Where to Start with AI When You Work Alone or in a Small Team

Start with a repetitive task you do yourself every week and where a mistake costs little: a data entry job, a first draft of an email, sorting your inbound requests. Writing help comes next, then analysing the data you already have. One entry point at a time, and you measure before adding another.

Everyone talks about AI, the tools keep multiplying, and "where do I even start?" still has no clear answer. When you work alone or with three colleagues, you have no IT department and no six months to train. You have a busy week and a list of jobs you'd rather not do.

To keep this concrete, I follow two made-up businesses. Samir is a freelance graphic designer: quotes, follow-ups, client emails. Lina's web agency has four people: meeting notes, project tracking, sales exports. The situations and the figures are invented, and I say so each time I give one.

I rate the risk by who sees the mistake first: you, a client, or a decision already made.

What you have that a big company doesn't

A large company has more budget, but also legacy systems, chains of approvals and teams to convince. You decide alone to change how you work, you can test an automation on one client this week, and you know your clients by first name. Those advantages are real.

Three mistakes at the start

Before the good entry points, three common traps:

• Starting from the tool instead of the problem. "I'll put an AI chatbot on my site" names a tool and no problem. "I keep answering the same three questions" is a project. The tool comes after.
• Underestimating the state of your data. If your clients are spread across a spreadsheet, your inbox and a notebook, the first job is to tidy them, not to add AI.
• Trying to automate everything at once. What lasts starts small: one task, one tool, two weeks. You measure, adjust, extend.

First entry point: repetitive tasks

For one week, write down every task you do by hand on autopilot: copying data from one tool to another, writing the same email, filing a receipt, chasing a client. At the end, look at the ones that come back most. A rule of thumb of mine: above two hours a week, the task deserves a look first.

Say Samir retypes the details of a quote into his invoice: fifteen minutes, twelve times a month, so three hours. With a template pre-filled by an assistant, it drops to five minutes per quote, so one hour: two hours recovered, as long as he rereads every invoice. These figures are invented, to show the method. In Lina's agency, formatting the meeting notes comes back every Friday.

How to tell whether a task can be automated

Three questions are enough:

1. Are the rules explicit? Can you describe the task in clear steps? If so, it's a good candidate.
2. Does the data exist in digital form? An email, a spreadsheet or a readable PDF counts. A note scribbled on paper doesn't.
3. Can you recover from a mistake? If you can fix it before it reaches a client, you can start. If it goes straight to them or touches an invoice, review it before sending.

A task that gets a yes on all three is your first test. For cases already worked out in detail, read Five Concrete AI Automations for Your Business. If you're unsure which workflow tool to pick, Automating Without Code: n8n, Make or Zapier, Which One to Choose as a Beginner? sorts them out.

Second entry point: writing help

Quotes, replies to enquiries, service descriptions, meeting notes, posts: you write a lot, and the first version is often the slow part. An assistant such as ChatGPT, Claude or Gemini can give you a first draft from your notes, rephrase a message to change its tone, or summarise a long document.

Here is a starting prompt for answering a quote request. Fill in the bracketed fields:

You're helping me reply to a quote request. Here is the request I received:
[paste the client's message, without their name or address]

Here is what I offer: [your service, your deadline, your rate and its unit].
Write a first draft of a reply in four sentences at most, in a direct and polite tone.
Don't invent any rate, deadline or guarantee that isn't in what I gave you.

What AI won't do for you

An assistant doesn't know your clients, what you've already told them, or how your market works. It produces plausible text, sometimes generic, sometimes wrong on one detail. The rule I'd give you: AI writes the first draft, you check and personalise it. Reread the figures, the dates and anything that commits your word. To see why a model states false things with confidence, read AI Invents Figures, Sources and Terms.

Third entry point: the data you already have

Sales history, support tickets, client feedback: this data hides signals you don't have time to spot by hand. Three uses within reach without a data team:

• Sales: see which services leave you the most margin, which clients no longer come back, which months are slow.
• Inbound requests: group the questions that repeat, so you know what to put in an FAQ or a reply template.
• Targeting: from your current clients, spot the ones who could buy a complementary service.

To start, a CSV export of your sales and a general-purpose assistant are enough: you ask your questions in plain language. Excel and Google Sheets also have assistant features, depending on your plan.

The prerequisite: clean data

This is the least pleasant step and you don't skip it. Duplicates, empty fields or mixed formats give unusable analyses. Before asking an AI anything, answer three questions:

• Is your data in one place, not scattered across five tools?
• Is it up to date, filled in regularly and not in a rush?
• Can you get to it without searching three inboxes?

If one answer is no, start there. It isn't wasted time: a tidy spreadsheet helps you with or without AI.

When the job goes beyond a manual test

A first test works with an assistant and a spreadsheet. When the job needs several tools wired together, has to run without you, or handles regular volumes, it becomes automation to build, sometimes a small custom AI agent. That is also what I build for freelancers and small businesses: if you want me to take it on, write to me. You don't have to go there to start: one task, one test, two weeks.

What you can do this week

Pick one entry point, and within it one task. Note how long it takes you today, set up your test, then let it run for two weeks before comparing. That is my own habit, not a rule: on a narrow scope, two weeks is often enough to know whether it holds. To work out what it earns you, read AI ROI: how to measure what a tool really earns you.

Frequently asked questions

Which entry point should I start with when I work alone? Repetitive tasks, because they're the easiest to scope: the rules are clear, you quickly see whether the result is good, and a mistake gets fixed before it reaches a client.

Can I paste my clients' messages into an assistant? Remove what identifies the person first: name, address, number. The CNIL points out that what you send to a generative AI tool can contain personal data, and that you should check where the tool hosts it and whether it reuses it. I expand on this in the article linked above.

When should I stop a test? On the review date you set yourself, if the time spent on the task hasn't dropped. A tool that doesn't move your number doesn't have to stay.

Do I need a budget to start? I can't give a figure: general-purpose assistants have free tiers and paid plans whose prices change. Check current prices on the vendors' official pages, and test with a free tier first.

The logical next step is From Solo to Small Team: What AI Replaces, and What It Will Not, which looks at what AI changes as your business grows. My next tests go out through my newsletter, you can sign up right below.