How to start implementing AI in a company: choosing the first process

This is the most common question we hear from companies: "We want AI, but where do we start?" The wrong answer begins with a tool name. The right one starts with the question: what decision or task do we want to perform faster, cheaper, or more accurately? Technology comes at the end of this path, not the beginning.

The temptation is obvious—someone saw a demo, read about a model, and wants to "have AI." But an implementation that starts with a tool usually ends up as a solution in search of a problem: a slick demo that no one uses because it doesn’t address a real need.

Reverse the order. First, list three processes that are painful today—consuming hours, prone to errors, or trapping people in dull, repetitive work. Only then ask whether and how AI can improve them. The tool is chosen to fit the process, not the other way around.

A good candidate for the first implementation has four characteristics:

• Repeatable — happens dozens or hundreds of times, not once a quarter. Repeatability is leverage.
• Time-consuming — eats up real hours that can be quantified. Without hours, there’s nothing to calculate ROI from.
• Measurable — the outcome can be evaluated numerically (percentage of correct classifications, handling time, number of tickets resolved without human intervention).
• Already exists manually — someone is doing it by hand today, so you have a baseline and "free" training data.

If a process meets these four criteria, it’s a good candidate—regardless of how "impressive" it seems from the outside. The first implementation is about proving value, not making an impression.

In Polish companies, the same ideas keep coming back—because they work:

Common denominator: narrow scope, clear outcome, existing manual process. It’s no coincidence that classifiers and extraction deliver the fastest ROI—they’re measurable by definition.

A successful implementation usually follows this path:

1. Process audit — list what’s repeatable, time-consuming, and error-prone. Check organizational readiness.
2. Choose one process — the smallest change with the biggest leverage. The automation finder can help narrow it down.
3. Calculate ROI — how many hours per month, what’s the rate, what’s the implementation cost. The ROI calculator gives you a number, not a hunch.
4. Fixed-cost pilot — a working system for one process, with measurable results, before asking for trust.
5. Gradual expansion — only after the first process proves its numbers do we add the next one.

First, data and order—then the AI layer. You don’t need perfect data, but it’s worth organizing a narrow slice for the first process, not the entire organization at once.

The rule of thumb is simple and calculable:

• If the process consumes less than a few hours per month → ROI in the first year is doubtful; start with something bigger.
• If the process consumes dozens of hours per month → implementation typically pays off in 2–4 months, and every subsequent month is savings.

That’s why we don’t choose the "coolest" process, but the one with the highest number of recoverable hours. You can calculate the return yourself in the ROI calculator—it’s deterministic math, not a promise.

Before implementing anything, describe the process, and the model will help break it down into steps suitable for automation (playground: PII masked, zero retention):

Start with a process audit, not tool selection. List three repeatable, time-consuming, and measurable processes, pick the one with the highest number of recoverable hours, and validate it with a fixed-cost pilot. The tool is chosen last, to fit the process—never the other way around.

The one that is repeatable, time-consuming, has a measurable outcome, and is already done manually today. In practice, classification and data extraction (invoices, tickets, documents) and handling common customer questions win most often—because the outcome is measurable by definition, and ROI can be calculated.

It’s even more worthwhile if the company performs a lot of repetitive manual work. If the chosen process consumes dozens of hours per month, implementation typically pays off in 2–4 months. If it’s less than a few hours—better to start with a bigger process or wait.

It doesn’t have to be perfect. We often start with a data audit and choose a process that works with what you already have (documents, FAQs, ticket history). RAG can work with existing knowledge without rebuilding everything—and where data is weak, we organize a narrow slice for the first process first.

A pilot for one process usually takes weeks, not months. We start with the smallest change with the biggest leverage, measure the outcome, and only then expand—with verification at every step.