Use AI at the lowest level that does the job. For most small businesses that means Level 2, where an AI model works from your own price list, intake forms, and notes, and you stay the one who sends, books, and signs. There are four levels in all, running from asking questions to letting software act for you, and climbing isn't the point. Level 4 suits one specific, repeatable job, and an owner who never gets there has not fallen behind.
Most AI coverage is written by people selling the deep end, so the plain uses get treated as a warm-up. A dental office that writes better recall letters at Level 2 is using AI correctly and completely.
How should a small business use AI?
Pick a job first and the level second. "Use AI" is a direction, and directions don't get done. "Draft replies to this week's Google reviews in our voice" is a job, and it tells you right away how much the model needs to know and whether it ever needs to touch anything of yours.
Then ask the cheapest question that fits. Does the model need your material, or only a good question? Must it read from a system you use, or can you paste the relevant part in? Does it need to act, or is a draft enough? Each "yes" moves you up a level, and each level asks more of you in setup, in attention, and in risk. If the answer to all three is no, you're at Level 1 and that's fine.
What are the levels of AI use?
Level 1 is asking. You open a chat window and use it like a much smarter search box. Nothing of yours goes in beyond the question. A plumber asks what a fair rate for a water-heater swap looks like. It costs nothing and risks little, since nothing private left the building. Its weakness is that the model knows nothing about you, so it answers for the average business, and the answers feel generic because they are. How a search habit turns into something more is the subject of beyond search, from prompts to projects to folders.
Level 2 is working with your material. You give the model your documents and it works from them. ChatGPT and Claude both offer projects, a saved workspace where files and standing instructions carry over to every chat inside it, so you explain your business once instead of every morning. OpenAI describes the feature in its ChatGPT projects help article. If you run a small accounting practice, you load your engagement letter template, your fee schedule, and a page of notes on how you like client emails worded, then draft a renewal in your own voice in the time it used to take to find last year's. Nearly every business finds most of its value here, and what your internal data can tell an AI model goes through which records are worth loading.
Level 3 is connected. The model reads from your systems directly through a connector, a prebuilt link between the AI tool and something like your calendar, inbox, QuickBooks, or point-of-sale system. A retail shop's owner asks which items sold through fastest last October and what's sitting in the back room now, and gets an answer from the actual sales data instead of from memory. At this level the model reads, and the acting stays with you. The plumbing has its own post, what MCP is, in plain English, and the reason connected data beats separate tools is in what data integration does for a small business.
Level 4 is delegated. The model acts for you. It sends the email, books the appointment, files the invoice, or changes the record, sometimes with your approval on each step and sometimes without. Software that does this is usually called an AI agent, and the rules for how much to let it touch are in how much access an AI tool should get.
| Level | What it looks like | What it needs from you | Where it goes wrong |
|---|---|---|---|
| 1, ask | A chat window, one question at a time | A clear question | Generic answers, because it knows nothing about you |
| 2, your material | A project holding your files and instructions | An afternoon of writing down how you work | Stale files, so it confidently repeats last year's prices |
| 3, connected | The model reads your calendar, books, or sales | Logins, and a decision about which systems it may see | It sees more than the job needs, including things you'd rather it didn't |
| 4, delegated | An agent that sends, books, or changes records | A tightly defined job, approvals, and a way to undo | One wrong action repeated fast, in your name |
Which level fits my business?
The one where your bottleneck lives. If the frustration is that the answers are generic, you need Level 2, because a connected model with no instructions is just a generic model with a view of your calendar. If the frustration is that you keep copying numbers out of your books and pasting them into a chat, that's a Level 3 problem. Approving the same correct action dozens of times a week, with no variation, means you may have found a Level 4 job, and AI automation for small business covers which jobs earn it.
Two things shift the answer regardless of the bottleneck. The first is what you hold. A dental practice with patient records, or a law office with client files, has duties about where that information goes, and that can change which tool you use at any level before it changes how high you climb. Whether you need private AI is the place to start, and the regulated-office rules are in AI for HIPAA and other regulated offices.
The second is which tool you pay for. The chat window you see is one program wrapped around the model, and the program around the model decides much of what you can actually do at every level. Which model sits inside matters less than people expect, though frontier and open-source models do differ.
My own week runs on more than one level at once, on purpose. Email to people I still write at Level 2, with my own notes loaded and my hand on the send button, and I look things up at Level 1 a dozen times a day. Nobody would be served by automating that email, because a wrong email to a person costs more than the minutes it saves.
What climbing costs
Every level up adds a cost that doesn't show on the invoice. Level 2 adds upkeep, since files describing your business go stale the day your prices change, and a model working from a stale file is wrong with total confidence. The third level adds exposure, because you're deciding what a piece of software may look at, and most tools default to seeing more than the job needs. Level 4 adds supervision. Someone has to check what the agent did and know how to switch it off, and in a five-person business that someone is usually the owner.
Climb only when the job pays for those costs. The owner who stays at Level 2 for a year and gets very good at it, with clean files, sharp instructions, and a habit of fixing the file whenever the model gets something wrong twice, is probably ahead of the one who wired up four connectors in a weekend and now isn't sure what any of them can see.
One place AI touches your business sits outside the levels entirely. Customers now ask AI models which plumber or dentist to pick, and AEO vs. SEO explains what changes when a model writes the answer.
A second set of eyes on which jobs belong at which level is what our services are for, and the rest of the series sits on the AI in plain English hub.
What I haven't settled is where Level 2 ends for a business that grows. A shop with one owner and one project folder is simple. A practice with six people each keeping their own project, each with slightly different notes about the same fee schedule, has a new problem, and I don't think anyone has a clean answer to it yet.