Where to start with AI in your company
How to choose your first use case, what it costs, and why most corporate AI projects end at the demo instead of reaching production.
Start with a process that meets three conditions at once: it's repetitive, has a countable time cost, and tolerates a mistake. Classifying incoming documents usually meets all three. "An assistant that knows the whole company" meets none of them — which is why so many rollouts that start there never finish.
Why pilots don't reach production
A demo shows a model handling a dozen or so examples chosen by the person who prepared them. Production is tens of thousands of examples nobody chose: scans crooked on the glass, documents in a format outside the test set, invoices with a handwritten note in the margin.
The difference between the two isn't model quality. It's handling the cases where the model isn't sure — and that's exactly what a rollout is actually made of. A pilot that skips it isn't an earlier version of the rollout. It's a different thing entirely.
How to choose your first use case
Make a list of candidates and score each on three dimensions.
Repeatability. How many times a month does someone do this? Below a hundred, the savings rarely justify a rollout.
Time cost. How many minutes does one instance take? Multiply by repeatability. That number sets the upper bound on a sensible budget.
Tolerance for error. What happens if the model gets it wrong? If the answer is "someone catches it at review" — good. If it's "money leaves an account" — you need a human approval step, which changes the maths.
The best candidates in a typical company: classifying and routing incoming documents, extracting data from invoices and orders, drafting first responses to repetitive support questions, searching internal documentation.
Set a success metric before you start
Before a prototype exists, write down one sentence: "This rollout makes sense if the model handles at least X% of cases on its own, at accuracy of at least Y%."
Without that number, there's no way to later decide whether the project succeeded. With it, the prototype ends in a clear decision: proceed, or stop. Stopping is also a good outcome — just a cheaper one than finding out a year in.
An architecture that survives production
A rollout that lasts has four elements a demo doesn't.
A confidence threshold. The model returns a result along with its own confidence score. Above the threshold, automation runs; below it, the case goes to a human. Setting that threshold is a business decision, not a technical one.
A queue for uncertain cases. The place where a human resolves cases below the threshold — and its size is the real measure of how well the system is working.
A decision log. A record of what the model received, what it returned, and with what confidence. Without it you can't diagnose degradation or answer an auditor's question.
Continuous measurement. Data changes. A model that was 94% accurate in March can be at 80% by November — not because it broke, but because the input stream changed.
Where your data goes
The question that comes up in every first meeting, rightly so. You have three options:
- A hosted model, no-training agreement. Data leaves the company, but the provider commits not to use it for training. Cheapest and fastest, sufficient for most use cases.
- A hosted model in your own private cloud. Data never leaves your infrastructure. More expensive, justified for sensitive data.
- A model on your own hardware. Full control, the highest entry and running cost. Makes sense under regulatory requirements that rule out the other options.
The choice is a decision about risk, not technology. You make it once, and it affects cost for the entire lifetime of the solution.
What it costs
- Prototype on your data: PLN 15–30k net, 2–3 weeks. Ends with a number you can base a decision on.
- Production rollout: PLN 60–200k net, depending on the number of integrations and approval requirements.
- Running the models: from a few hundred to a few thousand zloty a month, roughly proportional to volume.
- Maintenance and tuning: usually 15% of the rollout's value per year.
Four mistakes that come up most often
Starting with the most impressive use case. A conversational assistant impresses the board and is the hardest possible starting point — unbounded scope, no good success metric, high expectations.
No human in the loop. A system with no path for uncertain cases either performs badly or needs such a high confidence threshold that it only automates a small share of cases.
Measuring accuracy alone. 95% accuracy sounds great until you work out that at a thousand documents a day, that's fifty errors someone has to catch.
Treating the rollout as a closed project. A model needs watching, like any other piece of production infrastructure. A rollout with no measurement plan ages quietly.
The first step you can take this week
Pick one process. Count how many times a month it happens and how many minutes it takes. Write down what happens when someone gets it wrong.
Those three numbers are enough to judge whether a prototype is worth building — and they're worth more than a month of conversations about AI's possibilities.