How to Start with AI in a Mid-Sized Company: A Step-by-Step Plan

In short

Pick one repeating process with a named owner, available data and a written baseline. No company-wide strategy, no website chatbot.
Run a 6-8 week pilot with human review in one small group. Let AI only suggest, sort and check.
Set success thresholds before the start, such as time saved, customer satisfaction gain and a low rate of critical errors. Expand only if you meet them.
Train the people who work with AI. Article 4 of the EU AI Act requires companies to take measures that support AI literacy.

Start AI in a mid-sized company with one self-contained process that repeats and has a single owner. Write down how it works today, then run a 6-8 week pilot in which AI only assists and a person reviews the result. At the end, you decide on the numbers, not on impressions. You need neither a grand plan nor a platform purchase.

Where companies stand

According to the Czech Statistical Office, 18% of Czech businesses with 10 or more employees used at least one of eight tracked AI technologies in 2025. The EU average was 20%. Generative AI, meaning tools that create new text or code from a prompt, was used by 13% of Czech firms against 9% on average in the EU. The statistic does not break firms down by size, so it tells you nothing more about your own category.

Step 1: Choose the right process

The consultancy Pragmatic Coders recommends comparing ideas on two axes: business impact and how realistic implementation is. You start where both are high. An idea with a big payoff but no data or no owner should wait. This is a consultancy's text, not independent research, but the criteria are practical and you can test them in your own firm.

For your first use case, ask yourself these questions:

Question	What you want to hear
Is the process described, and does it repeat?	Yes, and not only in one person's head
Is there a written baseline and a target for 6-8 weeks?	Yes, in numbers
Does it have a single owner?	Yes, a name, not a committee
Can we get data and source material within two weeks?	Yes, such as folders, spreadsheets, CRM, procedures
Is the impact on the business noticeable?	Yes, time, errors, delays or lost deals

If you have no answer to one of these questions, do not pick that process yet. According to Pragmatic Coders, firms that cannot name the benefit often end up with an abandoned experiment.

In its guide for smaller businesses, AWS suggests starting where inputs and outputs are clear and the risk of connecting to your systems is low. Its examples include answers to frequent customer questions, search across internal documents, first drafts of texts, meeting summaries and regular reports.

What to avoid at the start

Pragmatic Coders warns against five typical mistakes:

a website chatbot just because everyone has one,
automating complex judgment calls right away,
rebuilding several systems at once,
a project whose only success is a demo for management,
the theme “AI is a company-wide matter”, which ends up belonging to nobody.

If the only argument is that an employee writes an email a few minutes faster, you have probably picked a weak case. A better one is a process where service to customers or the firm's results improve.

Step 2: Record the baseline and the success thresholds

Decide in advance what you will measure. AWS and Pragmatic Coders agree that without baseline numbers you cannot tell whether AI helped. Track four groups:

efficiency: time per task, backlog of open tasks,
quality: error rate, customer satisfaction (CSAT, the rating a customer gives after a case is resolved),
revenue: for example conversion, meaning the share of inquiries that end in an order,
risk: rule violations, handoffs of cases to a person.

AWS gives examples of decision thresholds: 20-40% time saved on the target task, a 10-point rise in CSAT and a critical error rate below 2% on important outputs. Treat them as inspiration. Your thresholds may differ, but they must be written down before the start.

Step 3: A 6-8 week pilot

AWS describes a schedule you can adopt almost unchanged:

Week 1: scope and rules. Choose one workflow. Define what AI may do, what it may not do, and when it must hand a case to a person.
Week 2: baseline and data. Measure the current state and prepare a small set of verified source material.
Weeks 3 to 6: operation with human oversight. Start in one group, for example one team or one request queue. Review outputs every week and adjust the prompts and the source material.
Weeks 7 and 8: decision. Compare the result with the thresholds. If you missed them, either adjust and repeat, or stop and try another process.

Pragmatic Coders recommends starting with assistance: AI suggests, sorts or checks, and a person decides. Increase autonomy only after the benefit is proven.

Step 4: Calculate the payback

AWS recommends a simple calculation. Monthly benefit is hours saved multiplied by the fully loaded hourly cost of labor, plus any revenue gain multiplied by margin. Payback period is the total pilot cost divided by the monthly benefit. Count three kinds of cost: licenses and usage, setup and integration, and people's time for training and quality control.

Fund only the pilot. Release further investment only when the result meets the agreed thresholds.

Step 5: Choose the tool last

According to Pragmatic Coders, payback does not depend on which model or tool you pick, but on which problem you solve. AWS lists three low-effort routes: paid cloud services with pay-as-you-go pricing, AI features you already have in your CRM, helpdesk or office suite, and simple automation that links steps together. When choosing, look at access control, data handling and the ability to export your data and prompts, so you are not locked in to one vendor.

Data, rules and people

AWS advises sending the tool only what the task really needs, masking sensitive data (personal, payment, health) wherever possible, and keeping a record of what data goes where and who has access. On top of that, write a one-page policy: what is allowed, what is forbidden and who checks what. Add three roles: the pilot owner, the keeper of the prompts and the person who spot-checks for errors.

Article 4 of the EU AI Act has applied since 2 February 2025. Companies that put AI systems into use must take measures to support AI literacy among their employees and other people who work with those systems. The measures take into account people's knowledge, education and the environment where AI is used. As it now stands, the article no longer requires guaranteeing a specific level of knowledge for every individual. The Commission and member states are also meant to give particular support to small and medium-sized enterprises and publish practical examples. For a start, AWS suggests two weeks of training: 60 to 90 minutes on the basics of writing prompts and what a good output looks like, plus short demonstrations in the very tools people use.

Checklist before you start

Can you name the process owner with one name?
Do you have today's number (time, errors, delays) written down, and a target for 6-8 weeks?
Do you know what AI may do and when it must hand over to a person?
Will you have source material and data within two weeks?
Is the pilot limited to one group and one process?
Do you know the thresholds at which you will expand, adjust or end the pilot?
Have the people in the pilot completed short training?

AI agent use cases for companies by industry Data security when you deploy AI in your company

Sources

Prepared by AI from the listed sources and checked automatically against them. Not legal advice.