When we talk to a business owner about automation, it almost never starts with artificial intelligence. It starts with the sentence "this eats an hour of my morning every day". Usually it is about retyping orders from the online shop into Pohoda (a Czech accounting system), or enquiries that someone copies from email into a spreadsheet so they don't get lost. Sometimes it is a report that gets painstakingly stitched together from four tools every Monday and that nobody reads anyway.
According to the Czech Statistical Office, 17 % of firms with ten or more employees were using some form of AI technology in 2025. Among large enterprises it was 54 %. Between those two numbers is the space where most of our clients live: a company with five to fifty people that has an online shop or a website, an accounting system, some kind of CRM, and in the middle of it all a person with four browser tabs open.
This article is about how to pick the process worth automating first, and what it costs. We build in n8n, so we won't get around a comparison with Make and Zapier, and we'll also get to where AI belongs in a workflow and where we don't want it.
What process automation means for a small business
Automating a business process is a simple thing: an event in one system triggers steps in other systems without a person sitting between them and clicking. An order in the online shop issues an invoice. A submitted form creates a contact in the CRM and sends a notification to Telegram. A new invoice in the accounting system is saved to a shared drive and logged in an overview.
Technically it is held together by webhooks, the APIs of the individual systems, and a tool that connects them into a workflow. It has nothing to do with robotic factories or enterprise RPA. We are talking about Fakturoid, Pohoda, Shoptet, WooCommerce, Raynet, Pipedrive, Google Workspace and Zásilkovna, the systems Czech companies actually use, and which, with a few exceptions, have usable APIs.
And one more thing worth saying right at the start. Most of the workflows we build have no language model in them at all. They are ordinary deterministic code that does the same thing every time, which is exactly what you want from them.
How to spot the process worth automating first
In consultations we use four questions. If a process passes all of them, it is a candidate for the first workflow.
- It repeats at least several times a week. A task you do once a quarter will cost more to automate than doing it by hand for the rest of your life.
- It has clear rules. It can be described as "when X, do Y". If the decision depends on an experienced colleague's gut feeling, leave it to them and automate what surrounds it.
- The data is already in systems. The order is in the shop, the contact in the CRM, the invoice in the accounting system. If the process starts with a piece of paper or a phone call, the input has to be sorted out first.
- A mistake is expensive, or on the contrary cheap. Both ends pay off. An expensive mistake (a wrong amount on an invoice) justifies a precise workflow with checks. A cheap mistake (a typo in an internal report) means you can automate without elaborate safeguards.
Then comes the arithmetic. Take the hours per month, multiply by the hourly rate of the person doing it, and compare with the price of the workflow. A real example: orders from an online shop retyped by hand into the accounting system, 20 minutes a day, 22 working days, 7 hours a month. At a cost of CZK 400 per hour that is CZK 2,800 a month and CZK 33,600 a year. A workflow for CZK 15,000 pays for itself in five and a half months and saves money from then on. That doesn't count the retyping errors, or the fact that the customer gets their invoice immediately instead of the next day.
| Process | Manual hours per month | Automation price | Payback |
|---|---|---|---|
| Order from the online shop to invoice and customer email | 7 h | from CZK 15,000 | 5 to 6 months |
| Website enquiry into the CRM with a notification to the salesperson | 4 h | from CZK 15,000 | 9 to 10 months |
| Weekly report from Analytics and the shop by email | 3 h | from CZK 15,000 | 12 to 13 months |
| Warehouse, shop and accounting connected both ways | 15 h | from CZK 25,000 | 4 to 5 months |
| Email triage and draft replies (AI) | 20 h | from CZK 40,000 | 5 months |
The hours are indicative, taken from projects we have done, at a rate of CZK 400 per hour. For enquiries and reports the payback looks worse on paper, but there the main gain is usually elsewhere. A salesperson who gets back to a lead within ten minutes instead of the next day closes more deals, and that doesn't fit into a table of hours.
Seven processes Czech companies automate first
Ordered by how often we come across them.
- Order to invoice. The online shop (Shoptet, WooCommerce) sends a webhook, the workflow issues an invoice in Fakturoid or Pohoda, attaches it to an email to the customer and flags the order. The most common first workflow of all.
- Enquiry into the CRM. A form on the website writes the contact into the CRM (Raynet, Pipedrive, HubSpot), assigns a salesperson and sends them a message on Telegram or in Teams. Enquiries stop getting lost in email.
- Regular reports. Every Monday at seven an email arrives with last week's revenue, traffic and number of enquiries.
- Documents in one place. A new invoice in the accounting system, or a received invoice in email, is saved to a shared drive under the right name and logged in a spreadsheet for the accountant.
- Warehouse and shop. A stock change in the ERP or warehouse system is written through to the shop and back. No more selling goods you don't have.
- Sorting incoming email. This is where AI starts to make sense: the model reads the email, classifies it (order, complaint, question, spam), pulls out order numbers and prepares a draft reply. A person checks it and sends it.
- A quote drafted from an enquiry. The enquiry becomes a draft quote with items and prices from the price list. The salesperson edits and sends it. An hour of work becomes ten minutes.
The first five items can be handled by a plain workflow without AI. The last two are work for an AI agent. We are a bit allergic to that term ourselves, because these days it gets stuck on anything with a language model inside. We mean something specific: a workflow in which the model is given tools, precisely limited permissions and an approval step before every action that has an effect on the outside world.
n8n, Make or Zapier
All three tools do the same thing: they connect applications into workflows. They differ in where they run, how you pay and what happens when a ready-made connector isn't enough.
| n8n (self-hosted) | Make | Zapier | |
|---|---|---|---|
| Where it runs | On your server or ours, in the EU | The provider's cloud | The provider's cloud |
| How you pay | Just the server, a few hundred CZK a month, no per-operation fees | Per operation, the Core plan is roughly USD 10 a month for 10,000 operations | Per task, the Professional plan is roughly USD 20 to 30 a month for 750 tasks |
| What counts as one operation | Not counted | Every step of a scenario | Every step of a Zap |
| Custom code | JavaScript or Python directly in a node | Limited | Limited |
| Where your data is | With you | With the provider, EU data centre on request | With the provider in the USA |
| Best for | More than a few scenarios, sensitive data, non-standard APIs | Quick start, low volumes | The widest range of connectors, simple scenarios |
Prices are from public price lists as of October 2026 (Make, Zapier, n8n), rounded. More important than the amount itself is how it is billed. A six-step scenario that runs fifty times a day makes 9,000 operations a month in Make. For a company with one scenario, that's fine. With ten scenarios you are two price tiers higher, and every additional workflow costs you money on the bill as well as the time to build it.
Honestly, Make is a good tool and for some of the companies that approach us it would be enough. If you have one simple scenario and low volumes, your own server is an unnecessary worry and we will happily say so. Once you have more scenarios, send personal data through your workflows, or need a step that no ready-made connector covers, n8n on your own server works out better. The community edition runs under the Sustainable Use licence, is free for internal use within a company and puts no limit on the number of executions. You pay for the server and, if needed, for someone to look after it. We can take over scenarios from Make or Zapier and migrate them; we do it quite often.
Where AI belongs in a workflow and where it doesn't
The rule we follow: AI works with text, code works with money. A language model is good where the input has no fixed structure, such as an email from a customer or an enquiry written in three sentences. It is bad at anything that requires one hundred percent reliability. We don't let it calculate the amount on an invoice.
That is why every workflow that sends money, publishes content or deletes data has an approval step. The AI prepares the material, a person clicks "approve" in Telegram or in an email. It takes five seconds and a model's mistake is stopped before it reaches a customer or the accounts.
To give an idea of costs: in our content pipeline, where AI prepares an article from several sources and runs it through a quality check, the model fees come to USD 0.15 to 0.50 per article. For email triage it is a fraction of a cent per message. What is expensive, on the other hand, is a workflow with no checks that repeats the same mistake two hundred times before anyone notices. That isn't a hypothesis; it is what most of the "broken automations" that come to us for repair look like.
A special case is the AI chatbot on a website. Technically it is a workflow too: a visitor's question, a search through prepared source material, an answer, and a handover to a human if needed. It pays off where the same questions keep coming in, typically for online shops or accommodation. We wrote about how AI is changing search itself in our article on getting cited by AI search engines.
What business process automation costs
The prices we build automations for, and what they include.
| What | Price | Delivery time | Example |
|---|---|---|---|
| One workflow | from CZK 15,000 | 1 to 2 weeks | Order from the online shop to invoice and email |
| System integration | from CZK 25,000 | 2 to 4 weeks | Online shop, warehouse and accounting exchanging data |
| AI chatbot for a website | from CZK 35,000 | 3 to 6 weeks | Answers from your own material with handover to a human |
| AI agent | from CZK 40,000 | 3 to 6 weeks | Email triage, quote drafts, content pipeline |
| Operation and monitoring | from CZK 2,500 per month | ongoing | Server, updates, monitoring, minor changes |
The server costs a few hundred CZK a month and AI fees are charged according to actual usage. We pass both on at cost, with no markup. You can leave the operation with us or take the workflows over. On handover you get documentation and training; a workflow in n8n is readable even for someone who doesn't program. A more detailed breakdown by type of integration is on our system integrations and API integration page.
For comparison with what a website costs a company: website maintenance comes to CZK 5,000 a month, so in price terms one workflow is three months of maintenance. For most companies it also pays for itself in that time.
What most often sinks an automation
The most common mistake has nothing to do with technology. A company automates a process that has no rules even in its manual form, and the workflow simply makes it faster, mistakes included. When two colleagues can't agree on when an enquiry should go to a salesperson and when to the bin, no tool will decide for them. Agree first, then build.
Second in line is silent failure. A workflow that crashes and tells nobody is worse than no workflow, because you trust it. Every run should be logged and every error should reach someone on Telegram or by email. Related to this are retries: the accounting system's API will occasionally drop out for a minute, and a good workflow tries again. A bad one loses the order and you hear about it from the customer.
Then there are credentials. A scenario built on the personal account of a colleague who leaves stops working the day they go. Use company accounts and keep credentials somewhere more than one person can reach. And finally the bill for operations that grows faster than the company does. That is the most common reason firms come to us from Make or Zapier, and it can be avoided by choosing the right tool at the start.
Data, GDPR and where things run
Workflows handle customers' personal data, so anyone in charge of GDPR at the company will ask about this. In a self-hosted setup, n8n runs on a server in the EU (we use Hetzner in Germany) or directly on your premises. Data passes through the workflow and is stored only where you send it. If a step uses an AI model, only the necessary content is sent to it, and we pick models whose API data is not used for training. In the documentation you get a description of what data flows where, so you have it for your records of processing activities.
How to start this week
- Write a list of things someone at your company does by hand and repeatedly. Ask your colleagues; they know best.
- Next to each one, note how much time it takes per month and which systems the data is in.
- Pick one that passes the four questions above and has the best ratio of savings to effort. Not the biggest one, the one that pays off soonest.
With that list, we can talk specifics. Send it to us via the contact form and within 48 hours we'll write back with what can be automated, how, and for how much. Everything we build is on our business process automation page.