"What will this cost?" is the first question in every AI project — and the hardest to answer, because the market quotes anything from 2 000 to 400 000 PLN without explaining what drives the difference. This article breaks the cost down: the levels of deployment, concrete ranges for SMEs in 2026, the factors that actually raise the price, and the situations where AI is not worth it at all.
Key takeaways: AI deployment in Poland costs from 3 000 PLN net for automating a single task to 60 000 PLN net for a system with an AI agent handling a process with exceptions. Maintenance runs at 10–20% of the project value per year. Three factors explain most of the price difference: the quality of your input data, the number of integrations with existing systems, and whether the process needs a model decision at all — or whether a rule would do. If a rule will do, AI is the more expensive way to achieve the same thing.
The three levels of AI deployment
The market prices AI work in three levels. The distinction matters, because most quotes mix them together and the buyer cannot tell what they are paying for.
| Level | What it covers | Time | Cost (PLN net) |
|---|---|---|---|
| 1. Rule-based automation | Connecting systems, data flow, notifications. No AI model, or a model used only to format text. | 1–3 weeks | 3 000–12 000 |
| 2. AI agent on one process | The model classifies, summarises or generates content within one clearly defined process. Exceptions handled by a person. | 4–10 weeks | 15 000–40 000 |
| 3. Multi-agent system | Several cooperating agents covering different stages, with context memory, quality control and reporting. | 3–8 months | 40 000–150 000+ |
For SMEs, levels 1 and 2 are realistic. Level 3 makes sense at volumes where a single mistake costs more than the system — handling thousands of complaints a month, for example.
What actually drives the price
The cost does not depend on "using AI" but on three variables you can assess before talking to any vendor
1. Input data quality. If the data lives in one system, in a consistent format, behind an API — the project is simple. If it has to be pulled from three places, cleaned and unified first, that preparation alone can consume half the budget. This is the most common source of overruns.
2. Number of integrations. Every connection to an external system (ERP, marketplace, accounting, email) is a separate scope of work: authentication, data mapping, error handling and API rate limits. The first integration is the most expensive, later ones are cheaper — but each one costs.
3. Type of task. Three categories, in ascending order of price
- Transformation — the model rewrites, formats, translates. Cheapest, because quality is easy to check automatically.
- Classification — the model assigns a category, priority or intent. Mid-range, because you need a test set and a measured accuracy threshold.
- Contextual decision — the model chooses an action based on several sources. Most expensive, because it needs validation, decision logging and an escalation path.
Maintenance — the part that gets forgotten
After deployment, recurring costs remain. The market standard is 10–20% of the project value per year, but the structure matters more than the number:
- Model costs (API): depend on volume. At 10 000 requests a month to a mid-tier model this is usually a few hundred PLN. At 500 000 requests, it is thousands. Project this alongside volume growth.
- Integration maintenance: external APIs change several times a year and need fixes.
- Quality monitoring: a model can drift after a version change. Without regression tests you will hear about it from a customer.
- Development: after the first deployment there are almost always more cases to cover.
If a vendor cannot break maintenance into these four lines, that is a signal they do not know what they will be maintaining.
When AI is not worth it
This is the section most price lists omit — and the most common source of burned budgets.
The process can be expressed as a rule. If the logic reads "if status = X, send email Y", you do not need a model. You need an integration. A model adds inference cost and unpredictability without adding value.
Volume is too low. At 20 cases a month, a 15 000 PLN deployment will never pay back. Calculate: deployment cost ÷ monthly saving = payback period. Above 24 months, the project should be redesigned or postponed.
The data is sensitive and there is no lawful basis. If the process handles personal, medical or financial data, you add GDPR compliance and audit work — which raises cost and lengthens the project. Sometimes it is cheaper to automate only the step that does not touch that data.
There is no process owner on your side. An AI deployment without a person who knows the process and can judge output quality ends up as a system nobody uses. This is the most common cause of failure, and it never appears in a quote.
How to compare quotes
Three questions worth asking every vendor before signing
- What exactly is in scope, and what is billed separately? Especially: data preparation, integrations, testing and team training.
- How will we measure that the deployment works? There should be a numeric criterion — "80% of tickets classified correctly" or "handling time down from 6 to 2 minutes".
- What happens when the model is wrong? Who handles the exception, how is it logged, and how does a decision go back to a human?
A quote that does not answer these three questions cannot be compared with any other, regardless of price.
Frequently asked questions
How much does AI deployment cost for a small company?
The simplest deployment — automating one task with a model used for formatting or classification — is 3 000–12 000 PLN net. A deployment with an AI agent handling a process with exceptions: 15 000–40 000 PLN net. Below 3 000 PLN we are usually talking about configuring an off-the-shelf tool, not a dedicated deployment.
Can AI be deployed without programming?
For some use cases, yes — no-code tools can connect a model to a form or a spreadsheet. The limit appears at integration with business systems (ERP, marketplace, accounting): those need an API and error handling, which requires a developer.
How long does AI deployment take?
- Automating a task: 1–3 weeks. An agent on one process: 4–10 weeks. A multi-agent system: 3–8 months. The longest stage is often not the model work but preparing and cleaning the data.
Are AI maintenance costs fixed?
No. They consist of API costs (growing with request volume), integration maintenance (external system changes), quality monitoring and development. Plan 10–20% of the project value per year, but at high volume API costs can outweigh everything else.
How do I calculate the return on AI deployment?
Work out the cost of the manual process: monthly cases × minutes per case × fully loaded cost per minute. Add the cost of errors you avoid. Divide the deployment cost by the monthly saving to get the payback period in months.
Can AI replace an employee?
In practice it replaces tasks, not jobs. The most common outcome is moving a person from manual work to handling exceptions and quality control. In companies with a well-designed process, the same team handles a larger volume without extra hiring.
Sources and tools
- OpenAI — API pricing (reference point for model costs)
- Anthropic — Claude API pricing
- NIST AI Risk Management Framework
- European Commission — data protection and GDPR
👉 Want to cost out your own process? See the Automation & AI service — we start with a process map and measurement, not a blind quote. Indicative ranges are in pricing. Get in touch via the contact form or at kontakt@informatix.com.pl.
👉 Not sure you need AI at all? Read what business process automation is and AI in business: prompts and data security.
📖 See also: e-commerce integrations, designing resilient API and webhook integrations and CI/CD and DevOps for Laravel.