Short answer: Start with one task you already repeat every week, measure how long it actually takes, and try the plainest possible AI version of it in a tool you already pay for. Write down what worked well enough that someone else on the team could repeat it, then decide whether to build on it, train people on it, or leave it alone.
Most small businesses that struggle with AI did not struggle because the tools are hard. They struggled because they started in the wrong place: a platform demo, a vendor pitch, or a headline about what some other company automated. None of that tells you what your business needs. The plan below does, because it starts with your own repeated work instead of someone else's.
Step 1: Pick one repeated task
Not a category. One specific task. Not "customer service," but "writing the reply to the three or four emails a day that ask the same pricing question." Not "marketing," but "drafting this week's social post from the notes I already jot down."
The task should meet three conditions. It happens on a schedule, at least weekly. You personally know what a good result looks like, because you have done it by hand many times. And it currently takes real time, not five minutes you would not miss.
If nothing comes to mind immediately, that is a sign to look harder before moving on, not a sign to skip this step. Ask two or three people on your team what they do every week that they find tedious. The answer is rarely a mystery once you ask.
Step 2: Measure how long it actually takes
Time the task honestly for a few instances before you touch AI at all. Most people are wrong about how long their own recurring work takes, usually by underestimating it, because it is spread across small chunks through the day instead of one visible block.
This matters for two reasons. It gives you a real baseline to compare against later, instead of a vague sense that something got faster. And it is often the moment you discover the task is not worth automating at all, because it turns out to take four minutes a week, not the twenty you assumed. That is a useful outcome too. It saves you from building something to fix a problem that was smaller than it felt.
Step 3: Try the simplest version, in a tool you already pay for
Resist the urge to shop for a platform at this stage. Open ChatGPT, Claude, or whatever AI feature is already sitting inside Google Workspace, Microsoft 365, your CRM, or your accounting software, and just try the task.
Give it real detail. Not "write a follow-up email," but the actual customer name, the actual job, the actual tone you use. The first attempt will not be perfect. That is fine. You are testing whether the shape of the task fits AI at all, not building a finished workflow yet.
Do this three or four times across a few days, on the real task, not a hypothetical version of it. By the third or fourth attempt you will know two things: whether it consistently saves you time, and what you had to tell it each time to get a result you would actually use. For a library of starting prompts, see ChatGPT prompts for business owners. If you want a live look at what a working version looks like before you try your own, the demos page shows a missed-call agent, inbox sorting, and a before-and-after from team training.
Step 4: Write it down so the team can repeat it
The most common failure point is not the AI part. It is that the good version of the task lives only in your head, or in a chat history nobody else can find. If it worked, turn it into something repeatable: the exact prompt you used, the tool it runs in, and the one or two details that always need to change.
One page is enough. A shared doc, a note in the CRM, a pinned message. The test is whether someone else on the team could follow it without asking you a question. If a task only saves time when you personally do it, it has not actually become part of how the business runs. This is also the point where you find out if the task is a genuine quick win others can pick up, versus something specific to how you personally think about the work.
Step 5: Decide build, train, or leave it
With a working, written-down version in hand, you have enough information to make a real decision instead of a guess.
Build when the task is valuable enough and repeatable enough to wire together properly, connecting it to your CRM, your calendar, or your accounting system so it runs with less manual copying and pasting. This is project territory, not a quick win, and it is where a custom build or a managed agent earns its keep.
Train when the task itself is simple enough to stay manual, but more than one person on the team should be doing it this way. The bottleneck is not the workflow, it is that only you know it exists. Team training and a structured session close that gap faster than a written page ever will, because people ask questions live and leave having actually done it. If adoption has stalled even after you showed people, read why your team isn't using AI before assuming the tool is the problem.
Leave it when the task turned out to be smaller than it felt, or when it depends on judgment specific to each situation that AI keeps getting wrong. Not every task should be automated, and knowing which ones to leave alone is as much a part of this process as knowing which to build.
What not to do first
Do not buy a platform before anyone has tried the free tools on real work. The order matters: using the tools first changes what you would actually ask a platform to do, and skips paying for a scope built on guesses.
Do not start with your hardest or most sensitive task. Save the process that touches customer payment details or a legal document for after you have a few easy wins and a clearer sense of where AI is reliable in your business. Get a sense of the guardrails first; the AI policy generator is a fast way to put a basic one in writing before you go further.
Do not try to fix five tasks at once. Parallel experiments blur together, and you end up unable to tell which prompt or which tool actually caused the improvement. One task at a time, done properly, beats five tasks half-tried.
Do not skip the measuring step because it feels like busywork. Without a real before-and-after, "this feels faster" is the only evidence you will have, and that is not enough to convince a skeptical team member or to know if a bigger investment is justified.
Use the free tool to structure this
If you have more than one candidate task and are not sure which to pick first, the free What to Automate First tool does steps one and two for you. List up to five things you do every week, and it ranks them by how well AI fits each one and gives you a starting point for each. It is the fastest way to turn "I know I should be using AI more" into an actual first task.
Once you have run this process on your own, you will already have a clearer picture than most vendor pitches will give you. For a longer list of what this can look like once you are past the first task, see 20 AI automation examples for small business. And before you decide the next task needs outside help, it is worth checking your own readiness first with the AI readiness checklist.
If you would rather have a second set of eyes on which task to pick and what a realistic first result looks like, that is what a two-week assessment is built for: it looks at your actual work, not a generic list.
Frequently asked questions
How do I start using AI in my small business?
Pick one task you repeat every week, time how long it actually takes, and try the simplest version of AI help inside a tool you already pay for. Write down what worked so someone else can repeat it, then decide whether it is worth building further, worth training the team on, or not worth the trouble.
What is the biggest mistake small businesses make when starting with AI?
Buying a platform before anyone on the team has spent real time with the free tools. That order almost always produces a subscription nobody uses, because the business bought what it read about instead of what its own work needed.
Do I need a big budget to start with AI?
No. ChatGPT and Claude both have free or low-cost plans, and most small businesses already own tools, like Google Workspace or Microsoft 365, with AI features built in. The starting cost is time, not money.
How long before I see results from starting with AI?
A single well-chosen task can show a result the same day: a drafted email, a faster reply, a summarized document. What takes longer is turning a personal habit into something the whole team relies on, which usually takes a few weeks of using it, writing it down, and adjusting.
Should I train my team or hire someone to build AI systems first?
Neither, first. Start by using a tool yourself on one real task. Training and building are both worth doing, but they work far better once you already know, from experience, what a good use case looks like in your own business.