How to start with AI in a small business: where to begin, costs, and avoiding overpaying

In a small business, you don’t have a budget for experiments "just because everyone’s doing AI." You have concrete employee time slipping away on repetitive work, and the question is whether you can reclaim it without risk or overpaying. As an independent research center, we see the same mistake repeatedly: a company buys a tool before defining the problem. This guide reverses the order—first the process and the numbers, only then the technology.

A good first project meets three conditions at once: it’s repeatable, measurable, and low-risk. Repeatable—because AI pays off on volume, not a single task. Measurable—because without a "before" number, you can’t prove there’s an "after." Low-risk—because the first mistake with a client costs more than the entire pilot.

Practical test: if the process happens daily, has clear input and output, and an error can be caught before it reaches the client—it’s a good candidate. Classic starting points include classification and routing of requests, preliminary responses in customer service, or an assistant for an internal knowledge base. Avoid starting with processes where the model decides on money, client rights, or health—there, human oversight is required, and the bar is much higher.

The cost of AI isn’t a single number. It consists of three parts, and most companies only see the first. The ranges below are realistic orders of magnitude for a small business in Poland in 2026—not a price list, just a reference point for discussion. A fuller breakdown is in the article how much an AI agent costs.

The biggest impact on the ongoing bill is matching the model to the task, not the model itself. A small, cheap model for classification and a powerful one only where truly needed—this is usually the biggest single cost lever. The number worth tracking is the cost per completed task (handling one request, classifying one document), because that’s what you directly compare to the cost of human labor.

Don’t deploy "straight to production." Run a narrow pilot on real but limited data, with a clear success criterion set before starting—for example, "the model misroutes fewer than one in ten requests" or "reduces first-response time by one-third." The pilot should answer one question: does this work well enough to scale? We’ve detailed the path from pilot to production in the article from AI pilot to production.

A small business rarely should build from scratch. A ready-made tool wins when your process is standard (e.g., a typical FAQ chatbot) and the data isn’t sensitive. A custom solution makes sense when the process is specific to your business, the data is confidential, or you need full control and predictable unit costs at higher volumes.

The third option, most common for small businesses we work with, is assembling from ready-made, open building blocks—your own logic, but based on proven models and embeddings, without writing everything from scratch. It provides control without the cost of full R&D.

• Hype over problem. If you can’t name the process and the number you’re improving, don’t buy anything yet. AI is a tool for specific work, not an end in itself.
• AI where it’s unnecessary. Some processes are better fixed with rules, templates, or data cleanup. If simple automation handles it cheaper and more reliably, the model is an unnecessary risk.
• Vendor lock-in. A solution you can’t exit without rewriting everything is a hidden cost. Ask about data export, open formats, and the ability to change models.
• No boundaries or control. A model without input and output guardrails will eventually say something it shouldn’t to a client. Boundaries, logging, and oversight aren’t add-ons—they’re prerequisites for launch.
• Ignoring personal data. Before sending anything to a model, determine what data goes there and whether it’s allowed. We expand on this in the article about company obligations under the AI Act and RODO.

1. List 3 repeatable processes and calculate for each: how often per month, how much time it takes, and the cost of an error.
2. Pick one—the most repeatable, measurable, and lowest-risk.
3. Set a success criterion in numbers before starting (time, accuracy, cost per task).
4. Run a narrow pilot on limited data, with human oversight.
5. Measure and decide—scale, improve, or drop. Every outcome is information.

Start with one process that’s repeatable, measurable, and low-risk—like sorting requests or an assistant for an internal knowledge base. First, calculate how often it repeats monthly and what it costs today, then choose the tool.

The first pilot of a simple process usually fits within a range of several to tens of thousands PLN for implementation, plus a small variable model cost that grows with volume. This is an order of magnitude, not a price list—the final amount depends on the number of integrations and data requirements.

If the process is standard and the data isn’t sensitive—buy ready-made. If the process is specific to your business or the data is confidential, consider a custom solution, preferably assembled from open, proven building blocks to avoid vendor lock-in and excessive R&D costs.

Start with one narrow process, measure the cost per completed task, and route all calls through a layer that matches the model to the task—a small, cheap one where sufficient. This is usually the biggest savings, larger than negotiating the tool’s price itself.

When you can’t name the process and the number you’re improving, or when a simple rule or template handles it cheaper and more reliably. AI can be an unnecessary risk where cleaning up data or deterministic automation is enough.