"How do I start using AI in my small business without hiring an expert?" has a better answer than "hire an expert", and it is not "buy a tool" either. You can start with nothing but a few conversations, one sentence of rules and one boring task, and you can run it yourself. This article is for the owner or CEO of a company with no IT department, perhaps twenty to a few hundred people, who typed that question, or "first AI project for a small business", or "where should a small business start with AI", and wants a plan that fits in a calendar. The overview of the decisions behind it is in how to start using AI in your business. This piece is what to do on which week.
I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them, which meant I watched many companies with no technical staff adopt, or fail to adopt, new systems. I am a lawyer by training. Now I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies through 013 Labs. Nothing here is legal advice; have your own lawyer review the final wording of any data rule.
Why a missing IT department is less of a problem than it sounds
The tools an owner would try first are ordinary web accounts. You sign in, type a request in plain English, and look over the answer. There is nothing to install, no server and no code. The skill that matters is the one you use to brief a new hire: say what you want, check what comes back, and decide who is responsible.
What an IT department would normally give you is not the software. It is three protections: someone who knows what is being used, someone who sets the rule about what data goes where, and someone who is accountable when something goes wrong. Those three jobs now fall to you, and they are small enough to do in a month. Week one finds out what is already happening, week two writes the rule, weeks three and four run one small piece of work, and the last day decides whether to keep it. There is no step called "choose the best tool", on purpose: picking a tool before you know the task and the data rule is how companies pay for something nobody uses.
Week one: find out what your staff already use
Somebody on your team almost certainly has an AI tool open on their screen right now, probably on a personal account, probably for something sensible like drafting an email or tidying a spreadsheet. You do not need to catch them, only ask.
Send one short, friendly note to your managers: "Who here uses an AI tool for work, and what for? Nobody is in trouble. I want to know what is already helping." For each, write down five things on a single sheet of paper:
The tool. Its name, and whether it is a personal account or one the company pays for.
The task. What they do with it, in their own words.
The data. What they paste in. Customer names, prices, contracts, employee details, or nothing sensitive at all.
The time it saves. A rough guess is enough.
The worry. Anything that makes them nervous.
Two things come out of this. You learn what your own people find valuable, a better guide to a first project than any article. And you learn who the natural leader is, the colleague everyone else asks. Note the name for week three.
If someone admits to pasting customer details into a free personal account, thank them and move on. Punishing the honest answer guarantees you will never hear another.
Week two: write the one-sentence data rule
A policy that runs to eleven pages is signed and forgotten. A rule short enough to repeat from memory gets followed. Your job this week is to write one sentence on what may go into an AI tool.
A common starting point is: "Nothing that identifies a customer or an employee goes into any tool the company has not approved." Make it stricter or looser as you like, but it must fit in one breath, a new hire must be able to say it back, and it must name who decides what counts as "approved", which is you or one person you choose.
Two practical points sit behind it. Decide whether the company will pay for a business account rather than rely on personal ones, since you can then see who has access and end it when someone leaves. The article on whether ChatGPT is safe for business use sets out the plan and terms questions to answer first, and you should write the date next to what you learn, because these products change often. And tell people: say the sentence at the next staff meeting and send it in writing the same day. A rule nobody has heard does not exist. If you later want something longer, the piece on an AI policy for employees that fits on one page shows where the sentence grows into a short document, but do not wait for the longer version before the shorter one is in people's heads.
By the end of week two you should be able to say two things: we know what is being used, and everybody knows the one rule.
Week three: pick one boring, repeated task
Now, and only now, choose the work. The temptation is to pick something impressive: a customer-facing assistant, a forecast, an automated sales machine. Resist it. The best first task has five boring qualities.
It is repeated, so you get many chances to see how it goes in a short time.
It is low-stakes. If the first draft is wrong, nobody outside the company is harmed, because a person checks it before it leaves.
It uses little or no sensitive data, so your one-sentence rule is easy to obey.
It has an obvious before. You can say today how long it takes and how often it goes wrong, which is the only way to know later whether anything improved.
It belongs to one person who already does it and wants help.
Look at your week-one sheet. The task your staff already said saves them time is usually sitting right there. If the task touches the books, read can AI do bookkeeping for a small business first, since money work has its own rules. For a longer discussion of how to compare candidates, picking your first AI project goes through it.
Then do three things before anyone starts.
Name the owner. One person, by name, is responsible for the task and for reporting back. Not a committee, and not "the team". It is often the colleague you noted in week one. Give that person the authority to say "this is not working" without having to defend the decision.
Write the before. On half a page: how long the task takes now, how often someone has to redo it, and what a good result looks like. Five minutes, no more.
Write the stop line. What would make you stop early? For example: any output that goes outside the company without a person checking it, any data that breaks your sentence, or a result that costs more time to fix than it saves.
Week four: run it for two weeks, with a check each Friday
Run the one task for two weeks. The owner of the task uses the tool for it every time it comes up, and checks every output before it is used. Nothing leaves the company unchecked.
Ask the named owner to keep a simple log, a line a day will do: what was the task, how long did it take, what did the tool get wrong, what did you have to fix. This is where you find out what is real. By the second week people usually know where the tool helps and where it is confidently mistaken, and the second is the one you cannot learn from a demonstration.
Spend ten minutes with the owner each Friday on three questions: is it saving time, what went wrong, and did anyone put in anything your rule says should stay out? If so, fix the rule or the habit that week.
When you keep a task and want more people doing it, the next step is showing them how, on your own work: how to train your employees to use AI.
The decision on the last day: keep, change or stop
At the end of the two weeks, sit down with the named owner for thirty minutes and compare the log to the half page you wrote before. There are only three outcomes, and all of them are good ones.
Keep it. The task is faster or better, the checking is quick, and nobody broke the rule. Write down how it is done, who owns it, and what gets checked, and make it the way the task is done from now on. Then, and only then, choose the next task and run the same four steps again.
Change it. It helped, but not reliably. Adjust one thing, such as the part of the task you hand over, and run it another week.
Stop it. It did not save time, or the checking cost more than the task. Stop, write one paragraph on what you learned, and pick another boring task. Stopping is not failure. It cost a few hours and told you something true about your business.
Whatever you decide, write a dated note: what you tried, who owned it, the result, and your decision. When a lender, a buyer or an insurer asks what the company does with AI, that note is the answer.
Picture the thirty days going wrong
Picture a distribution company of ninety people with no technical staff. The owner, impatient, skips week one and buys a business subscription for the whole office after a vendor demonstration. Nobody has been asked what they already use, so a sales manager keeps using her own personal account, and the company now has two accounts and no sense of which holds what. There is no data rule, so a customer service lead pastes a long supplier contract, including pricing the supplier asked to keep private, into a tool to summarise it. With no task and no named owner, after a month four people use it for four different things and no one can say whether it saved a minute. When a major customer sends a questionnaire asking how the company handles their data in AI tools, the owner cannot answer it. The repair takes a month: the same steps, in the right order, done late. This is a hypothesis, a composite and not a client.
Mistakes that repeat
Starting with a tool instead of a task. Skipping the question to your own staff. A data rule too long to remember. A first project that touches customers directly. No named owner. No written before, so no way to judge the after. No end date, so the experiment drifts on. Treating a stop as a failure.
Frequently asked questions
How do I start using AI in my small business without hiring an expert?
Run a thirty-day plan yourself. In week one ask your staff what they already use. In week two write one sentence about what data may go into a tool. In weeks three and four choose one boring repeated task, give it a named owner, and run it for two weeks with every output checked by a person. On the last day decide to keep it, change it or stop.
What is a good first AI project for a small business?
A task that is repeated, low-stakes, light on sensitive data, easy to measure and owned by one person who wants help. Avoid anything customer-facing or anything that touches money on the first try.
How much does it cost to try this?
Very little: a low-cost business subscription or a tool you already pay for, plus a few hours of one person's time. If it needs more, the first task is too big.
When should I bring in outside help?
When the owner's own month has shown you something worth repeating, or when a lender, a buyer, an insurer or a large customer asks questions you cannot answer in a few sentences. Until then, the plan above costs you attention and nothing more.
Not sure where you stand? Take the 3-minute AI readiness scorecard.