Bringing AI to a small business doesn't take a pilot project with a steering committee, it takes one single role that you test for a week before you add the next one. Most guides online describe six phases with a target picture, a pilot team, and a scaling stage, built for companies with their own IT department. For a business with five people, that's too much process for too little payoff.
I run my own business with an AI workforce of eight AI employees built on Claude, and along the way I learned how a rollout actually works without a project budget. What an AI employee is in the first place, and how it differs from a chatbot, is on the AI employees page. Here I walk through the rollout as four weeks with four roles, the same way I'd recommend it to a small team.
Take a business with five people as the example: the owner, two in sales or client work, one in admin, one apprentice. No IT department, no budget for a consulting project, but a full inbox and proposals that keep piling up. That's exactly the business this plan is built for.
Why a six-phase roadmap is too much for five people
Roadmaps with a baseline assessment, use case selection, ownership mapping, a multi-week pilot phase, training, and only then scaling make sense for companies with a hundred or more employees. There, several departments have to coordinate, and a wrong call quickly affects a lot of people at once.
In a five-person business, that's different. The owner decides alone, there's no pilot team because the whole team is the pilot team. A multi-week baseline assessment there costs more time than the first AI role saves afterward, before anything is even running. What actually matters in a business like that is a role that makes a visible difference in week one, not a target picture on a slide.
That's why the plan below replaces the six phases with four weeks, one new role each. Every week ends with a decision: keep going, adjust, or stop. This isn't a project, it's a series of small trials.
AI for small business: the 4-week plan with no IT department
The order of the roles follows what eats the most time and is easiest to check: first the office, then proposals, then finding customers, then the business's own website. The table below shows the flow.
| Week |
Role |
Task |
What you approve that week |
| 1 |
Frieda |
Sorting the inbox, setting up appointments, filing receipts |
every reply that goes out to a customer |
| 2 |
Anton |
Drafting proposals and invoices from the sorted inquiries |
every proposal and every invoice before it's sent |
| 3 |
Olaf |
Finding and pre-qualifying leads that fit the business |
every outreach to a new contact |
| 4 |
Ralf |
Building a modern demo version of the business's own website |
the decision whether the demo goes live |
Week 1: Frieda takes over the inbox. A concrete day looks like this: twelve new emails sit in the inbox in the morning, one customer inquiry, three follow-ups on running jobs, one supplier invoice, the rest newsletters. Frieda attaches a short note to every email saying what it is and how urgent, turns the meeting request into a calendar entry, files the supplier invoice as a receipt, and writes a reply draft for each of the three follow-ups. The admin person in the example business reads the drafts, fixes what's off, and sends them herself. By the end of the week, the team decides from its own experience whether the sorting held up and whether Frieda should also take over scheduling for next month.
Week 2: Anton takes over proposals and invoices. From the sorted inquiries, Anton drafts a proposal, with structure and prices following the business's own logic, no invented line items. The owner reads every draft, changes what's wrong, and sends the proposal herself. If a customer accepts, the invoice draft follows the same flow. By the end of the week, the team knows how many proposals went out without major edits, and whether the second step is worth it.
Week 3: Olaf finds matching leads. Instead of the sales team searching directories and networks themselves, they get a weekly list of pre-qualified contacts with a short note on why each one might fit. Which of them actually get contacted is still a human decision on the team, not Olaf's.
Week 4: Ralf builds a website demo. From the existing, dated website, a modern demo version comes together to look at, before anyone spends money on a redesign. The decision whether that demo actually goes live stays with the business.
Approval rules before the first AI employee starts
Before Frieda reads a single real email in week one, two things have to be settled. First the approval rule: anything that leaves the building, a reply to a customer, a proposal, an invoice, an outreach message, goes out only after a human approves it. Second the privacy and access rules: which mailbox or folder the AI can reach at all, what data is allowed to sit there, and what happens when it finds something sensitive. That belongs before the first access to real customer data, not in week three.
Internal steps then run without a check, but not all of them carry the same weight. A file put in the wrong folder is fixed in a minute; a missed receipt, a missed deadline, or a wrong calendar entry can get expensive. So I would rather start narrow: drafts instead of execution in the first week, test data or a clearly bounded read-only area instead of the whole mailbox, and more room only after a few clean runs. What also matters is that the rules apply to everyone on the team, not just the role that got introduced first.
Week 4: measuring results and deciding whether to expand
At the end of week four, the team sits down once and answers the same question for each of the four roles: did the work visibly change, and was the quality of the drafts good enough that correcting them took less time than the original task? Where the answer is yes, the role stays in daily use. Where the answer is no, the team either sharpens the briefing or drops the role again, and that doesn't count as a failure.
I wouldn't invent fixed numbers like hours saved at this point, because they differ too much between businesses and roles. What counts is the direct comparison: did the task feel noticeably lighter than before, or not. The team can make that call from its own experience after four weeks, no spreadsheet required.
Two typical mistakes when starting out
A common mistake is starting two roles at once because both feel urgent. By the end of the week you cannot reliably say whether either one worked. Start only one role per week, even when a second one feels pressing.
The second mistake is briefing too broadly, on the assumption that more tasks mean more value. The result is an employee who helps a little with a lot of things and gets really good at none of them. Start with a narrow, clearly scoped task and only widen it once the first one runs reliably. How you set a role up technically as its own project with its own instructions is shown in the Claude help center.
FAQ
Which task should I choose first when bringing AI to my business?
The one that comes up most often and whose result you can check the fastest. For most small businesses, that's the inbox, because it happens every day and a reply draft can be judged in seconds.
Do I need a dedicated budget to get started?
For the four roles in this plan, a single AI tool subscription is enough. Exact prices, and when a ready-made agency solution is worth it instead, are in What an AI Employee Really Costs.
Does my team need training for this?
A short walkthrough on how to read and approve or reject a draft is enough. A multi-day training isn't needed for the four roles in this plan, because the real work happens in writing the briefing, not in operating a tool.
When do I keep expanding after the four weeks?
As soon as a role runs reliably and the team trusts it in daily use. After that, the next role follows the same pattern: test for a week, then decide, rather than starting several new tasks at once.
What if one of the four roles just doesn't fit my business?
Then you drop it, or sharpen the briefing and try it for another week. Not every role fits every business, and dropping a role that isn't paying off isn't a step back, it's the outcome of the week-four check-in.
Where to go from here
Before you start week one, it's worth looking at your own workflows: which task repeats the most and can be judged quickly? The method for finding and evaluating candidates is in Process Optimization With AI: The Method. Once you've picked a role, Hiring an AI Employee: The Process shows how you get from an empty chat to a permanent role.
Frieda, Anton, Olaf, and Ralf are four of eight ready-made packages you load and onboard, in my community Claude Practitioners.
