AI agent & RAG cost 2026 — chatbot vs production

The AI delivery market in Poland has started publishing price lists — “chatbot from a few thousand złoty setup plus a monthly fee” and similar entry rates. That is good news for the buyer — transparency beats fog. The catch: a price without scope means nothing. The same label “AI agent” can mean a single-prompt script or a system with integrations, quality control and production maintenance. The price gap is not margin — it is a difference in product.

This article is not a price list. It is a map: what you actually buy at each level, what drives the cost, and where the cheap option stops being enough. If you want to understand where the cost in a production system really sits, we covered that separately, on production data, in what GenAI really costs in production.

Six levels — what you buy, not what you pay

The table below orders the market from the simplest chatbot to a system that has to be maintained. We give market rates only as rough context — not as our quote. Our pricing always follows scope, never a price list.

Level	What it is	What you get	What drives the cost	Where it breaks
Simple chatbot	A model on a prompt, without your data	Answers from the model’s “memory”, one channel	Configuration, prompt content	No sources, hallucinations, no auditability
Point agent	A script doing one task	Automation of one narrow action	Task logic, one integration	No boundaries, no logs, brittle on change
RAG over documents	Answers from your files, with citations	Reliable answers tied to a source	Data preparation and versioning	Data quality decides answer quality
RAG at organisation scale	RAG with access control and many sources	Company knowledge with per-role permissions	Permissions, security, retrieval scale	Without governance, risk grows, not value
Automation with integrations	An agent acting inside your systems	AI that not only answers but acts	Integrations, permissions, fallback, tests	Without evals and human oversight it acts blind
Maintenance and governance	A standing layer under all of the above	Monitoring, evaluations, compliance, fixes	Storage, retrieval, quality evaluation, audit	Skipped — the cost returns after go-live

Market rates for simple chatbots start at a few thousand złoty for setup plus a monthly fee. The further down the table you go, the less the price says on its own — because the difference is made by what a price list does not show: data, integrations, security and maintenance.

Why a cheap chatbot is often the most expensive

A low entry cost only makes sense with controlled quality. A chatbot that answers from the model’s “memory” has no source to hold on to — and in most business use cases an “almost right” answer with no citation is a risk, not a saving. It is the same principle we see in our own production data: a cheap scan that gets it wrong is expensive, because you pay twice — for fixing the error, for lost credibility, sometimes for compliance.

The cost that shows up after go-live is usually hidden at quoting time:

Data — ETL, cleaning, knowledge versioning. Usually the largest, most rarely quoted line item.
Integrations — connecting to your systems, permissions, fallback.
Maintenance — monitoring, quality evaluations, fixes when data changes.
Compliance — risk classification and documentation for the AI Act, human oversight.

The right tool for the job

This is not an argument that a cheap chatbot is bad. For simple FAQ handling on a single channel a simple chatbot is often exactly what you need — and overpaying for a production system would be the opposite mistake. The line is functional, not financial: the simple option is enough until you need reliability tied to a source, integration with systems, an audit trail and maintenance. The moment one of those requirements appears, the upfront saving turns into debt.

The two kinds of system we build answer two different requirements: RAG over company documents — when AI must answer only from your sources and cite them; AI agents — when AI must act inside your systems within set boundaries, with permissions and logs. Both assume human oversight and evaluations, because without them “production” is just a word.

What really drives the price of a production system

Counterintuitively, it is not tokens. In production the cost of a single model call can be a fraction of a cent — we covered the cost structure with numbers in a separate piece. The price is driven by three layers: data (usually most of the budget), inference (the smallest line) and maintenance (a fixed cost that grows with scale). A quote that looks only at the model is incomplete — and usually understated.

What to measure instead of comparing price lists

A price list compares entry prices. A buying decision should compare total cost and risk. Before you pick a vendor, work out:

Unit cost — what one answer or one action costs, not just the setup.
Cost of error — what one wrong answer costs in your process.
Maintenance cost — who maintains the system, and for how much, when data changes.
Time to production — when the system actually runs for you, not in a demo.
Compliance risk — whether the system is ready for AI Act obligations or bolts them on later.

That is the real “ROI calculation” — not a widget on a page, but five numbers you know best for your own process.

Don’t buy a price list — buy scope

The cheapest price list is not the cheapest delivery. The cheapest delivery is the one that works the first time and does not have to be built twice. That is why our pricing follows scope: we start with an audit that establishes what you actually need — and what it costs in your case, before you spend a single złoty on the build itself. See how we work or book an AI audit.