AI plan for 2027: how to set AI strategy and budget in a mid-sized company

Polish companies still use AI less often than the average EU company. According to Eurostat, 8.4% of companies with at least 10 employees did so in 2025, against an EU average of 20%. A year earlier the Polish figure was 5.9%.

These figures cover companies that use AI in their business, not employees using a chatbot on their own initiative. For a 2027 plan, they point to one conclusion: in most industries it is not too late. Competitors are only just starting too, so care matters more than speed.

What has changed is the legal and technical environment. The points below show what this means for next year’s plan.

The annual plan is a good place to collect the dates that will affect AI projects and their costs. The table below lists those that apply to many mid-sized companies. Not every item will apply to you, so go through them with your legal team or adviser.

The most important change concerns HR. Under the AI Act, AI used to screen candidates, assess employees or decide on promotions is a high-risk system. From 2 December 2027, such a system must meet additional requirements, including on human oversight and documentation, so build them into the project now.

Some companies are also covered by the amendment to the Polish National Cybersecurity System Act (KSC Act), which implements the EU NIS2 Directive. It applies to selected sectors and to companies from medium size upwards. If your company is one of them, AI projects must fit within the security rules that have to be in place by April 2027.

The simplest approach is to split the year into quarters, because the board reviews results at that pace anyway. Each stage ends with a decision: continue, improve or stop. Without one, an AI project easily turns into an endless experiment whose purpose nobody remembers.

Quarters are a guide, not a rigid rule. If choosing processes takes six weeks, the pilot can start earlier. What matters is that every stage has an expected result and a decision date written down in advance.

This stage is not a technology review but a review of your processes. You are looking for places where people do a lot of repetitive work on documents, emails or data, and where mistakes are costly. For each candidate, measure the current situation: the number of cases per month, the time needed to handle one case and the number of errors.

Take a hypothetical distribution company with 150 employees. Its candidates are orders received by email, preparing quotations and the monthly sales report. It chooses orders, because several hundred arrive every week and the data goes into a single ERP system.

A good candidate for a first project meets most of the criteria in the list below. If a process fails three of them, put it aside for later – even if it sounds attractive.

A pilot means running the solution on some of the cases, for example for one type of email or one team. Its job is to check whether AI works with your data and your people. A demonstration of what the tool can do is not enough.

Before you start, write down three things: the measure of success, the threshold for stopping and the decision date. For example: “the draft reply needs corrections in fewer than half of cases, and handling time falls by at least 30%”. If the result is worse by the agreed date, you analyse the causes and decide whether to improve the solution or stop the project.

During the pilot, a person checks every AI result before it goes any further. This takes time, but it gives you data for the decision: how many results were good, how many needed corrections and which errors keep recurring.

Do not buy licences for the whole company before the pilot ends. If the pilot uses the Microsoft Copilot assistant, Microsoft 365 Copilot licences for the test group are enough.

Scaling up means more than switching the solution on for everyone. You need to train further teams, write instructions, decide who corrects errors and review access rights to data. Microsoft Copilot shows users everything they have access to, so a payroll folder open to everyone becomes a real risk.

In Q3 it is worth starting a second pilot, because the team already has experience from the first. Make sure, however, that no more than two pilots run at the same time. Each one takes up the time of the process manager and the most experienced employees.

Q4 is for improvements and a review of the year. You check whether a cheaper AI model would give the same quality, where AI is still making mistakes and whether the company’s rules for using AI keep up with practice. The conclusions become the basis for the 2028 budget.

The most common mistake is a budget set from the top: “we are putting this much into AI”. Such a figure has no link to the benefit, so it is hard to defend and to account for later. It is better to calculate from the cost of the processes you want to change.

The expected benefit sets the upper limit of the budget. Suppose handling orders from emails takes 2,500 working hours a year, and a third of that – about 830 hours – can realistically be saved. At a labour cost of, say, PLN 60 an hour, that is about PLN 50,000 a year, or PLN 100,000 over two years. This is the break-even limit for the first two years: a sensible budget should be clearly below it. You will compare every offer against this figure.

Break the budget down into the stages and components in the list below, and release the money in tranches. The board approves the whole plan, but funds for scaling up reach the project only after a positive decision following the pilot. If the pilot fails, you lose the cost of one stage, not of the whole year.

Licences can be calculated from the price list. For example, 30 people on Copilot Business cost 30 × USD 21 × 12 months = USD 7,560 a year, based on the October 2026 price list. The other components depend on the scope, and the provider prices them after analysing the process.

The level of ambition depends more on the company’s experience than on its size. A 300-person company with no projects behind it should start cautiously. An 80-person company with two successful pilots can afford to do more.

Skipping levels rarely pays off. Without a pilot, a company does not yet know how much of its people’s time a project will take, and it usually underestimates this. The table below describes the levels by scope of work rather than by amount, because the amount depends on the processes chosen.

The board does not need a report on the number of queries sent to the AI model. It needs a few figures that show whether processes are running more cheaply, faster and with fewer errors. Compare each indicator in the list below with the baseline measured in Q1.

Financial indicators such as margin or revenue per employee are also worth tracking. However, many factors affect them at once, so it is hard to attribute a change to AI alone.

The mistakes below appear in plans regardless of industry. Each of them can be avoided with a single line in the plan before the board approves it.

A year-long programme does not always make sense. If the company has no process that recurs often and has clear rules, it is better to plan a single project. Return to decisions about further projects once it is finished.

The same applies when nobody in the company has time to run a project, or when the company is struggling with basic problems such as late payments or stock shortages. In that case, a budget line for staff training and written rules for using AI is enough for 2027. If employees use AI, Article 4 of the AI Act requires measures to build their knowledge of the technology anyway, and training is the simplest of them.