What to Actually Use AI For in a Small Business

Most founders I talk to are using AI as a faster Google.

That’s not a criticism, it’s the data. In a study of just over five thousand US knowledge workers, around six in ten uses were AI as a replacement for a search engine. Only about one in six had used an agentic tool at work at all. So the typical experience is: ask it a question, get a decent answer, feel mildly impressed, close the tab. Nothing in the business changes.

The instinct after that is to go looking for a better tool. It’s almost never the tool.

IBM surveyed two thousand CEOs and found that 86 percent of employees either have the skills to use AI or could pick them up with light training. 25 percent actually do. That’s a sixty-one point gap between what your people can do and what they do. You don’t close a gap like that with a better subscription. You close it by connecting a specific capability to a specific job somebody actually has on a Tuesday afternoon.

So this isn’t a list of tools. It’s how to tell which jobs to hand over.

The rule that sorts almost everything

Here’s the one I use, and it takes about five seconds to apply.

If the task has one correct answer, use a deterministic tool. If it has many good answers, use AI.

File moves, counts, format checks, renaming a batch, reconciling a number against a source: those have exactly one right outcome. Use a script, a formula, a checklist, a piece of ordinary software. Summaries, themes, first drafts, tradeoffs, “what are the three things this long document is really saying”: those have many good answers and no single obvious one. That’s what AI is for.

There’s one place that rule needs a second question, and it’s worth thirty seconds because it’s where most real work actually sits. The rule assumes the machine can already see your data. A script can only act on structure it can address, so the moment the input is messy prose, a script is out regardless of how many right answers there are. Pulling a renewal date out of a forty-page contract has exactly one correct answer, and you still need a model, because nothing else can read it.

So the honest version is two questions, not one. Is the input structured? That decides whether a machine can read it at all. Does the output have one right answer? That decides what you do afterward. One right answer doesn’t mean don’t use AI. It means you can check it, so check it. Many good answers means nobody can check it for you, which is exactly where your judgment has to stay.

I learned this one the annoying way. I once used an AI agent to make folders and copy files. It worked. It also took longer, cost more, and carried a risk a plain script never would, because it could misread a path, skip a file, and still report back that everything was done. Which it then did. It told me it had processed every item. It had missed eighteen of them.

The fix wasn’t a smarter agent. The fix was one line of boring code that counts things.

That’s the whole trap in one story. A model wrapped around a task that only needed plumbing looks impressive and costs you the one cheap check that would have caught it. Trust the smart tool for judgment. Trust the dumb tool for counting.

Tier one: jobs where the answer is easy to check

This is where almost every small business should start, and where most of the available value is sitting right now.

The test for this tier isn’t “is it important.” It’s: can I tell in about ten seconds whether this is right? If yes, hand it over.

First drafts of anything. Not finished work. First drafts. The blank page is the expensive part, and a mediocre first draft that you cut in half is faster than a good one you write from nothing. The thing that separates this from generic slop is giving it a specific role and real source material instead of a blank instruction. “Write me a blog post about X” gets you fifteen hundred words of polished nothing. Every time.

Summarizing and extracting. A long report, a contract, a call recording, a thread with forty replies. Pulling out the decisions, the dates, the commitments, the numbers. This is the single most underused capability in small business and it’s almost risk-free, because the source document is right there to check against.

Preparation. This is my favorite because it’s small and it changes a whole week. Before a call I get a short brief built from the person’s name, their LinkedIn, and their email history. Two minutes of reading instead of ten minutes of scrambling, on every call, forever. Nobody sells this as an AI use case because it isn’t impressive. It just quietly gives you back an hour a week.

Triage and sorting. Read the inbox, flag what genuinely needs a reply today, draft the routine responses for approval. Nothing sends without you.

Tier two: work you already do, redesigned

The step most people skip is mapping the process before assigning any tool.

I spent two hours with a nonprofit team who wanted help using AI, and we went the first hour without opening a single tool. That wasn’t a clever plan on my part. We just kept hitting questions nobody could answer, so we stopped and wrote down how their content actually got made: what they were researching, where the trusted sources were, who wrote, in what voice, how they caught the parts that sounded robotic, and what they did with the numbers afterward.

They ended up with five steps on a whiteboard, and two of them had never been said out loud before. That was the useful part of the afternoon, and none of it was about AI. Once the five steps existed, picking tools was almost boring, and each step wanted a different one because they were different kinds of job.

Two things came out of that afternoon that I now say to everyone.

Break it into the smallest pieces, get one working, then copy it. A professional audio mixing desk looks terrifying until you notice every row is identical. It’s channel one, repeated. The complexity is an illusion made of repetition. Don’t build the whole desk. Build one channel.

Go through it manually first. The instinct is to automate immediately, connect everything, build the pipeline. Don’t. Run the loop by hand three or four times, clunky and slow. You’ll discover that step three should have come before step two, or that there’s a sub-step nobody had ever said out loud. You don’t actually know what your process is until you’ve watched yourself do it. Automating a process you haven’t examined just means being wrong faster.

Tier three: the things you wrote off

This is the tier nobody talks about, and it’s where the real change is.

Somewhere in your business is a list of problems you stopped treating as problems. Not because you solved them. Because the fix required a person you could not hire, or a budget that never made sense, so the problem quietly became weather.

I have a personal one. I’m not a debater. When my son needed help preparing for debate, I couldn’t coach him on cross-examination or rebuttal. What he needed was a sparring partner available at ten at night, on any motion, who would give him real criticism instead of encouragement. That person doesn’t exist at a price a family pays. I’d written it off completely.

So I stopped trying to be the coach and built him something to practice against instead. Counter-arguments on demand, structure and clock enforced, honest feedback on clarity and logic.

You don’t need to be an expert in the content to be an expert in the container.

Look for the same shape in your business. The follow-up nobody has time to do. The onboarding that would be great if someone had two hours per client. The analysis you would run monthly if it didn’t take a day. The training that never happens because the one person who could deliver it is billing. Those were all capacity problems, and capacity is the thing that just got cheap.

That’s a very different question from “which AI tool should I use.” It’s: what did I give up on, and was it actually impossible or just expensive?

Where a person has to stay

None of the above works unattended, and the reason is specific rather than sentimental.

A model only ever has the general case. It doesn’t know that this client is fragile this quarter. It doesn’t know you can’t afford to be wrong in front of that particular account. It doesn’t know which of your people is one bad month from leaving. That knowledge is the thing your years in the business bought you, and it’s not for sale at any subscription price.

So the rule for where people stay isn’t about seniority or trust. Keep a person wherever being wrong is expensive, and wherever the output is about to touch a customer.

And check outputs, not statuses. A completed task is only a claim that instructions were followed. Following the wrong instructions correctly is the most common way this goes bad in a real company, and a green check won’t tell you that happened.

What to do this week

Pick one job from tier one. Something small, checkable, and genuinely on your plate. Run it by hand a few times, write down what you actually did, and hand over one step. Not the whole thing.

Then look at the list of things you gave up on and ask which one was a capacity problem.

That second question is usually where the money is, and almost nobody asks it, because you have to remember what you stopped wanting first.

If you want help working out which jobs in your company fit which shape, that’s most of what I do. KyberFive works with founders of 5 to 50 person companies as an embedded Chief AI Officer. Start a conversation.