AI Process Automation: Where to Start in Your Company

AI process automation sounds like something every company needs right now. The problem is that most businesses start from the wrong end: they buy a tool first, then go looking for something to use it on. This piece flips that order. We show you how to find a process actually worth automating, then test it in a quarter before you commit to anything bigger.

Why most automation projects end in disappointment

Here's the typical scenario: someone at the company sees a ChatGPT demo, or an "AI agent" tool, gets impressed, and wants it implemented. The team gets a vague brief: find a use case. A month later they have a working prototype that solves a problem nobody particularly felt. Budget spent, enthusiasm gone, AI gets shelved for a year.

We see this pattern regularly: a company asks about "AI automation" in general terms instead of pointing to a specific process. It's like asking for a home renovation without saying which room actually needs work. Good automation starts with diagnosing the process, not picking the tool.

Four criteria for choosing your first process

Before deciding what to automate, run every candidate through these four checks.

Repeatability. The process needs to look roughly the same every time. If every customer request is different and needs a different approach, AI will make more mistakes than it prevents. Look for tasks you do over and over, following the same pattern.

Volume. Automating something that happens once a month won't pay off, even if it takes a long time each time. Scale matters: the more repetitions per day or week, the faster automation pays back the cost of building it.

Cost of error. This is the criterion most often skipped. If AI misfiles an email into the wrong folder, nothing much happens: someone fixes it manually. If it sends an invoice to the wrong vendor, or misclassifies a warranty claim, the cost of a mistake is high. Start with a process where an error is cheap and easy to catch.

Data availability. AI needs data in a usable form: emails, PDFs, records in a system, historical examples of past decisions. If the data lives in people's heads or in paper folders, you need to organize it first, and that's a separate project on its own.

A process that passes all four checks is a strong pilot candidate. If it only passes two out of four, you're probably better off looking for something else to start with.

Examples of processes worth automating

A few areas that tend to work well as a first step, because they're repeatable and high-volume.

• Expense invoice processing. Reading data from a PDF or scan, assigning it to a cost category, checking it against the purchase order, and routing it for approval. A classic paper-heavy process that can largely be automated.
• Classifying requests and emails. Incoming customer messages can be sorted automatically by topic (complaint, product question, invoice request) and routed to the right person, instead of sitting in a shared inbox someone checks by hand.
• Drafting the first version of a proposal. Based on a client brief and past proposals, AI can put together a draft document that a salesperson then refines. The time saved is obvious right away, since the draft always had to be written from scratch before.
• Recurring reports. Sales summaries, marketing metrics, or project status reports that someone currently pulls manually from several systems and pastes into Excel or a presentation. A perfectly repeatable task, low error cost, high time volume.

How to calculate whether it's worth it

Before anyone writes a prompt or implements a tool, run the numbers on paper. You need three figures.

1. Hours per month the process currently takes across everyone who touches it.
2. Hourly cost of the person doing the work (salary plus overhead, divided by working hours in a month).
3. Realistic time reduction after automation. It's rarely 100%, since someone usually still needs to check the output. In our experience, 50-70% time savings on a well-chosen process is a realistic result to expect at the start.

Multiply these three numbers and you get a monthly saving in PLN. Compare it against the implementation cost: a proof-of-concept prototype that checks whether a process can even be automated typically costs somewhere in the range of tens of thousands of PLN (roughly a few thousand EUR, at 1 EUR ≈ 4.3 PLN). A full production implementation, integrated with your systems, starts at around 100,000-150,000 PLN (about 23,000-35,000 EUR), depending on integration complexity. If the monthly saving doesn't pay back the prototype cost in a reasonable time, look for a different process or wait until volume grows.

How to run a pilot in one quarter

A pilot makes sense when it has a clear start and end point, not when it's an open-ended "let's try AI" project.

Weeks 1-2: selection and data preparation. Pick one process using the criteria above and gather examples: 50-100 real cases from recent months, along with how they were previously handled. This is the data the solution will learn from or be tested against.

Weeks 3-6: building the prototype. A working version takes shape that runs on real data but isn't yet connected to any production system. You test it against past examples and compare the output to what a person actually did.

Weeks 7-9: parallel testing. The prototype runs alongside the existing process, on new, current cases. A person still makes the final call, but you track how often AI agrees with that decision and where it gets things wrong.

Weeks 10-12: decide to implement, refine, or shut down. Based on the parallel test results, you make a call: move to production, keep refining and testing, or drop this process and look for another candidate. All three outcomes count as a successful pilot, because each gives you an answer based on data rather than a gut feeling.

Common mistakes in AI automation projects

Mistake	Consequence	What to do instead
Starting from the tool, not the process	The solution goes looking for a problem, enthusiasm fades fast	Choose the process first, using the four criteria
No baseline measurement before starting	You can't prove whether the automation paid off	Measure the process's time and cost before the pilot
Automating a high-error-cost process first	One serious mistake undermines trust in the whole project	Start with a process where errors are cheap and visible
No exit plan	The company stays stuck in a project that doesn't work, reluctant to write off the investment	Set the shutdown criteria upfront, before starting
Ignoring the people affected	Team resistance, sabotaged rollout, data entered carelessly on purpose	Involve the team in choosing the process and testing from day one

Data security in automation projects

AI process automation often means routing company data, sometimes sensitive, through external models. A few rules worth setting before you start.

Check where the data physically goes and how long the model provider retains it. For customer personal data, you need a data processing agreement, the same as with any other IT subcontractor. Consider anonymizing or masking sensitive data before sending it to the model where possible: a customer ID instead of a full name is often enough for the process to work. And don't give the automation broader access than it actually needs: a tool that classifies support requests has no business seeing the company's financial data.

Frequently asked questions

Do you need a big budget to start with AI automation?

No. A sensible pilot testing one process typically costs tens of thousands of PLN (a few thousand EUR), not hundreds of thousands. Only after confirming the process is a good fit and delivers savings does it make sense to invest in a full production rollout.

How long before automation pays for itself?

With a well-chosen process, high volume, and real time savings, the pilot's return typically shows up within a few months of going live in production. What matters most is calculating the savings accurately before you start, not after the fact.

Can AI fully replace a person in a given process?

Rarely at the start, and rarely worth pushing for right away. A safer model has AI prepare the first draft of a decision or document, with a person approving or correcting the result. This limits the cost of errors and builds trust in the system across the company.

How do you choose between an off-the-shelf tool and a custom solution?

Off-the-shelf tools work well for standard tasks like sorting emails or transcription. When the process is specific to your industry or needs integration with internal systems, a custom solution usually delivers a better result, though it requires a bigger upfront investment.

If you're trying to figure out which process in your company is the right candidate for a first pilot, we can help you evaluate it and build a prototype before you commit to a full rollout. More on our approach to AI projects at DOPEST.software.