it always needs to be posted into this account. So that can be a formula or that can be an automation. And term 22, AI agent, an AI that can use tools, make decisions, and adapt when things don't go as planned. And that's the key distinction between a workflow automation, an agentic workflow, and an AI agent. It's that now you are giving it tools. You're telling it, hey, you have access to my inbox, you have access to my project management tool, you have access to our SOPs. I'm gonna give you a goal and you'll figure it out. That's really it. It's not gonna have a predefined path that you set for him. And an example might be, what is our most senior accountant working on this week? It may not know who your senior accountant is. It might not know how it can know what it's working on, but it might at that point start to check, hey, let me see what tools I have available. Oh, look at that. I have access to a project management tool. Let me see if I can go in and list the tasks that people have, are working on this week. Well, I don't know who is the senior accountant. How can I know that? Let me see if I can pull the members and see if they have their title. Perhaps that's a way I can know. And you can start to think about and figure it out, try and fail through the process of answering the question or achieving the goal you gave it. It reasons through to try to attempt to do that on its own without a predefined path. So now you speak the language, but knowing the terms and actually building with them are 2 entirely different things. So in this next video, I wanna walk you through 10 actual AI automations we built for accounting firms using some of these concepts so you can see in practice how they work. Check that out next, and until then, stay efficient.