Business process automation does not start with a tool — it starts with a process map. Most deployments I have seen stalled not on technology, but because nobody had measured what the manual process actually costs. This article covers the basics: how automation differs from digitalisation and from AI, which processes to tackle first, how to calculate the return, and what it really costs in 2026.
Key takeaways: Business process automation (BPA) replaces repetitive, rule-based steps with a flow executed by a system — no human decision at every step. The fastest payback comes from processes with high volume, low variability and an expensive error: order handling, invoicing, stock synchronisation and reporting. For an SME, automating a single process typically costs 3 000–25 000 PLN net, with 10–20% of that amount per year for maintenance. Start with one process and measure the time before you deploy — otherwise you cannot prove the return.
What business process automation actually is
Business process automation means connecting applications, data and rules into a single controlled workflow where the system moves from step to step on its own, and people handle only the exceptions.
The key word is "rule-based". If the instruction can be written as "if X, then do Y", the process can be automated. If every case requires a judgement based on context that is not in the data, it cannot — at least not entirely.
A practical e-commerce example: an order from a marketplace has to reach the store, update stock, generate an invoice and send the tracking number back to the marketplace. Four steps, four systems, zero business decisions along the way — a textbook automation candidate. The exception that stays with a human is, for instance, an order with an unusual delivery address.
Automation, digitalisation and AI are three different things
These terms are used interchangeably, but they describe entirely different levels of intervention
- Digitalisation: moving a process from paper into a system. The document exists as a file, but a human still performs every step. Example: a PDF invoice instead of a paper one, issued by hand.
- Automation: the system performs steps according to rules. Example: the invoice generates itself when the order status changes.
- AI: the system makes decisions where rules are not enough — classifying, summarising, recognising intent. Example: a model recognises that a customer message is about a return and routes it to the right queue.
The order matters. Deploying AI on a process that was never rule-automated is the most common mistake I see — the model is handed a task that a simple rule would solve in 90% of cases, so you pay for inference where a condition would have been enough.
Which processes to automate first
Do not automate everything at once. Use a simple ranking based on four criteria — score each process from 1 to 5 and add the points:
| Criterion | Control question | Weight |
|---|---|---|
| Volume | How many times per month does the process run? | high |
| Variability | What share of cases needs a human decision? | high (lower is better) |
| Cost of error | What does a mistake cost — complaint, penalty, lost customer? | medium |
| Manual time | How many minutes does one run take? | medium |
A process with high volume, low variability and an expensive error is your first candidate. In practice the winners are usually:
- Order handling across multiple channels (store, marketplace, phone)
- Invoicing and passing data to accounting
- Stock and price synchronisation between channels
- Periodic reporting assembled from several systems
- Repetitive enquiries (order status, availability, documents)
A process with low volume and high variability — pricing a bespoke project, for example — should stay with a human. Automating it costs more than it saves.
What it costs and how to calculate the return
Market ranges for SMEs in 2026
| Scope | Time | Cost (PLN net) |
|---|---|---|
| Single simple process (form, notification, data export) | 1–2 weeks | 3 000–8 000 |
| Process with two-system integration and exception handling | 3–6 weeks | 8 000–25 000 |
| Process with an AI agent (classification, summarising, content generation) | 6–10 weeks | 20 000–60 000 |
| Maintenance and development | monthly | 10–20% of project value per year |
Calculate the return as simply as possible
Monthly saving = number of runs × minutes per run × cost per minute of work
If order handling takes 6 minutes, you run it 400 times a month, and the fully loaded cost of a minute is 0.80 PLN — the saving is 1 920 PLN per month. A 12 000 PLN deployment pays back in just over six months.
Then add the cost of the errors you avoid. In e-commerce, overselling is not just a complaint — it is a drop in seller quality on the marketplace, which affects listing visibility.
A plan for the first 30 days
Week 1 — map and measure. List the process steps, measure the real time of one run (with a stopwatch, not from memory) and count the volume. Without these two numbers there is no case for the project.
Week 2 — choose and scope. Pick the process with the highest ranking score. Define what counts as an exception and who handles it. Write down the success criterion: "handling time drops from 6 to 1 minute".
Week 3 — technical design. Determine which systems must connect and whether they expose an API. This is where surprises usually appear — no API means file-based work, which lengthens the project and increases risk.
Week 4 — pilot deployment. Run the process on a slice of data (say 10% of orders), compare the result against the success criterion, and only then roll it out.
The most common mistakes
- Automating chaos. If the process is badly defined, automation will cement the mess. Simplify first, automate second.
- No measurement before deployment. Without a baseline you cannot prove the return or win support for the next stage.
- Chasing 100% automation. The last 10% of cases usually costs more than the first 90%. Leave the exceptions to people.
- Starting with AI. A model at the front of a process that has no rules is cost without benefit.
- Ignoring data compliance. Automation moves data between systems, so what you send must comply with GDPR. I cover the details in the article on AI in business and data security.
Frequently asked questions
Is process automation the same as RPA?
No. RPA (Robotic Process Automation) mimics a human clicking through an interface — useful when a system has no API. Process automation works at the level of integrations and data, which is more stable and cheaper to maintain. Choose RPA only when there is no other route to the data.
Where should a small company start with automation?
Start with the process you run most often and which is the most repetitive — usually order handling, issuing documents or reporting. Measure the time of one run and multiply it by the monthly number of runs. If the result is more than a few hours a month, automation makes economic sense.
How long does an automation project take?
A simple process with one system-to-system connection: 1–2 weeks. A process with two-system integration and exception handling: 3–6 weeks. A process with an AI agent: 6–10 weeks, because model output quality has to be tested.
Will automation replace employees?
In practice it replaces tasks, not jobs. The usual outcome is moving a person from manual work to handling exceptions and quality control. In companies that did this well, the same team handles a larger volume.
Do I need AI to automate a process?
In most cases, no. AI makes sense where data is unstructured (email content, ticket descriptions, documents) or where the rules are too complex to write down. If the process can be expressed as conditions, a plain integration will do and will be cheaper to maintain.
How do I measure the return on automation?
- Three numbers: the time of one run before deployment, the number of runs per month, and the fully loaded cost of a minute of work. Their product is your monthly saving. Divide the deployment cost by that figure to get the payback period. Add the value of the errors you avoided.
Sources and tools
- Google Cloud — Business Process Automation
- n8n — workflow automation documentation
- Make — integrations and scenarios
- European Commission — data protection and GDPR
👉 Want to automate a specific process? See how the Automation & AI service works — starting with a process map, then system integrations, then an AI assistant handling the exceptions. Get in touch via the contact form or at kontakt@informatix.com.pl.
👉 Connecting sales systems? See e-commerce integrations — Baselinker, Allegro, Amazon, ERP and accounting systems. For indicative ranges, see pricing.
📖 See also: AI implementation cost, designing resilient API and webhook integrations and CI/CD and DevOps for Laravel.