Blog · Costs

How much does AI implementation cost in a company

Updated: 18 September 2026 · reading time about 7 minutes

"It depends" is true but useless. Below we break the cost of AI implementation into parts you can actually calculate: what pushes the price up, what is one-off, what is monthly and where a sensible budget for a small company ends. The figures are based on our own packages – other providers may split things differently, but the components are the same.

Two billing models – and why this is the most important decision

Project model. You pay for the implementation up front, usually PLN 30,000–150,000 depending on scope, and then separately for maintenance. The upside: the system is yours. The downside: a high entry cost, and AI needs ongoing tuning – after six months without care the bot's effectiveness drops, because your offer has changed and new questions have appeared.

Subscription model. You pay monthly, with implementation included in the price. The upside: a low entry cost, and someone has a stake in the system still working in the sixth month. The downside: you pay for as long as you use it. We work in this model and believe it is the fairer one for companies with fewer than a hundred people.

What really pushes the price up

• Integrations. The biggest single factor. A bot that only answers questions is cheap. A bot that checks slot availability in your calendar, creates a record in your CRM and reserves goods in the warehouse requires work on each of those systems. Especially if one of them has no API.
• Voice channel. A phone line costs more than a chat – you add telephony, per-minute costs, more testing and stricter response-time requirements. A call where the bot thinks for three seconds is unacceptable, while in a chat nobody notices.
• State of the knowledge base. If your offer, price list and procedures are written down, the implementation is short. If the knowledge lives in the heads of three people, it first has to be drawn out and organized. This can be half the project and is the most common cause of delays.
• Local model instead of cloud. A model running on your own hardware means buying a machine and maintaining it. It makes sense for sensitive data – in a law firm, a clinic, a financial company. In other cases you are usually overpaying.
• Number of languages and channels. A second language adds a few percent to the cost, not double – the knowledge base is shared. The same goes for another messaging app.

Orders of magnitude

Our packages are subscription-based and have no separate implementation fee. Roughly, it looks like this:

• Basic package – AI chatbot on your website and in messaging apps, knowledge base, a promotional website or a simple store included in the subscription.
• Standard package – additionally an AI voice agent for phone service and integrations with company systems, website included.
• Maximum package – the full scope with process automation and extensive integrations.
• Mindplix / local LLMs – priced individually, because the cost depends on hardware and security requirements.

Current amounts are on the pricing page. Below we show what happens to them once a grant is taken into account.

A worked example with a grant

75% is the upper limit, not a guaranteed rate – the level of support depends on the call for applications, company size and voivodeship. We describe the details on the page about grants for artificial intelligence.

How to calculate the return

The simplest calculation that works in customer service: count how many requests your team handles per month and how long each one takes. If that is 400 requests at 6 minutes each, that is 40 hours a month. The bot usually takes over 50–70% of them in the first months – that is 20–28 hours of recovered time.

The second item is usually bigger but harder to count: the requests that are lost today. A missed call after 5 pm, a weekend message that gets an answer on Monday. If you know your conversion rate from inquiry to order, multiplying it by the number of lost contacts gives the most convincing figure in the whole comparison.

What not to buy at the start

1. "Just in case" integrations. Connecting all your systems in the first month raises the cost and lengthens the implementation, and usually half of them turn out not to be needed. Start with one process.
2. A local model for no reason. If you do not process sensitive data and have no industry requirement, a cloud model is cheaper and better in quality.
3. The voice channel before the chat. Chat is cheaper and shows faster what customers really ask. Those transcripts are then the best material for building a phone agent.

Summary

The cost of AI implementation in a small company does not start in the hundreds of thousands – it starts with a subscription comparable to a fraction of a salary, as long as the scope is matched to one specific process. The biggest money goes into integrations and organizing knowledge, not into the model itself. And with a grant, the math changes enough that the first question should not be "how much does it cost" but "do we qualify".